Imagine you want to buy a laptop.

You could open half a dozen reviews, compare battery tests, read repairability scores, check user forums, and decide which experts you trust.

Or you could ask an AI assistant:

Which laptop under $1,500 has the best battery life and is likely to last five years?

Seconds later, you have an answer.

That answer may depend on reviewers who obtained the laptops, ran tests, measured displays, opened the chassis, compared warranties, interviewed manufacturers, and updated their recommendations over time.

You receive the benefit of their work.

But you may never visit them.

Now repeat that behavior across recipes, travel guides, medical explainers, financial information, software documentation, restaurant reviews, product testing, local news, investigative journalism, and millions of everyday questions.

That is the economic problem hiding underneath AI search.

The threat is not that artificial intelligence suddenly “kills the internet.” It is subtler — and more plausible:

AI can make information dramatically cheaper to retrieve while making high-quality information harder to finance.

AI changed what search is for

For most of the modern web, search engines and publishers had an imperfect but mutually useful relationship.

A publisher invested in an article, review, database, or guide. Google indexed it. Someone searched for information, discovered the page, and sometimes clicked through.

That visit could become advertising revenue, an affiliate purchase, a subscription, a donation, a customer, or simply a reader who returned later.

The system was hardly generous. Publishers complained about platforms for years, and there was never a guarantee that being crawled would produce traffic.

But the underlying exchange was understandable:

Publish useful information → get discovered → receive visitors → monetize some of those visits.

Generative search changes the transaction.

Instead of primarily helping users find information, AI systems increasingly read, combine, and explain information for them.

The source page and the AI-generated answer can now satisfy the same demand.

And the early evidence suggests that when the answer is already on the screen, many people do not click.

Pew Research Center analyzed 68,879 Google searches made by 900 U.S. adults. When an AI summary appeared, users clicked a traditional search result in just 8% of visits, compared with 15% when no AI summary appeared. They clicked sources cited inside the AI summary in only 1% of visits (Chapekis & Lieb, 2025).

That finding alone does not prove AI caused the difference.

But a randomized field experiment by Saharsh Agarwal and Ananya Sen offers stronger evidence. When Google AI Overviews appeared, outbound organic clicks fell by 39.8%, while zero-click searches increased by 34.5%. Search frequency itself did not significantly decline (Agarwal & Sen, 2026).

That distinction matters.

People were still searching.

They were simply less likely to leave the platform.

Synthesis is becoming cheap. Production is not.

This is where the debate becomes larger than publisher traffic.

Artificial intelligence is extraordinarily good at reducing the cost of retrieving, combining, summarizing, and explaining information.

But those are not the same activities as producing the information in the first place.

A restaurant review still requires someone to visit the restaurant.

A serious product comparison requires products, testing, time, expertise, and often specialized equipment.

A medical reference page requires expert review and updating.

Investigative journalism can require months of reporting before a single story appears.

A local newspaper must pay someone to attend the city council meeting before an AI model can summarize what the council decided.

AI can make the downstream process — finding and synthesizing those facts — vastly more efficient.

It does not automatically make discovering and verifying the facts cheaper.

That creates a strange economic asymmetry.

The better AI becomes at extracting useful answers from the web, the less often a user may need to visit the websites that financed the underlying work.

Over time, the system risks rewarding synthesis while weakening the economics of production.

And information production needs something that AI systems cannot generate from an archive alone:

replenishment.

Yesterday’s web cannot tell us what happened at tomorrow’s council meeting.

AI referrals are not replacing search

One possible answer is that AI assistants will simply become the next generation of referral engines.

There is evidence that AI referrals are growing rapidly.

But they remain tiny compared with traditional web traffic.

Chartbeat, which measures billions of pageviews across thousands of media sites, reported that referrals from ChatGPT grew more than 200% year over year. Yet AI chatbots still accounted for less than 1% of total pageviews across its network.

At the same time, Google Search pageviews fell 34% between December 2024 and December 2025. The decline was especially steep for smaller publishers: 22% for large sites, 47% for medium-sized sites, and 60% for small publishers (Chartbeat, 2026).

Those numbers should be interpreted carefully. Search traffic changes for many reasons: algorithms, changing news cycles, user behavior, Google Discover, publisher strategy, and AI can all play roles.

But they expose a structural vulnerability.

Large publishers can diversify. They can build subscriptions, events, apps, commerce businesses, licensing deals, newsletters, and recognizable brands.

A small specialist publication may have none of those advantages.

Yet niche publishers are often exactly where highly specific expertise lives.

If the economics deteriorate, the first casualties may not be the biggest media companies.

They may be the small sites whose disappearance barely makes news — but whose work quietly made the web useful.

But AI search is genuinely better for users

There is a powerful counterargument.

Why should consumers preserve an inefficient business model simply because publishers became dependent on it?

If five webpages contain pieces of the answer to a question, why should someone have to open all five, navigate cookie banners, scroll past advertising, and reconstruct the answer manually when an AI system can do it in seconds?

They should not.

AI search solves a real problem.

Google also disputes the idea that AI search is simply destroying the web’s traffic economy. In August 2025, Google Search head Liz Reid said total organic click volume from Search to websites had remained relatively stable year over year and that Google was sending slightly more “quality clicks” — visits in which users did not immediately return to the search page (Reid, 2025).

That argument deserves to be taken seriously.

AI may create new searches that previously never happened. It may filter out low-value clicks while sending more motivated users to publishers.

Fewer clicks and better clicks can coexist.

But even if that is true in aggregate, it does not eliminate the underlying incentive problem.

The question is not whether every lost click deserves to be preserved.

It is whether the organizations paying to produce valuable information can continue capturing enough value to keep producing it.

That is a different question.

And the answer cannot simply be: make people click more.

The web needs a new bargain

The encouraging part is that a replacement economic model is beginning to emerge.

Microsoft’s Publisher Content Marketplace, announced in February 2026, is explicitly designed to let publishers license premium content to AI systems, define usage terms, receive usage reporting, and be paid based on delivered value. Microsoft describes the traditional web as an implicit exchange in which publishers supplied content and distribution channels helped people find it — an exchange that does not translate cleanly into an AI-first world (Frank, 2026).

Cloudflare has experimented with another model: Pay Per Crawl. Instead of forcing publishers to choose between allowing AI crawlers free access or blocking them entirely, the system allows participating sites to charge crawlers for access (Allen & Newton, 2025).

Neither model is guaranteed to succeed.

But they reveal the direction of travel.

The future may include direct licensing, collective marketplaces for smaller publishers, paid APIs, usage-based revenue sharing, crawler payments, clearer opt-out controls, and better measurement of how AI systems use and refer to publishers.

Publishers themselves will also have to adapt.

Search dependency is becoming a business risk. Newsletters, memberships, communities, apps, direct traffic, events, and recognizable brands matter more when platforms control an increasing share of discovery.

The goal should not be to preserve the economics of 2015 forever.

It should be to create an economic architecture appropriate for 2030.

Citations are not compensation

AI companies are right to emphasize citations.

People should know where claims come from. Original sources should remain visible. Attribution is essential for trust and accountability.

But citation solves a provenance problem.

It does not necessarily solve an economic problem.

Pew’s 1% source-click finding makes the distinction difficult to ignore.

A citation can identify the journalist.

It cannot pay the journalist.

A link can credit the product-testing site.

It cannot buy the next ten devices that need to be tested.

An attribution can point toward the local newspaper.

It cannot finance someone to attend next week’s planning meeting.

If AI becomes the primary interface through which people consume information, the industry will need something more durable than the hope that enough users voluntarily click the footnotes.

AI has a supply-chain problem

The real danger is not that artificial intelligence destroys the open web overnight.

It is that parts of the web slowly become less worth producing.

At first, users may barely notice.

There will still be billions of webpages. AI systems will still produce polished answers. Decades of accumulated human knowledge will remain available to retrieve and summarize.

But quality information is not a static natural resource.

Someone has to conduct the experiment.

Someone has to test the product.

Someone has to interview the witness.

Someone has to update the database.

Someone has to investigate the company.

Someone has to attend the meeting.

Someone has to discover the fact before anyone — or any machine — can summarize it.

The question is therefore not whether AI should make information easier to access.

It should.

The question is whether we can build an economy in which making information easier to consume does not make valuable information progressively less rational to produce.

An AI ecosystem that consumes the open web without helping replenish its economics is not merely disrupting an obsolete middleman.

It is eating its supply chain.

References

Agarwal, S., & Sen, A. (2026). The impact of Google AI Overviews on publisher traffic and user experience: Evidence from a field experiment. SSRN.
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513059

Allen, W., & Newton, S. (2025, July 1). Introducing pay per crawl: Enabling content owners to charge AI crawlers for access. Cloudflare Blog.
https://blog.cloudflare.com/introducing-pay-per-crawl/

Chapekis, A., & Lieb, A. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center.
https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

Chartbeat. (2026). Pageviews are down, but AI’s impact is complicated.
https://chartbeat.com/resources/articles/pageviews-down-ai-impact/

Frank, T. (2026, February 3). Building toward a sustainable content economy for the agentic web. Microsoft Advertising.
https://about.ads.microsoft.com/en/blog/post/february-2026/building-toward-a-sustainable-content-economy-for-the-agentic-web

Reid, L. (2025, August 6). AI in Search is driving more queries and higher quality clicks. Google.
https://blog.google/products-and-platforms/products/search/ai-search-driving-more-queries-higher-quality-clicks/

For most of the last thirty years, enterprise software has been built around a simple assumption:

People operate software.

Salespeople open CRM systems. Finance teams work inside ERP platforms. HR operates HCM software. Customer-support teams live in ticketing systems. Project managers open project-management applications. Employees move information between documents, dashboards, inboxes, databases and business systems.

The architecture of the modern enterprise reflects that assumption.

At its simplest:

employee → application → database

The cloud changed where those applications ran.

SaaS changed how they were bought.

Mobile changed where people could access them.

Generative AI initially changed how people interacted with them by adding copilots, chat interfaces and content generation.

But AI agents introduce a more fundamental possibility.

Instead of requiring an employee to understand which applications contain the necessary information, navigate those systems, interpret the data, decide what to do and manually execute each step, an AI system can increasingly:

• interpret an objective,

• retrieve the relevant organizational context,

• identify the systems required,

• reason across data and policies,

• call software tools,

• execute approved actions,

• maintain workflow state,

• evaluate results,

• and escalate decisions that require human judgment.

The architecture begins to change from:

employee → application → database

to:

business intent → governed AI agents → workflows → systems of record

That shift is much more consequential than putting a chatbot inside Salesforce, SAP, ServiceNow or Microsoft 365.

It changes where value sits in the enterprise technology stack.

And it suggests that the next major computing architecture inside companies will not simply be another generation of SaaS.

It will be an Enterprise Execution Fabric: a governed layer through which intelligent systems interpret business intent, access organizational context, exercise bounded authority and execute work across applications.

The enterprise application does not disappear.

But it stops being the center of the experience.

The workflow becomes the center.

AI Adoption Is Exploding. Enterprise Transformation Is Not.

The numbers explain why enterprise leaders are paying attention.

McKinsey’s 2025 survey found that 88% of respondents said their organizations regularly used AI in at least one business function, up from 78% the previous year. Generative-AI use increased from 71% to 79%.

But only roughly one-third of organizations had begun scaling their AI programs.

For agentic systems specifically, 23% reported scaling an agentic system somewhere in the enterprise, while another 39% were experimenting.

Gartner tells a similar story.

It expects as many as 40% of enterprise applications to contain task-specific AI agents by the end of 2026, versus less than 5% in 2025.

Yet Gartner’s 2026 research found only 17% of organizations had actually deployed agents, while more than 60% expected to do so within two years.

That is the central contradiction of enterprise AI:

AI is mainstream as a tool, but still immature as an operating model.

The money flowing into the market is enormous.

IDC estimated worldwide AI spending at nearly $235 billion in 2024, growing to more than $631 billion by 2028.

Gartner uses a much broader definition and forecasts $2.52 trillion in worldwide AI spending in 2026, up 44% year over year.

Within the model-and-platform layer alone, Gartner expects spending to rise from $39.3 billion in 2025 to $64.3 billion in 2026, a 63.4% increase.

Its AI-agent software forecast is even larger:

• $86.4B in 2025

• $206.5B in 2026

• $376.3B in 2027

These estimates use different taxonomies and should not be treated as measurements of the same market.

But the direction is unmistakable: AI investment is expanding simultaneously across infrastructure, models, development platforms and enterprise applications.

Regional growth is broad as well.

IDC’s comparable global forecast estimated:

• North America: ~$131B in 2024 → ~$353B in 2028

• EMEA: ~$50.5B → ~$147B

• Asia-Pacific and Japan: ~$45.3B → ~$110B

• Latin America: ~$7.8B → ~$21.4B

Newer IDC regional forecasts put European AI spending at $144.6B by 2028, a 30.3% CAGR, and Asia-Pacific spending at $175B by 2028, with a 33.6% CAGR from 2023.

The adoption curve is moving faster than the economic-value curve.

And that distinction matters.

The Model Is Not the Transformation

The evidence already suggests that AI can materially improve productivity.

A large NBER field study involving customer-support workers found that AI assistance increased productivity by nearly 14% on average.

Less-experienced and lower-skilled workers saw gains of approximately 35%.

A Procter & Gamble field experiment found that individuals using generative AI could perform at roughly the level of two-person teams without AI. Individuals using AI spent 16.4% less time on the task, while AI-enabled teams spent 12.7% less time.

Morgan Stanley says more than 98% of its financial-adviser teams actively use its internal AI assistant.

Controlled GitHub studies have also reported faster task completion and improvements in some code-quality measures among developers using Copilot, although vendor-produced research should not be treated as equivalent to independent evidence.

The productivity signal is real.

But the more important lesson is easy to miss.

Installing intelligence does not automatically redesign work.

The highest-performing organizations appear to be doing something more fundamental.

McKinsey has repeatedly found workflow redesign to be strongly associated with enterprise AI value. AI high performers are nearly three times as likely as other organizations to report fundamentally redesigning workflows.

A 2026 McKinsey analysis found that roughly 79% of organizations skip the step of decomposing workflows and explicitly deciding what AI should do, what conventional software should do and what humans should do.

That suggests a more realistic equation for enterprise AI value:

AI ROI = model capability × workflow redesign × usable context × adoption × governance × unit economics

It is multiplicative rather than additive.

If the model is excellent but the data is unusable, value collapses.

If the data is excellent but employees refuse to use the system, value collapses.

If employees adopt it but the workflow remains badly designed, value collapses.

If the workflow works but every transaction costs more than the human alternative, value collapses.

If the economics work but the system cannot be trusted with meaningful authority, value collapses.

This is why so many impressive AI demonstrations struggle to become durable enterprise systems.

They optimize the intelligence layer while leaving the operating model untouched.

The Enterprise Execution Fabric

The emerging enterprise architecture can be understood through five layers:

1. Intent

What outcome is the organization trying to achieve?

Not:

“Open the CRM.”

But:

“Reduce renewal risk in our highest-value accounts.”

Not:

“Run the invoice report.”

But:

“Identify invoices that are likely to miss payment terms and resolve the exceptions.”

Intent is expressed in business language rather than application language.

2. Context

What does the system need to know to act correctly?

That includes:

• authoritative customer information,

• contracts,

• transaction history,

• policies,

• product data,

• pricing,

• organizational roles,

• previous workflow actions,

• semantic definitions,

• permissions,

• business rules,

• and exceptions.

The model may provide intelligence.

But context determines whether that intelligence understands the enterprise.

3. Authority

What is the system allowed to do?

Can it:

• read a record?

• update it?

• send an email?

• issue a refund?

• approve an expense?

• generate a purchase order?

• schedule a technician?

• move money?

• change production parameters?

Agentic AI turns authorization from a security detail into a central product-design problem.

4. Action

Which software tools and business systems can execute the required work?

The agent may call:

• CRM,

• ERP,

• HCM,

• supply-chain systems,

• email,

• collaboration tools,

• databases,

• analytics systems,

• customer-support software,

• industrial systems,

• payment infrastructure,

• or specialized APIs.

The applications remain important.

But increasingly they operate behind the workflow rather than defining it.

5. Accountability

How does the organization know whether the system worked correctly?

This requires:

• logging,

• evaluations,

• observability,

• approvals,

• traceability,

• exception handling,

• cost measurement,

• policy enforcement,

• incident response,

• and human accountability.

This fifth layer is what separates an AI demonstration from an enterprise operating system.

Together, these five layers form an Enterprise Execution Fabric:

Intent → Context → Authority → Action → Accountability

That may become a more useful way to think about enterprise AI than organizing technology primarily around individual applications.

What This Looks Like in Practice

Consider a renewal workflow.

Today, an account manager might need to:

1. open CRM;

2. identify upcoming renewals;

3. inspect product usage;

4. check recent support cases;

5. review customer sentiment;

6. calculate account economics;

7. check contract terms;

8. determine discount authority;

9. draft outreach;

10. schedule follow-up activity;

11. update CRM;

12. escalate risky accounts.

Each task may involve a different system.

The company’s process exists partly inside software and partly inside the employee’s head.

Now imagine the employee instead says:

“Find the five renewal accounts at greatest risk this quarter. Explain why each is at risk, propose the strongest retention offer permitted under our margin policy, draft outreach and schedule follow-ups after I approve.”

Behind that request, an agent may:

• query CRM,

• retrieve product telemetry,

• analyze support history,

• inspect billing information,

• check contractual terms,

• apply pricing and margin policy,

• rank account risk,

• draft recommended actions,

• request approval,

• send communications,

• schedule tasks,

• and update the systems of record.

The employee no longer needs to know where every piece of information lives.

The agent does.

That is the architectural shift.

The Same Pattern Appears in Finance

Consider invoice reconciliation.

A traditional process may involve:

• ERP,

• bank records,

• purchase orders,

• email,

• document storage,

• tax information,

• supplier communications,

• and manual exception queues.

A useful agent does not simply summarize invoices.

It can:

• compare invoice data with purchase orders,

• identify mismatches,

• retrieve supplier contracts,

• classify the discrepancy,

• determine whether the difference falls within an approval threshold,

• request supporting documentation,

• reconcile permitted exceptions,

• and send uncertain cases to finance staff.

The AI does not need unrestricted autonomy.

It needs enough authority to resolve routine cases and enough governance to prevent dangerous ones.

That distinction leads to one of the most important principles of enterprise AI.

The Sweet Spot Is Bounded Autonomy

Enterprise AI is often discussed as though the end state is maximum autonomy.

That is unlikely to be the winning design.

The economic sweet spot is bounded autonomy.

Give AI enough authority to complete useful work.

Constrain that authority according to risk.

An agent summarizing an internal meeting may need almost no supervision.

An agent drafting customer communication may need policy checks.

An agent issuing a $50 service credit might act automatically under a predefined rule.

An agent approving a $50,000 payment may need explicit authorization.

An agent terminating employment, transferring millions of dollars or controlling industrial equipment should operate under substantially stricter controls.

The governing principle should be:

Autonomy should increase with reversibility and decrease with risk and blast radius.

That principle sounds simple.

Its implications are enormous.

A hallucinating chatbot produces a bad sentence.

A hallucinating agent with purchasing authority can produce a bad transaction.

A poorly governed recommendation system can influence a decision.

A poorly governed execution system can change reality.

That means governance can no longer live primarily in policy documents, compliance reviews or legal checklists.

Governance becomes part of the runtime architecture.

Enterprise AI Needs a Control Plane

The durable enterprise architecture will not be “pick one LLM and connect company data to it.”

It is more likely to involve:

Users and events
↓
AI assistants and agents
↓
Agent orchestration — planning, routing, memory and state
↓
Governance and policy — identity, permissions, approvals and audit
↓
Model routing — frontier models, smaller models, domain models, multimodal models, edge models
↓
Enterprise context — data, documents, semantic definitions, metadata and permissions
↓
Business systems — ERP, CRM, HR, IT, supply chain and physical operations

The LLM becomes one replaceable component inside that architecture.

Gartner expects spending on domain-specific and specialized generative models to grow 210% in 2026, suggesting enterprises will increasingly route different tasks to different models based on accuracy, privacy, latency, language and cost.

A frontier model may make sense for difficult reasoning.

A smaller model may be better for classification, extraction or routing.

An edge model may be necessary inside a factory, hospital or retail location where latency, sovereignty or connectivity matter.

The model layer therefore becomes a portfolio, not a monopoly.

Context Becomes the New Software Moat

Early enterprise GenAI focused heavily on retrieval-augmented generation: search documents, retrieve passages and feed them into an LLM.

Agents need much richer context.

They need to understand:

• which customer record is authoritative,

• which price is current,

• what permissions apply,

• what happened earlier in the workflow,

• which employee has approval authority,

• which policy governs the transaction.

So the future data moat is not: “We have more PDFs.”

It is: “We know what our data means, which version is authoritative, who may access it and what actions may follow from it.”

That changes enterprise software competition.

In traditional SaaS:

moat ≈ features + UX + workflow + switching cost

In AI-native enterprise software:

moat ≈ proprietary context + workflow permissions + system-of-record depth + distribution + evaluations + trust

This is why application companies are moving into data, data companies are moving into AI applications, workflow companies are building agents and model companies are creating enterprise tool integrations.

Everyone is converging on context.

The Return of Platform Consolidation

For twenty years, best-of-breed SaaS benefited from the assumption that companies could simply integrate specialized applications through APIs.

AI introduces a counterforce.

Every system boundary creates additional:

• context loss,

• authentication complexity,

• governance complexity,

• integration work for agents.

That does not mean enterprises will return to one monolithic vendor.

A more plausible equilibrium is: consolidated core, composable edge.

Large enterprise platforms are already moving in this direction.

SAP is positioning Joule around assistants and agents operating across business processes.

Oracle’s Release 26B introduced 22 Fusion Agentic Applications across ERP, HCM, SCM and CX, describing the transition from a “system of record” to a “system of outcomes.”

ServiceNow describes its AI Platform as combining AI, data and workflows into an orchestration layer.

Salesforce is combining CRM, Data Cloud, Slack, Tableau and Agentforce, including usage-based Flex Credits.

Microsoft can connect Dynamics 365 with Microsoft 365, Azure, Fabric, Power Platform and its broader Copilot ecosystem.

And newer AI-native platforms such as T1U are attempting to start with the integrated architecture itself. T1U markets an AI-native “business operating system” spanning CRM, finance, HR, inventory, projects, documents, support, automation and other functions through its NeoMind AI layer.

The strategic logic is credible.

If CRM, finance, service and operational data sit inside the same environment, an agent can reason across a richer business context without stitching together as many independent SaaS databases.

Potential advantages include:

Less integration tax.
Fewer systems need to exchange and duplicate data.

Better context.
AI can reason across multiple business functions.

Faster automation.
Cross-functional workflows can be built inside one environment.

Simpler governance.
Identity, permissions and audit can potentially share one model.

How the Platforms Compare

Public pricing as of August 19, 2026 gives a rough sense of positioning, although these offerings are not functionally equivalent:

Pricing excludes implementation, services, usage overages, negotiated discounts and other costs.

Adjacent competitors include Workday, Oracle NetSuite, HubSpot, monday.com, Databricks, Snowflake and NVIDIA AI Enterprise.

The important competitive question is no longer:

Who has AI?

Almost everyone does.

It is:

Who can combine the broadest useful context with the safest ability to act?

AI Will Also Break SaaS Pricing

Traditional SaaS assumes revenue scales with human seats.

Agents complicate that model.

If ten agents perform work previously handled by fifty employees, charging purely per human user becomes increasingly misaligned with delivered value.

Enterprise vendors are therefore experimenting with:

• human-seat pricing,

• suite subscriptions,

• premium AI seats,

• token or usage pricing,

• credit systems,

• infrastructure subscriptions,

• outcome-based pricing.

Salesforce’s Agentforce packaging already combines conventional subscriptions with Flex Credits and Data Cloud credits.

The likely long-term model is hybrid:

platform fee + human seats + agent consumption + premium context/data + outcome-linked pricing

And the next FinOps question will not be: “How many tokens did we consume?”

It will be: “What did it cost to resolve this ticket, reconcile this invoice, qualify this lead or investigate this security incident compared with the human baseline?”

Cost per completed business process becomes the metric that matters.

Regulation Makes Governance a Buying Criterion

The EU AI Act is already turning governance into a procurement requirement.

The Act entered into force in 2024.

General-purpose AI obligations became applicable in August 2025.

Transparency requirements took effect on August 2, 2026.

Following the 2026 AI Omnibus changes, Annex III high-risk requirements are scheduled for December 2, 2027, while requirements for high-risk AI integrated into regulated products are scheduled for August 2, 2028.

That means enterprises increasingly need answers about:

• model provenance,

• documentation,

• human oversight,

• logging,

• data rights,

• incident procedures,

• transparency,

• vendor responsibilities,

• deployer responsibilities.

NIST’s AI Risk Management Framework and GenAI Profile reinforce the same basic principle:

governance is a lifecycle process, not a pre-launch legal review.

Gartner predicts that more than 40% of agentic-AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

It separately forecasts that 40% of enterprises could demote or decommission autonomous agents following production governance failures.

Those are forecasts, not observed outcomes.

But the warning is important.

The future is likely to involve high automation and low tolerance for unconstrained autonomy.

AI Changes Jobs by Breaking Them Into Tasks

McKinsey’s 2025 survey found limited AI-driven workforce reduction so far.

Across functions, the median share reporting workforce decreases attributable to AI during the preceding year was 17%, while a median 30% expected decreases during the following year.

At the same time, companies continued hiring AI-related engineering talent.

That suggests job decomposition may be more useful than thinking only about job elimination.

Most jobs combine four categories:

1. repeatable information processing;

2. judgment based on known rules;

3. judgment under ambiguity;

4. human relationships, accountability or physical execution.

AI reaches the first category fastest and increasingly moves into the second.

It assists the third.

The fourth remains harder to automate.

So the transformation unit should not be the job title.

It should be the workflow.

The managerial role changes too.

AI can increasingly compile status updates, synthesize information, draft documents, schedule activity and track exceptions.

But prioritization, coaching, conflict resolution, context and accountability remain human responsibilities.

Management therefore shifts from: moving information through hierarchy

toward: designing and supervising human-machine systems.

What Enterprise Software Could Look Like in 2031

By 2031, the distinction between “AI software” and “software” may largely disappear.

Systems of record will remain.

Transactional integrity still matters.

Accounting systems still need authoritative ledgers.

Customer systems still need authoritative records.

HR systems still need controlled employee data.

Industrial systems still need reliable operational state.

But humans may increasingly interact with those systems indirectly.

Instead of navigating applications, employees describe outcomes.

Instead of manually coordinating workflow, they supervise execution.

Instead of learning where information lives, they ask systems that already understand organizational context.

A plausible 2031 enterprise could include:

• bounded agents executing multistep workflows;

• portfolios of models rather than one universal model;

• enterprise semantic and context layers;

• machine identities alongside human identities;

• centralized policy and execution controls;

• fewer redundant SaaS products;

• open tool and agent standards;

• continuous model and agent evaluation;

• edge AI inside physical environments;

• hybrid seat, usage and outcome pricing;

• and humans managing exceptions rather than routine orchestration.

The CIO’s mandate begins to change.

Historically, CIOs managed an application portfolio.

Increasingly, they may govern an:

Enterprise Execution Fabric

The organization’s core technology question stops being:

“Which applications do we own?”

and increasingly becomes:

“Through what governed infrastructure does intelligent work happen?”

And by 2036?

By 2036, enterprise software may look less like a collection of interfaces and more like an organizational operating system.

Policies define authority.

Systems of record preserve truth.

Models interpret information.

Context layers explain what organizational information means.

Agents plan and execute.

Humans set objectives, resolve ambiguity, manage relationships and remain accountable.

None of this requires artificial general intelligence.

Many enterprise processes are not intellectually extraordinary.

They are organizationally fragmented.

Consider:

• invoice reconciliation;

• employee onboarding;

• supplier procurement;

• scheduling;

• compliance evidence gathering;

• account renewal;

• service triage;

• inventory management;

• security investigation.

Much of the difficulty comes from moving information between systems and people.

AI agents can transform those processes long before machines possess generalized human-level intelligence.

And that changes the fundamental design question for enterprise software.

For decades, vendors asked:

“How easy is this application for a person to use?”

Increasingly they will also need to ask:

“How safely and effectively can machines use this system on behalf of people?”

That changes product design.

APIs become as important as screens.

Machine identities become as important as human accounts.

Agent permissions become as important as user roles.

Evaluations become as important as QA.

Auditability becomes as important as usability.

Semantic clarity becomes as important as interface design.

Software must increasingly be designed for two kinds of users:

humans and intelligent machines acting under human authority.

Five Principles for the Agentic Enterprise

The transition can be summarized in five principles.

Principle 1: The workflow becomes more important than the application.

Applications remain systems of record and execution.

But business value increasingly comes from coordinating work across them.

Principle 2: Context determines competence. Authority determines risk.

A model cannot act intelligently without organizational context.

But context alone is insufficient.

The enterprise must also define what the system is permitted to do.

Principle 3: Autonomy should be proportional to reversibility.

Low-risk, reversible actions can be highly automated.

High-risk or irreversible actions require stronger controls.

Principle 4: Models become interchangeable faster than organizational context.

The best model will change.

Prices will change.

Providers will change.

Enterprise semantics, permissions, workflows, history and trust are harder to replicate.

That is where durable advantage increasingly sits.

Principle 5: Measure business completion, not AI activity.

Tokens consumed are not value.

Prompts submitted are not value.

Agents created are not value.

Completed workflows, improved economics and reduced risk are value.

What Leaders Should Do Now

The answer is not to write a ten-year AI roadmap.

The technology is changing too quickly.

Instead, companies should redesign the foundations on which multiple generations of AI systems can operate.

1. Choose domains, not demos.

Do not start with:

“Where can we deploy an agent?”

Start with:

“Which end-to-end business domain has enough economic value to justify redesign?”

Prioritize processes connected directly to:

• revenue,

• cost,

• cycle time,

• customer experience,

• capital efficiency,

• or risk.

2. Redesign workflows before automating them.

Decompose the process.

For every step, decide:

• what AI should do;

• what deterministic software should do;

• what humans should do;

• what requires approval;

• what context is required;

• what can go wrong.

Do not automate a broken process simply because AI makes automation possible.

3. Build the control plane.

Standardize:

• machine identity,

• human identity,

• permissions,

• model access,

• tool access,

• approvals,

• logging,

• observability,

• evaluations,

• policy enforcement,

• exception handling,

• and cost controls.

This should become shared infrastructure rather than something rebuilt independently for every agent.

4. Fix contextual data.

The future does not belong simply to companies with more data.

It belongs to companies whose data can be used safely and meaningfully.

Create:

• authoritative definitions,

• lineage,

• permissions,

• semantic layers,

• business rules,

• metadata,

• ownership,

• and source-of-truth policies.

5. Architect for multiple models.

Avoid hard-wiring the enterprise operating model to one LLM.

Build routing and evaluation mechanisms that allow different models to compete for workloads.

Treat models as a changing portfolio.

6. Govern autonomy by risk.

Define explicit authority boundaries.

Tie those boundaries to:

• reversibility,

• financial exposure,

• safety,

• regulatory requirements,

• customer impact,

• and blast radius.

7. Consolidate selectively.

Do not consolidate because “all-in-one” sounds efficient.

Consolidate where fragmentation creates genuine context, governance or integration costs.

Preserve specialized systems where they deliver unique value.

8. Measure unit economics.

Track:

cost per successful workflow

rather than merely:

• AI usage,

• token consumption,

• prompts,

• seats,

• or agent counts.

Compare the system against the human baseline.

9. Preserve exit rights.

Integrated AI platforms will become deeply embedded in operations.

That increases the importance of:

• data export,

• documented APIs,

• model portability,

• workflow portability,

• auditable logs,

• permission export,

• and contractual exit rights.

Integration without portability can become dependency.

10. Stop Funding AI Use Cases

Perhaps the most important change is managerial.

Companies still tend to create lists of AI use cases:

• AI for sales.

• AI for finance.

• AI for HR.

• AI for customer support.

• AI for documents.

• AI for analytics.

That framing encourages pilots.

It does not necessarily produce transformation.

A better approach is to choose a business domain and redesign it from first principles.

For example:

Not:

“Add AI to accounts payable.”

But:

“Redesign invoice-to-payment so routine exceptions resolve automatically, ambiguous cases reach the correct human and the cost per successfully processed invoice falls materially.”

Not:

“Give salespeople an AI assistant.”

But:

“Redesign renewal management so risk identification, research, offer construction and follow-up preparation happen automatically while account strategy and customer negotiation remain human.”

Not:

“Deploy a support chatbot.”

But:

“Redesign issue-to-resolution so machines handle every reversible step they can reliably complete and humans focus on ambiguity, relationships and exceptions.”

That is the operating-model shift.

Stop funding isolated AI use cases. Fund redesigned business domains.

The Question That Matters

The wrong question for enterprise leaders is:

“Where can we add AI?”

That question assumes the current organization is basically correct and intelligence should simply be inserted into it.

The better question is:

“If intelligent, low-cost digital labor could understand our information and safely use our software, how would we redesign this process from first principles?”

That question produces very different answers.

For thirty years, enterprise software largely forced people to learn how computers organize work.

Employees learned which application to open.

Which screen contained the information.

Which fields had to be updated.

Which system contained the truth.

Which steps had to happen next.

The next generation begins to reverse that relationship.

Software increasingly learns the enterprise’s:

• language,

• data,

• policies,

• permissions,

• processes,

• goals,

• relationships,

• and exceptions.

And then it acts within those boundaries.

The screen becomes less important.

The context becomes more important.

The individual application becomes less important.

The workflow becomes more important.

The model becomes more interchangeable.

The governance layer becomes more strategic.

The seat becomes less important.

The business outcome becomes more important.

Enterprise software is not disappearing.

It is becoming something deeper.

It is becoming the infrastructure through which intelligent work happens.

The companies that win will not simply possess the most powerful AI models.

They will be the companies that can convert organizational knowledge into execution without losing control.

In other words:

The ultimate advantage will belong to those who can turn trusted context into governed action faster than everyone else.

Sources and References

Anthropic (2024), Introducing the Model Context Protocol — https://www.anthropic.com/news/model-context-protocol

Anthropic (2025), Donating the Model Context Protocol and establishing the Agentic AI Foundation — https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation

Autio et al./NIST (2024), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Brynjolfsson, Li & Raymond, Generative AI at Work, NBER Working Paper 31161 — https://www.nber.org/papers/w31161

Dell’Acqua et al. (2025), The Cybernetic Teammate — https://www.hbs.edu/faculty/Pages/item.aspx?num=67197

European Commission (2026), AI Act: Regulatory Framework — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

European Commission (2026), AI Act Implementation Timeline — https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act

European Commission (2026), AI Act FAQ — https://ai-act-service-desk.ec.europa.eu/en/faq

Gartner (June 2025), Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

Gartner (August 2025), 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 — https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

Gartner (January 2026), Worldwide AI Spending Will Total $2.5 Trillion in 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026

Gartner (April 2026), 2026 Hype Cycle for Agentic AI — https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai

Gartner (May 2026), Autonomous Business and AI Layoffs May Create Budget Room but Do Not Deliver Returns — https://www.gartner.com/en/newsroom/press-releases/2026-05-05-gartner-says-autonomous-business-and-artificial-intelligence-layoffs-may-create-budget-room-but-do-not-deliver-returns

Gartner (May 2026), Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure — https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure

Gartner (July 2026), Worldwide AI Platforms and Models Market to Grow 63% in 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-07-20-gartner-forecasts-worldwide-ai-platforms-and-models-market-to-grow-63-percent-in-2026

GitHub (2024/2025), Does GitHub Copilot Improve Code Quality? — https://resources.github.com/videos/write-more-code-by-writing-less-code-with-github-copilot/

Harvard Business School AI Institute (2025), The Cybernetic Teammate — https://aiinstitute.hbs.edu/the-cybernetic-teammate-how-ai-is-reshaping-collaboration-and-expertise-in-the-workplace/

IDC (2024), Worldwide AI and Generative AI Spending — Industry Outlook — https://www.idc.com/resource-center/blog/idcs-worldwide-ai-and-generative-ai-spending-industry-outlook/

IDC (2025), European AI Spending to Reach $144 Billion by 2028 — https://my.idc.com/getdoc.jsp?containerId=prEUR253256125

IDC (2025), Asia/Pacific AI Spending to Reach $175 Billion by 2028 — https://my.idc.com/getdoc.jsp?containerId=prAP53348125

McKinsey (2025), The State of AI: How Organizations Are Rewiring to Capture Value — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-how-organizations-are-rewiring-to-capture-value

McKinsey (2025), The State of AI: Global Survey 2025 — https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

McKinsey (2025), The New Economics of Enterprise Technology in an AI World — https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-new-economics-of-enterprise-technology-in-an-ai-world

McKinsey (2026), Designing an End-to-End Technology Workforce for the AI-First Era — https://www.mckinsey.com/capabilities/mckinsey-technology/our-insights/designing-an-end-to-end-technology-workforce-for-the-ai-first-era

McKinsey (2026), The Operating Model Advantage — https://www.mckinsey.com/industries/industrials/our-insights/the-operating-model-advantage-why-ai-winners-are-rewiring-their-organizations

McKinsey (2026), What AWS’s Agentic Journey Can Teach CEOs About Rewiring for AI — https://www.mckinsey.com/about-us/new-at-mckinsey-blog/lessons-from-our-alliances-what-aws-agentic-journey-can-teach-ceo-about-rewiring-for-ai

McMahan et al. (2017), Communication-Efficient Learning of Deep Networks from Decentralized Data — https://arxiv.org/abs/1602.05629

Microsoft (2026), Dynamics 365 Business Central Pricing — https://www.microsoft.com/en-us/dynamics-365/products/business-central/pricing

Microsoft (2026), Intelligent Factories — https://learn.microsoft.com/en-us/industry/manufacturing/enable-intelligent-factories

Microsoft (2025), Azure AI Models on Siemens Industrial Edge — https://learn.microsoft.com/en-us/industry/manufacturing/architecture/azure-ai-siemens-industrial-edge

NIST (2023), AI Risk Management Framework 1.0 — https://www.nist.gov/itl/ai-risk-management-framework

Odoo (2026), Pricing — https://www.odoo.com/pricing

OpenAI (2026), Morgan Stanley Uses AI Evals to Shape the Future of Financial Services — https://openai.com/index/morgan-stanley/

Oracle (2026), Oracle AI for Fusion Applications — https://www.oracle.com/uk/applications/fusion-ai/

Oracle (2026), New Fusion Agentic Applications — https://blogs.oracle.com/fusioninsider/new-fusion-agentic-applications-details-and-demos

Reuters (2024), Klarna Using GenAI to Cut Marketing Costs by $10 Million Annually — https://www.reuters.com/technology/klarna-using-genai-cut-marketing-costs-by-10-mln-annually-2024-05-28/

Reuters (2024), Klarna Says AI Chatbots Help Shrink Headcount — https://www.reuters.com/technology/artificial-intelligence/swedens-klarna-says-ai-chatbots-help-shrink-headcount-2024-08-27/

Reuters (2025), Klarna Shifts AI Focus From Cost Cuts to Growth — https://www.reuters.com/business/swedens-klarna-shifts-ai-focus-cost-cuts-growth-2025-09-10/

Salesforce (2026), Sales Cloud Pricing — https://www.salesforce.com/uk/sales/pricing/

SAP (2026), Joule Agents and Joule Assistants — https://www.sap.com/products/artificial-intelligence/ai-agents.html

SAP (2026), The Autonomous Enterprise — https://www.sap.com/products/business-suite.html

ServiceNow (2026), ServiceNow AI Platform — https://www.servicenow.com/platform.html

ServiceNow (2026), Real-Time Data Foundation for Autonomous AI — https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-launches-the-real-time-data-foundation-that-puts-autonomous-AI-to-work-across-the-enterprise/default.aspx

T1U (2026), Mission, Journey & AI Business Operating System — https://www.t1u.com/about

T1U (2026), NeoMind AI — https://www.t1u.com/neomind-ai

T1U (2026), Pricing — https://www.t1u.com/pricing

T1U (2026), Security — https://www.t1u.com/security

T1U (2026), Privacy Center — https://www.t1u.com/privacy-center

Zoho (2026), Zoho One Pricing — https://www.zoho.com/one/pricing/

On January 27, 2025, Nvidia lost roughly $589 billion in market value in a single trading day, at the time the largest one-day destruction of market capitalization in U.S. stock-market history. DeepSeek had reported that the final training run for its V3 model consumed 2.788 million Nvidia H800 GPU-hours, corresponding to an estimated compute cost of around $5.6 million under commonly used rental assumptions.

That figure was repeatedly misinterpreted as the total cost of building DeepSeek’s model. It was not. It excluded research salaries, failed experiments, infrastructure, data, previous models and the accumulated engineering knowledge necessary to make the final training run work.

Yet the broader lesson survived the correction.

DeepSeek had demonstrated something the AI industry had not fully priced in:

The relationship between intelligence and capital expenditure was not fixed.

A company operating with fewer frontier chips and dramatically less capital than America’s leading AI laboratories had still produced a model capable enough to shake global markets.

That was the first DeepSeek shock.

It was about cost.

The next one may be about power.

DeepSeek is now reportedly preparing significant price increases for some AI services. At the same time, the company has continued using aggressive low prices elsewhere in its product lineup.

That apparent contradiction is precisely what makes the story interesting.

The question is no longer merely:

Can China produce sophisticated AI more cheaply than the United States?

It is becoming:

Can Chinese AI companies charge more — and keep their customers?

Because if they can, DeepSeek will have moved beyond demonstrating efficiency.

It will have begun demonstrating pricing power.

And that would tell us something far more important about the future economics of artificial intelligence.

Cheap AI was never as simple as it looked

At launch, DeepSeek-R1 was priced at approximately $0.55 per million uncached input tokens and $2.19 per million output tokens. OpenAI’s o1 pricing was around $15 and $60, respectively.

That made DeepSeek roughly 27 times cheaper on the headline API prices.

The temptation was to conclude that Chinese companies had simply discovered how to produce AI at one twenty-seventh of the American cost.

Reality was more complicated.

A product can be cheap because it is genuinely more efficient, but it can also be cheap because a company is willing to accept lower margins, because it is subsidizing adoption, because it wants to attract developers, or because winning an ecosystem is considered more valuable than maximizing short-term revenue.

Those explanations are not mutually exclusive.

This is why rising prices at DeepSeek should not automatically be interpreted as evidence that its efficiency advantage has disappeared. Some of the discount embedded in the original pricing may simply have been strategic. The company may also be facing stronger demand, tighter inference capacity or higher costs from increasingly compute-intensive reasoning models.

But there is another possibility that is much more interesting.

DeepSeek may believe it no longer needs to be quite so cheap.

That changes the question investors should be asking. In 2025, the question was: Can China offer competitive AI at a fraction of American prices?

The more important question now is: Can Chinese AI companies raise those prices and keep their customers?

If they can, that is not necessarily a sign that the disruption is ending. It may be evidence that the disruption is maturing.

A competitor that can win only by dramatically underpricing an incumbent has demonstrated efficiency or a willingness to subsidize. A competitor that can later raise prices without losing demand has begun to demonstrate something much more valuable: pricing power.

But there is an even broader point here. The AI market may not converge toward a single “price of intelligence” at all. As models become cheaper to operate but capable of performing more valuable work, pricing is likely to fragment. Commodity inference can remain brutally price-sensitive, while advanced agents, regulated deployments and mission-critical workloads command premiums for reliability, reasoning, latency, security or sovereignty.

The emerging economics of AI pricing

Falling costs do not necessarily produce uniformly falling prices. Greater adoption, more demanding workloads and different customer needs can instead push the market toward increasingly segmented pricing.

This is why DeepSeek’s rising prices do not necessarily contradict the broader trend toward cheaper AI. The underlying cost of producing intelligence can continue falling even as providers become more selective about how they price it. Small models and high-volume inference may become increasingly commoditized, while frontier reasoning, enterprise guarantees, sovereign deployment and specialized workloads capture substantially higher margins.

In other words, cheap AI may not be disappearing. The market for intelligence may simply be splitting into cheap intelligence and expensive intelligence.

DeepSeek changed the meaning of compute advantage

DeepSeek’s deeper contribution was not simply cheap inference. It demonstrated how aggressively engineering could improve the productivity of expensive computing resources.

DeepSeek-V3 contained 671 billion parameters, but only around 37 billion were active for each token because of its mixture-of-experts architecture. The company combined this with techniques including Multi-head Latent Attention, FP8 training and extensive optimization of communication between GPUs.

DeepSeek did not prove that scaling was obsolete. It demonstrated that scaling could be made far more efficient.

That changes the strategic question from who possesses the most GPUs to who produces the most useful intelligence from each GPU.

This distinction matters enormously for the United States and China because America retains an extraordinary advantage in capital, advanced semiconductor design, cloud infrastructure and installed compute. China, meanwhile, has been forced to operate under tighter semiconductor constraints and with substantially less private investment.

The American advantage is therefore largely one of abundance.

China’s competitive pressure is increasingly one of optimization.

Scarcity is obviously not preferable to abundance, but constraints can change what engineers optimize for. A laboratory capable of solving a performance problem partly by adding another cluster of GPUs faces a different incentive structure from one that cannot easily obtain those GPUs.

This is one reason DeepSeek complicated the logic of U.S. semiconductor restrictions. Export controls can impose very real costs on Chinese companies without freezing Chinese innovation. Companies adapt, substitute and optimize, sometimes developing technologies specifically designed to make constrained resources more productive.

The lesson is not that compute no longer matters. It is that compute alone is not a permanent moat.

Cheap intelligence may actually increase the demand for compute

There is another paradox at the center of this story.

AI was already becoming dramatically cheaper before DeepSeek arrived. Stanford’s 2025 AI Index estimated that the cost of querying a model performing at roughly GPT-3.5 level on MMLU fell from around $20 per million tokens in November 2022 to approximately $0.07 by October 2024, a decline of more than 280-fold.

It would be easy to look at that curve and conclude that falling inference costs must eventually reduce the need for enormous AI infrastructure.

Economic history suggests the opposite may happen.

William Stanley Jevons famously observed that improvements in the efficiency with which coal was used did not necessarily reduce society’s consumption of coal (Jevons paradox). Greater efficiency made coal economical for more applications, sometimes increasing overall demand.

AI could experience the same phenomenon.

If an AI task becomes ten times cheaper, a company may not simply reduce its AI expenditure by 90 percent. It may deploy AI across ten business functions instead of one. Developers may run agents continuously. Banks may analyze every transaction. Factories may add intelligent systems to individual machines, while robots and autonomous devices could eventually perform inference almost constantly.

In that world, the economic chain becomes:

greater efficiency → lower prices → more applications → more inference → potentially more total compute.

This explains how DeepSeek can simultaneously challenge the economics of individual AI models while strengthening the case for enormous long-term investment in AI infrastructure.

The unit price of intelligence can collapse while the total market becomes vastly larger.

We have already seen this pattern in computing, bandwidth and storage.

AI may be next.

The most important chart in the AI race is not about benchmarks

The geopolitical implications become clearer when we compare investment with capability.

According to Stanford’s 2026 AI Index, private AI investment in the United States reached approximately $285.9 billion in 2025, compared with around $12.4 billion in China.

That is more than a 23-to-one advantage.

The comparison is imperfect because private-investment data do not fully capture China’s state-backed capital and industrial-policy mechanisms. Even so, the difference is enormous.

If AI leadership were primarily determined by private capital deployed, the technological outcome should be overwhelmingly one-sided.

It is not.

Stanford reported that American institutions produced 59 notable AI models in 2025, compared with 35 from China. More strikingly, its 2026 assessment concluded that the U.S.–China model-performance gap had effectively closed, with models from the two countries trading the lead at different points.

That discrepancy deserves much more attention than it receives.

America is spending vastly more private capital and possesses substantially greater compute infrastructure, yet China has compressed the capability gap far more rapidly than those input differences would predict.

This does not mean the American advantage is imaginary. The United States remains extraordinarily strong in frontier research, semiconductor design, cloud computing, capital markets and AI infrastructure.

But it suggests something uncomfortable:

capital intensity is not the same thing as technological defensibility.

A 23-to-one investment advantage has not produced anything resembling a 23-to-one capability advantage.

There is no longer one AI race

The phrase “AI race” itself is becoming misleading because the United States and China are competing across several different dimensions simultaneously.

There is a frontier race to build the world’s most capable models; an efficiency race to produce comparable intelligence using less compute and energy; a distribution race to become the technological foundation used by developers, businesses and governments; and an industrialization race to embed AI into manufacturing, robotics, logistics, science, medicine, energy and defense.

America may lead comfortably in one dimension while China closes the gap or gains ground in another.

DeepSeek is particularly important in the distribution race because its open-model strategy makes its technology easier to download, modify, fine-tune and deploy outside proprietary American cloud ecosystems.

That matters geopolitically.

Countries and companies outside the United States and China may not ask simply which nation has the highest benchmark score. They may ask which ecosystem gives them the greatest capability at the lowest cost and with the least technological dependence.

Open models therefore compete not only through performance but through diffusion.

And diffusion can eventually become power.

This is another reason DeepSeek’s pricing strategy is so interesting. Extremely cheap AI can accelerate distribution. Distribution attracts developers. Developers create applications and dependencies. Once an ecosystem becomes sufficiently embedded, the company behind it may have greater freedom to monetize.

Seen through that lens, the transition from extremely low prices toward higher prices is not necessarily a retreat.

It could represent a transition from buying distribution with price to monetizing distribution through price.

So is cheap AI ending?

Probably not.

What may be ending is the simplistic assumption that the cheapest AI provider will always remain the cheapest, or that every improvement in efficiency will be passed directly to customers.

AI capability is likely to continue becoming cheaper over time, but the economics will become more complex. Providers with strong ecosystems may retain a larger portion of efficiency gains as margin. Premium reasoning models may become more expensive even while older capabilities become commodities. Meanwhile, lower costs may create so many new applications that aggregate spending on inference continues rising.

That is why DeepSeek’s pricing changes matter.

The original DeepSeek shock suggested that China could produce sophisticated AI surprisingly efficiently despite having fewer frontier computing resources than the United States.

The next test is whether Chinese companies can convert that engineering advantage into something more durable: developer loyalty, distribution, ecosystem depth and eventually pricing power.

If they can, the U.S.–China competition becomes considerably more complicated.

America still possesses extraordinary strengths in chips, capital, cloud infrastructure and frontier research. China possesses extraordinary strengths in manufacturing, supply chains and increasingly in efficient and openly distributed AI models.

The ultimate competition will therefore not be decided by a single benchmark or by who spends the most money training the next model.

It will be decided by who can connect the entire stack — chips, energy, models, software, developers, manufacturing, robotics and capital — and turn rapidly improving intelligence into economic and industrial advantage.

DeepSeek did not prove that scaling was dead. It showed that inefficient scaling was vulnerable.

It did not prove that China had won the AI race. It showed that America’s lead could not safely be measured simply in dollars spent or GPUs accumulated.

And rising prices do not necessarily mean that the era of cheap AI is over.

They may mean something more interesting: that the company which disrupted the market by making intelligence extraordinarily cheap is beginning to test how valuable that intelligence has become.

The first AI price war was about who could make intelligence cheaper.

The next one may be about who has enough power to stop giving it away.

References

Alibaba Group. (2025, February 20). Alibaba to scale up AI investments in three areas over the next three years. https://www.alibabagroup.com/en-US/document-1829422257118642176

Alibaba Group. (2026, May 20). Letter from our chairman and our CEO. https://www.alibabagroup.com/en-US/document-1993785120221298688

Alibaba Cloud. (2026). Model Studio model pricing. https://help.aliyun.com/zh/model-studio/model-pricing

Alibaba Cloud. (2026, August 3). Qwen3.7-Max model information. https://help.aliyun.com/zh/model-studio/qwen3-7-max

Anthropic. (2024, March 4). Introducing the next generation of Claude. https://www.anthropic.com/news/claude-3-family

Anthropic. (2026). Claude Platform pricing. https://platform.claude.com/docs/en/about-claude/pricing

Anthropic. (2026, April 20). Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute. https://www.anthropic.com/news/anthropic-amazon-compute

Bureau of Industry and Security, U.S. Department of Commerce. (2026, January 13). Department of Commerce revises license review policy for semiconductors exported to China. https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china

Bureau of Industry and Security, U.S. Department of Commerce. (n.d.). Export controls imposed on advanced computing and semiconductor manufacturing items to the People’s Republic of China. https://www.bis.gov/press-release/bis-updated-public-information-page-export-controls-imposed-advanced-computing-semiconductor

Cyberspace Administration of China. (2023, July 13). Interim Measures for the Management of Generative Artificial Intelligence Services [生成式人工智能服务管理暂行办法]. https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm

DeepSeek. (2024, August 2). DeepSeek API introduces Context Caching on Disk, cutting prices by an order of magnitude. https://api-docs.deepseek.com/news/news0802/

DeepSeek. (2025, August 21). DeepSeek-V3.1 release. https://api-docs.deepseek.com/news/news250821/

DeepSeek. (2025, September 29). Introducing DeepSeek-V3.2-Exp. https://api-docs.deepseek.com/news/news250929/

DeepSeek. (2026, April 24). DeepSeek V4 Preview release. https://api-docs.deepseek.com/news/news260424/

DeepSeek. (2026). Models & pricing. Retrieved August 12, 2026, from https://api-docs.deepseek.com/quick_start/pricing/

DeepSeek-AI et al. (2024). DeepSeek-V3 technical report. arXiv. https://arxiv.org/abs/2412.19437

Google. (2026). Gemini Developer API pricing. https://ai.google.dev/gemini-api/docs/pricing

OpenAI. (2023, March 14). GPT-4. https://openai.com/index/gpt-4-research/

OpenAI. (2024, May 13). GPT-4o API announcement. https://community.openai.com/t/announcing-gpt-4o-in-the-api/744700

OpenAI. (2025, November 5). 1 million business customers: The fastest-growing business platform in history. https://openai.com/index/1-million-businesses-putting-ai-to-work/

OpenAI. (2026, April 8). The next phase of enterprise AI. https://openai.com/index/next-phase-of-enterprise-ai/

OpenAI. (2026). Compare models: GPT-5.6 Sol, Terra and Luna. https://developers.openai.com/api/docs/models/compare

Reuters. (2024, May 21). Chinese tech giants slash prices of language models used to power AI chatbots. https://www.reuters.com/technology/chinese-tech-giants-slash-prices-language-models-used-power-ai-chatbots-2024-05-21/

Reuters. (2025, January 27). DeepSeek sparks AI stock selloff; Nvidia posts record market-value loss. https://www.reuters.com/technology/chinas-deepseek-sets-off-ai-market-rout-2025-01-27/

Reuters. (2025, February 26). DeepSeek cuts off-peak pricing for developers by up to 75%. https://www.reuters.com/technology/chinas-deepseek-cuts-off-peak-pricing-by-up-75-2025-02-26/

Reuters. (2026, January 6). Governments, regulators increase scrutiny of DeepSeek. https://www.reuters.com/legal/litigation/governments-regulators-increase-scrutiny-deepseek-2026-01-06/

Reuters. (2026, February 12). A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models. https://www.reuters.com/world/china/year-deepseek-shock-get-set-flurry-low-cost-chinese-ai-models-2026-02-12/

Reuters. (2026, March 31). Zhipu accelerates pivot to domestic chips amid AI boom in China. https://www.reuters.com/world/asia-pacific/chinas-zhipu-posts-132-rise-annual-revenue-ai-boom-2026-03-31/

Reuters. (2026, April 29). Big Chinese tech firms scramble to secure Huawei AI chips after DeepSeek V4 launch, sources say. https://www.reuters.com/world/china/big-chinese-tech-firms-scramble-secure-huawei-ai-chips-after-deepseek-v4-launch-2026-04-29/

Reuters. (2026, May 23). China’s DeepSeek to make permanent 75% price cut on flagship V4-Pro AI model. https://www.reuters.com/world/china/chinas-deepseek-make-permanent-75-price-cut-flagship-v4pro-ai-model-2026-05-23/

Reuters. (2026, June 25). After Anthropic shutdown, China’s Z.ai closes frontier gap as it plans dual listing. https://www.reuters.com/world/asia-pacific/after-anthropic-shutdown-chinas-zai-closes-frontier-gap-it-plans-dual-listing-2026-06-25/

Reuters. (2026, August 3). DeepSeek’s new AI model is by far the cheapest of well-known models to run, research firm says. https://www.reuters.com/business/retail-consumer/deepseeks-new-ai-model-is-by-far-cheapest-well-known-models-run-research-firm-2026-08-03/

Reuters. (2026, August 6). DeepSeek resumes funding round seeking nearly $8 billion, Bloomberg News reports. https://www.reuters.com/world/asia-pacific/deepseek-resumes-funding-round-seeking-nearly-8-billion-bloomberg-news-reports-2026-08-06/

Savov, V., & Rai, S. (2026, August 6). DeepSeek plans ‘significant’ price increase for AI services. Bloomberg News, republished by Moneyweb. https://www.moneyweb.co.za/news/ai/deepseek-plans-significant-price-increase-for-ai-services/

Bloomberg. (2026, August 7). DeepSeek’s pricing moves have a whole industry watching. Bloomberg. https://www.bloomberg.com/news/newsletters/2026-08-07/deepseek-s-plan-to-raise-prices-have-a-whole-industry-watching

Z.AI. (2026). GLM model pricing. https://docs.z.ai/guides/overview/pricing

Z.AI. (2026). GLM-5.2 overview. https://docs.z.ai/guides/llm/glm-5.2

Imagine a world in which nearly everyone has access to an intelligence more capable than the best doctor, tutor, programmer, lawyer, scientist, or strategist they could ever afford.

It helps a child learn in her own language. It gives a small business capabilities once reserved for multinational corporations. It accelerates medical discoveries, translates knowledge across borders, and allows ordinary people to create, organize, and solve problems at extraordinary speed.

Now imagine that the same intelligence is controlled by a handful of institutions.

They decide what it may say, what it may know, whose interests it serves, which countries may access it, how much it costs, what happens to the data it collects, which workers it replaces, and how the wealth it creates is distributed.

The intelligence might be abundant.

Power would not be.

That is the contradiction at the center of the superintelligence debate.

We keep discussing advanced AI as though it were primarily an engineering question.

How capable will it become?

When will it arrive?

Will it reason, deceive, invent, manipulate, or escape?

Will it cure disease, end scarcity, displace workers, or threaten civilization?

Those questions matter. But beneath them lies an even more consequential political question:

Who will control intelligence itself?

One camp promises personal superintelligence in everyone’s glasses, phones, homes, and pockets. Another promises cognitive capability made as cheap and universal as electricity. A third sees a geopolitical and industrial race that must be won through extraordinary national mobilization. A fourth argues that frontier systems should remain behind increasingly sophisticated security controls. A fifth wants development halted until superintelligence can be shown to be safe, controllable, and democratically legitimate.

These positions are usually presented as technical philosophies.

They are also competing constitutions for the future.

They ask who will own machines that may outperform humanity’s finest scientists, engineers, entrepreneurs, strategists, and institutions.

They ask who will decide what those machines may do, whose values they will follow, who will receive their economic gains, and who will bear the consequences when they fail.

My answer is simple:

Open at the edge.
Guarded at the frontier.
Shared at the upside.
Human at the center.

Ordinary, broadly beneficial AI should be affordable, interoperable, competitive, multilingual, and widely available.

Capabilities that could cause catastrophic or irreversible harm should require independent evaluation, controlled deployment, strong security, traceability, and public authorization.

The wealth generated from public science, infrastructure, energy, knowledge, and data should be shared with the societies that made it possible.

And every AI system should ultimately be judged not only by whether it makes machines more capable, but by whether it leaves human beings more free.

No corporation should become the sovereign of human intelligence.

No government should obtain an unaccountable monopoly over it.

No small group of engineers should be able to make civilization-level decisions in private.

And no downloadable file should make catastrophic capabilities permanently available to anyone with an internet connection.

Superintelligence, if it arrives, must have no single master.

It must operate inside a system of distributed access, democratic accountability, verifiable safety, pluralistic governance, and shared ownership.

That is the superintelligence compact humanity now needs.

The five futures competing to govern us

The current debate contains at least five competing visions of the future.

Each sees something clearly.

Each also contains a serious danger.

1. Personal intelligence through corporate platforms

Mark Zuckerberg’s vision of personal superintelligence promises to put increasingly capable AI directly into people’s hands, allowing individuals to pursue their own goals rather than merely consuming the products of centrally automated systems.

In this world, context-aware glasses, phones, and other personal devices may become the primary interface between human beings and machine intelligence.

The promise is real.

Intelligence should not become a priesthood available only to governments, multinational corporations, and wealthy institutions. Broad access could lower barriers to education, entrepreneurship, scientific research, translation, creativity, and political participation.

But placing AI in a person’s pocket does not necessarily place power in that person’s hands.

A system may feel intimate while its objectives, updates, infrastructure, interface, and data remain controlled by a corporation.

A machine that “knows us deeply” could empower individuals while also creating the most intimate surveillance architecture ever constructed.

The decisive question is not whether AI is physically close to us.

It is whether the person using it possesses meaningful rights over the model, the data, the objectives, the interface, and the institution behind it.

2. Abundant intelligence through private infrastructure

A second vision promises intelligence made widely available and inexpensive through privately built platforms.

OpenAI and others increasingly speak in the language of decentralization, broad access, democratic participation, and universal prosperity. The idea of a “right to AI” recognizes that capable machine intelligence may become foundational infrastructure — something people need to participate fully in education, work, commerce, and civic life.

That vision also contains an important truth.

But access alone is not sovereignty.

A free chatbot running on somebody else’s servers is not decentralized power.

A system can be available to billions while remaining governed entirely by one company.

An open-weight model may be downloadable while depending on chips, clouds, energy, and distribution networks controlled by a handful of firms.

A product can be cheap while the capacity to create it remains extraordinarily concentrated.

Access is not ownership.

Affordability is not sovereignty.

Abundant intelligence could create abundance in outputs while producing extreme scarcity in power.

3. National mobilization

A third vision, articulated most forcefully by Leopold Aschenbrenner in Situational Awareness, treats advanced AI not as an ordinary consumer technology, but as a strategic industrial system capable of reshaping the global balance of power.

In this view, the decisive assets are not merely algorithms. They are semiconductor fabrication, advanced chips, data centers, energy systems, cybersecurity, specialized talent, and enormous amounts of capital. Governments therefore cannot afford to treat frontier AI as just another startup market. They may need to secure model weights, protect chip supply chains, expand electrical infrastructure, coordinate research, and prepare for a competition that could shape military and economic power for generations.

Aschenbrenner sees something the consumer narrative often misses: intelligence is becoming industrial capacity, and industrial capacity is geopolitical power. A country that neglects the foundations of AI could become dependent on private monopolies or strategic rivals.

But the language of a race to be won creates dangers of its own. Arms-race logic rewards secrecy, compresses decision time, discourages safety disclosures, and treats restraint as weakness. It makes extraordinary state powers easier to justify and harder to remove.

A project created to defend democracy could eventually subordinate democracy to the project.

4. Controlled access through frontier laboratories

A fourth vision, associated most clearly with Dario Amodei and Anthropic’s Responsible Scaling Policy, argues that openness and deployment should depend on capability.

Amodei’s Machines of Loving Grace presents advanced AI as a potential accelerator of medicine, science, economic development, and human flourishing. But Anthropic’s safety framework also recognizes that increasingly capable systems may cross thresholds at which ordinary product safeguards are no longer sufficient.

Under this approach, lower-risk models can be released broadly, while systems capable of materially assisting biological-weapons development, automating sophisticated cyberattacks, manipulating critical infrastructure, or accelerating their own research should face progressively stronger controls.

This principle is reasonable. A system capable of causing catastrophic harm is not equivalent to an ordinary software tool. Capability-sensitive safeguards are not inherently censorship.

But private laboratories cannot be permitted to serve simultaneously as inventor, evaluator, regulator, and court of appeal.

A company may publish a thoughtful safety framework. It is still regulating itself, under competitive and commercial pressure. Responsible scaling should therefore become a matter of public law, independent evaluation, and democratic oversight — not private discretion alone.

5. Prohibition until legitimacy exists

A fifth vision argues that superintelligence should be prohibited until there is broad scientific agreement that it can be developed safely and strong public support for proceeding.

The Future of Life Institute’s Statement on Superintelligence represents this position. As of August 2026, its public website displayed more than 71,000 signatures.

This camp is right about consent.

No private company, government agency, or technical elite — however brilliant or well financed — automatically possesses the moral authority to transform the conditions of human agency.

Technological capability does not create a democratic mandate.

But prohibition is not a complete governance system.

A meaningful moratorium would need measurable thresholds, monitoring, international verification, rules distinguishing research from deployment, mechanisms for dealing with open models already in circulation, and a clear standard for deciding whether development could resume.

A pause without verification becomes symbolic.

A pause observed only by democracies could shift development toward less accountable regimes.

Each of these five visions contains part of the answer.

None is sufficient on its own.

The central mistake is the demand that humanity choose one value and sacrifice the others.

Access without safety is reckless.

Safety without democracy is technocracy.

State capacity without civil liberty is authoritarianism.

Innovation without shared ownership is oligarchy.

A pause without verification is symbolism.

The challenge is not to select one camp.

It is to build institutions that preserve what each camp sees clearly while constraining what each fails to see.

The facts are outrunning the slogans

“AGI” and “superintelligence” remain contested terms.

There is no universally accepted test for the moment a system becomes generally intelligent, nor a settled timeline for when machines might outperform the best humans across almost every cognitive domain.

But uncertainty about the definition should not be confused with a lack of progress.

The 2026 International AI Safety Report, developed with more than 100 experts and an advisory process involving nominees from more than 30 countries and international organizations, concluded that general-purpose AI capabilities were improving faster than many experts had expected.

Systems have demonstrated gold-medal-level performance on International Mathematical Olympiad problems, major gains in coding and autonomous operation, and growing usefulness in real-world cyber activity.

Their capabilities, however, remain jagged.

A model may solve a problem that defeats almost every human being while failing at something apparently simple.

AI is neither “just autocomplete” nor a dependable digital god.

It is an expanding archipelago of superhuman abilities separated by surprising seas of incompetence.

METR’s research provides one measure of this progress. Its “task-completion horizon” estimates the length of software, machine-learning, and cybersecurity tasks that an AI system can complete at a specified level of reliability, expressed in equivalent human working time.

Historical measurements found that the 50-percent-reliability horizon approximately doubled every seven months over several years.

But METR also warns that these benchmarks cover a limited class of well-specified technical work, that capability differs sharply across domains, and that longer-duration estimates remain uncertain.

The combination of exponential-looking improvement and persistent unreliability is the defining technical fact of the present era.

The frontier is becoming more powerful — and more concentrated

The physical scale behind advanced AI is becoming extraordinary.

Stanford’s 2026 AI Index estimated that global AI compute capacity reached the equivalent of 17.1 million H100 accelerators, after growing approximately 3.3-fold per year since 2022.

Industry produced more than 90 percent of notable AI models in 2025.

At the same time, the most capable systems became among the least transparent. Leading developers increasingly disclosed less about their parameter counts, training datasets, training code, and training duration.

Epoch AI estimates that frontier-language-model training compute has increased approximately fivefold annually since 2020, while training costs have risen around 3.5-fold per year.

It also estimates that just five hyperscalers control more than two-thirds of global AI compute.

A typical gigawatt-scale AI data center may require roughly $38 billion in upfront capital.

These numbers reveal a distinction that optimistic narratives about democratization often obscure.

Even if the price of using intelligence falls dramatically, the capacity to create the frontier may remain concentrated in a very small number of firms and states.

A free interface does not eliminate the power of the institution controlling the servers behind it.

A downloadable model does not decentralize the chip supply chain.

Cheap intelligence can still be governed by expensive and highly concentrated infrastructure.

AI infrastructure has public costs

The infrastructure race extends far beyond Silicon Valley.

Data centers consume electricity, water, land, grid capacity, and public resources. Communities may face higher energy prices, weakened grid reliability, water pressure, or opaque tax concessions while much of the resulting economic value flows elsewhere.

The International Energy Agency reported that data-center electricity consumption rose 17 percent in 2025, while demand from AI-focused data centers rose by approximately 50 percent.

Its central projection places global data-center electricity consumption at roughly 950 terawatt-hours by 2030, up from approximately 485 terawatt-hours in 2025.

AI-focused data-center consumption could triple over that period.

Grid constraints may delay roughly one-fifth of planned global data-center capacity.

Infrastructure policy is no longer peripheral to AI policy.

It is AI policy.

A society cannot claim that intelligence is being democratized if communities subsidize the infrastructure while private actors capture nearly all of the upside.

The productivity story is more complicated than either side admits

AI is often presented as an automatic engine of productivity.

The evidence is far less uniform.

The International Monetary Fund estimates that nearly 40 percent of global employment is exposed to AI. In advanced economies, the figure rises to approximately 60 percent.

Exposure does not mean elimination.

Some workers will be complemented, made more productive, or given access to expertise they previously lacked.

Others may experience lower demand, weaker bargaining power, reduced wages, restructuring, or displacement.

Early productivity research demonstrates why sweeping claims are premature.

A study of more than 5,000 customer-support agents found that an AI assistant increased productivity by approximately 14 to 15 percent, with much larger benefits among less experienced workers.

A controlled GitHub Copilot experiment found that participants completed one defined coding task 55.8 percent faster.

Yet a later randomized METR study involving experienced open-source developers working on their own repositories found that early-2025 AI tools made participants approximately 19 percent slower.

The developers had expected the tools to help.

They did not.

AI therefore has no single productivity effect.

It can transfer elite knowledge to novices.

It can accelerate clearly bounded tasks.

It can create capabilities that did not previously exist.

It can also introduce verification, correction, and coordination costs that overwhelm its benefits.

The effect depends on the task, the worker, the model, the workflow, and the institution controlling deployment.

Anyone promising an automatic productivity miracle is claiming a certainty the evidence does not support.

The upside and the danger are both real

The potential benefits are extraordinary.

AlphaFold’s database contains more than 200 million predicted protein structures and is used by millions of researchers across more than 190 countries.

It is a powerful demonstration of what can happen when advanced AI is directed toward a clearly defined scientific bottleneck and its outputs are broadly shared.

The risks are also increasingly visible.

Stanford recorded 362 documented AI incidents in its latest reporting period, up from 233 in 2024.

The International AI Safety Report found that safety testing is becoming more difficult as systems learn to distinguish evaluations from real-world deployment, exploit weaknesses in tests, and acquire capabilities that existing benchmarks fail to predict reliably.

Although 12 companies published or updated frontier-safety frameworks during 2025, most of the relevant governance remained voluntary.

The expert community itself remains deeply uncertain.

A survey of 2,778 researchers who had published at leading AI conferences assigned an aggregate 10 percent probability to unaided machines outperforming humans in every task by 2027, and a 50 percent probability by 2047.

Depending on how the question was phrased, approximately 38 to 51 percent of respondents assigned at least a 10 percent probability to outcomes as severe as human extinction.

Yet most respondents still believed that good outcomes from superhuman AI were more likely than bad ones.

That is not a consensus that catastrophe is imminent.

It is evidence of profound uncertainty surrounding consequences too large to ignore.

When experts disagree about whether transformation is years or decades away — and whether the result will be shared prosperity or civilizational disaster — the rational response is neither panic nor complacency.

It is institutional preparedness.

Human sovereignty should be the goal

The central political objective of the intelligence age should not be to maximize intelligence in the abstract.

It should be to maximize human sovereignty under conditions of increasingly powerful machine intelligence.

An economy can become more productive while most people become less powerful.

A society can generate more knowledge while citizens lose the ability to understand or contest the systems governing them.

Consumers can receive increasingly personalized services while becoming dependent on providers they cannot leave.

A nation can win an AI race while its people lose control over the resulting state.

GDP is therefore an inadequate measure of success.

So are benchmark scores, model size, tokens per second, scientific output, corporate valuation, and the number of users.

The essential question is whether people retain practical control over their lives.

Human sovereignty means that every person should possess a meaningful baseline of cognitive capability, privacy, economic security, political voice, and recourse against automated power.

It does not require humans to perform every task manually.

It does not mean machines may never make decisions.

It means institutions must remain accountable to people — and that people must remain capable of understanding, contesting, leaving, and changing those institutions.

A sovereign person in the intelligence age needs at least five forms of power.

They need cognitive power: access to capable systems for learning, communication, creation, and economic participation.

They need personal power: control over their data, identity, digital memory, and choice of provider.

They need economic power: a claim on wealth generated through public science, infrastructure, knowledge, and social investment.

They need political power: the right to know when automated systems are being used and to challenge consequential decisions.

And they need institutional power: accountable human judgment in domains where moral responsibility cannot legitimately be delegated.

Human sovereignty is not anti-technology.

It is the condition under which technological progress remains politically legitimate.

The Superintelligence Compact

Turning this principle into reality requires five commitments.

1. Make baseline intelligence a civic right

Every person should have affordable access to capable AI for education, communication, creativity, entrepreneurship, and participation in economic and civic life.

Schools, libraries, universities, healthcare systems, nonprofits, and community institutions should receive public compute budgets and access to multiple providers.

Meaningful access requires more than a free interface.

It requires connectivity, training, accessibility, multilingual support, and institutions capable of helping people use the technology effectively.

People should be able to move their data and digital history between providers.

Personal agents should be interoperable.

Systems should be accessible to disabled users.

A person’s AI should be loyal to that person within the law.

It should not be secretly optimized for advertising, platform lock-in, compulsive engagement, political influence, or behavioral extraction.

A model may be hosted by an institution.

It should not quietly convert its user into a resource to be mined.

2. Replace the open-versus-closed debate with a capability ladder

The argument between “open” and “closed” AI is too crude for the systems now being developed.

Research, evaluation methods, code, and lower-risk models should be presumptively open.

The burden of proof should fall on those seeking restriction.

But as systems approach defined thresholds in catastrophic domains, the burden should reverse.

Developers should have to demonstrate that a release is sufficiently safe.

At lower capability levels, the default should be freedom, transparency, competition, experimentation, local adaptation, and broad availability.

At higher levels, safeguards should progressively include independent evaluation, secure hosting, containment, rate limits, cybersecurity requirements, traceability, and restrictions on reproducing dangerous model weights.

The standard should depend on what a system can demonstrably do — not simply how much compute was used to train it, how large the company is, or whether the model is marketed as open source.

No frontier developer should certify its own safety.

Systems above designated capability thresholds should undergo independent evaluation.

Developers should publish safety cases explaining:

• what the system can do;

• where the evidence remains weak;

• which failure modes remain;

• what safeguards are in place;

• and why the proposed deployment is justified.

Governance should also include mandatory reporting of serious incidents, protected whistleblowing, external red teams, cybersecurity standards, liability rules, reproducible evaluations, and public authority to delay or condition deployment.

A regulator should not approve every software update or dictate how every model is trained.

It should ensure that institutions capable of creating civilization-scale externalities cannot conceal or unilaterally approve them.

The principle is straightforward:

The greater the potential consequence, the greater the required evidence.

3. Give the public an ownership stake

Universal access gives people tools.

Shared ownership gives them power.

Frontier AI rests on decades of publicly funded science, university research, electrical grids, internet infrastructure, semiconductor policy, public procurement, copyright systems, and knowledge created by billions of people.

The public is not an uninvolved spectator.

It is an early investor whose contribution rarely appears on a capitalization table.

Governments should establish independently governed public wealth funds that hold diversified stakes in the AI economy.

A Public Intelligence Trust could be capitalized through:

• equity warrants attached to public subsidies;

• royalties from publicly financed intellectual property;

• stakes in strategic public-private projects;

• payments for preferential access to energy or infrastructure;

• and levies on extraordinary and concentrated AI rents.

The fund should not become a political slush account or a speculative bet on a handful of laboratories.

It should hold diversified assets, operate transparently, and follow clear rules protecting it from day-to-day political interference.

Part of its long-term returns could be distributed as a universal intelligence dividend.

The remainder could support education, healthcare, science, public infrastructure, resilience, and future generations.

The purpose is not to punish successful companies.

It is to recognize that the intelligence economy is being built on extensive public and social investment.

A society that helped finance the intelligence age should not be asked to accept only cheaper subscriptions in return.

4. Give workers voice, value, and automatic protection

Employers should not be allowed to introduce high-impact workplace AI solely as a system of surveillance, intensification, deskilling, or displacement.

Workers and their representatives should participate in decisions involving monitoring, performance evaluation, automation, workload, safety, scheduling, and job redesign.

Large organizations should publish AI transition assessments explaining:

• which tasks are expected to change;

• which roles may expand or contract;

• what productivity gains are anticipated;

• how those gains will be measured;

• and how workers will share in the resulting value.

Sharing could mean higher wages, shorter working hours, profit participation, retraining guarantees, portable benefits, stronger retirement contributions, or employee ownership.

Governments should also establish real-time labor-market observatories tracking hiring, wages, hours, displacement, occupational transitions, regional effects, and job quality.

When disruption passes predefined thresholds, wage insurance, unemployment support, healthcare coverage, retraining accounts, mobility assistance, and direct financial support should expand automatically.

Waiting for a political crisis before building the response would be negligent.

Education policy must also stop pretending that one mid-career retraining course can prepare a person for permanent technological change.

Every adult should have access to a lifelong learning account.

Institutions receiving public funding should be rewarded for verified employment, capability, and income outcomes — not enrollment numbers alone.

The goal should not be to help workers endure a transition designed without them.

It should be to give them power over the transition itself.

5. Create enforceable rights against automated power

Every person should have the right to know when they are interacting with an AI or being materially evaluated by one.

They should have the right to meaningful explanation and human appeal in high-impact decisions.

They should have the right to access, correct, and move their data.

They should have the right to privacy-preserving alternatives.

They should have the right to refuse manipulative personalization.

No AI system should become the final, unappealable authority over a person’s liberty, citizenship, access to essential healthcare, or fundamental political rights.

Children should receive heightened protections against emotional dependency, pervasive profiling, behavioral manipulation, and commercial exploitation.

Intimate AI systems should not be permitted to exploit attachment in order to drive purchases, ideology, engagement, or disclosure.

Some decisions should remain subject to accountable human judgment.

Not because human beings are infallible.

But because moral responsibility cannot be delegated completely to an entity that cannot be punished, voted out, shamed, forgiven, or required to sacrifice its own interests.

Communities also deserve rights against the physical costs of AI infrastructure.

Major data centers should disclose projected energy and water use, finance the generation and transmission capacity they require, participate in demand-response systems, and provide enforceable benefits to communities carrying concentrated costs.

New capacity should, wherever practicable, be additional rather than diverted from households and existing industries.

Public land, tax credits, energy access, grid priority, and infrastructure should not be offered without public participation in the upside.

Competition must be bounded by cooperation

Democratic states should protect semiconductor supply chains, secure frontier systems, invest in energy and research, and prevent the most dangerous capabilities from empowering hostile actors.

But strategic competition must be bounded by strategic cooperation.

Major AI powers should negotiate common standards for frontier evaluation, serious-incident reporting, compute-cluster security, emergency communication, and prohibited uses.

The world needs an International Frontier AI Agency.

It should not be a global ministry of algorithms.

It should be a technical institution capable of coordinating common evaluations, investigating serious incidents, verifying selected commitments, and maintaining trusted channels among rival states.

Nuclear governance offers both inspiration and warning.

Humanity developed hotlines, inspections, test restrictions, command safeguards, and arms-control agreements not because rivals trusted one another, but because certain failures were too dangerous to leave entirely to competition.

The same principle applies here.

States can compete fiercely in beneficial AI while agreeing that no one benefits from unsecured self-replicating systems, automated biological-weapons design, attacks on nuclear command infrastructure, or autonomous systems capable of initiating strategic conflict without meaningful human authorization.

Every extraordinary power created in response to AI risk should also expire unless renewed.

Emergency restrictions, surveillance authorities, export controls, and centralized government powers should contain sunset clauses, judicial review, legislative renewal, and public reporting.

The danger of superintelligence does not justify creating permanently superintelligent states.

Who is entitled to decide?

The debate over superintelligence is usually framed as a forecast.

When will it arrive?

Will it think?

Will it escape?

Will it replace us?

Will it cure disease?

Will it end scarcity?

Those questions matter.

But they distract from a question that can already be answered:

Who is entitled to decide?

Today, a remarkably small number of institutions control the frontier.

Industry produces more than 90 percent of notable models.

Five hyperscalers control most of the world’s advanced AI compute.

The most capable systems are becoming among the least transparent.

Hundreds of billions of dollars are being invested before the world has built institutions capable of measuring or governing the resulting power.

This is not evidence of conspiracy.

It is evidence of institutional lag.

The people building frontier AI are not uniquely malevolent.

Many are motivated by scientific ambition, public benefit, fear of rivals, fear of catastrophe, or a sincere belief that intelligence can improve human life.

But history does not require villains to produce dangerous concentrations of power.

It requires only incentives, momentum, secrecy, dependency, and institutions that arrive too late.

Superintelligence should not belong to the company that trains it.

It should not belong to the government that seizes it.

It should not belong to the billionaire who finances it.

It should not belong to the engineer who discovers it.

And it should not belong to the machine itself.

Its benefits should belong broadly to humanity.

Its dangerous capabilities should be constrained by law.

Its economic returns should strengthen the public that made it possible.

Its governance should be pluralistic, contestable, and democratic.

Its deployment should preserve the dignity, agency, and meaningful role of human beings.

We do not need to choose between innovation and safety, because unsafe innovation eventually destroys the public legitimacy on which progress depends.

We do not need to choose between open and closed AI, because different capabilities justify different degrees of openness.

We do not need to choose between markets and government, because markets create dynamism while democratic institutions define the boundaries within which that dynamism remains legitimate.

We do not need to choose between abundance and meaning, because the purpose of abundance should be to give people greater freedom to pursue lives they consider meaningful.

And we do not need to wait for AGI to begin.

The infrastructure is being built now.

The ownership is being assigned now.

The standards are being written now.

The habits of secrecy or transparency are forming now.

The public’s bargaining position is being determined now.

The manifesto war should not end with one chief executive defeating another.

It should end with a constitutional settlement that no chief executive can overrule.

Open at the edge.
Guarded at the frontier.
Shared at the upside.
Human at the center.

That is not a compromise between acceleration and fear.

It is a program for freedom in the age of intelligence.

References

Altman, S. (2021, March 16). Moore’s law for everything.

https://moores.samaltman.com/

Altman, S. (2025, February 9). Three observations. https://blog.samaltman.com/three-observations

Altman, S. (2025, June 10). The gentle singularity. https://blog.samaltman.com/the-gentle-singularity

Amodei, D. (2024, October). Machines of loving grace: How AI could transform the world for the better. https://darioamodei.com/essay/machines-of-loving-grace

Anthropic. (2025, May 22). Activating AI Safety Level 3 protections. https://www.anthropic.com/news/activating-asl3-protections

Anthropic. (2026). Anthropic’s Responsible Scaling Policy. https://www.anthropic.com/responsible-scaling-policy

Anthropic. (2026, April 7). Project Glasswing: Securing critical software for the AI era. https://www.anthropic.com/glasswing

Aschenbrenner, L. (2024). Situational awareness: The decade ahead.

https://situational-awareness.ai/

Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at work. National Bureau of Economic Research Working Paper Series, №31161. https://www.nber.org/papers/w31161

Epoch AI. (2026). Trends in artificial intelligence. https://epoch.ai/trends

Future of Life Institute. (2025). Statement on superintelligence.

https://superintelligence-statement.org/

Google DeepMind. (n.d.). AlphaFold. https://deepmind.google/science/alphafold/

Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., & Brauner, J. (2025). Thousands of AI authors on the future of AI. Journal of Artificial Intelligence Research. https://www.jair.org/index.php/jair/article/view/19087

International AI Safety Report. (2026). International AI Safety Report 2026. https://internationalaisafetyreport.org/sites/default/files/2026-02/international-ai-safety-report-2026_1.pdf

International Energy Agency. (2025). Energy and AI. https://www.iea.org/reports/energy-and-ai

International Energy Agency. (2026). Key questions on energy and AI. https://www.iea.org/reports/key-questions-on-energy-and-ai

International Monetary Fund. (2024, January 14). AI will transform the global economy: Let’s make sure it benefits humanity. https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity

METR. (2025, March 19). Measuring AI ability to complete long software tasks. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

METR. (2025, July 10). Measuring the impact of early-2025 AI on experienced open-source developer productivity. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

METR. (2026, May 8). Task-completion time horizons of frontier AI models. https://metr.org/time-horizons/

Meta. (2025, July 30). Personal superintelligence. https://www.meta.com/superintelligence/

OpenAI. (2023, February 24). Planning for AGI and beyond. https://openai.com/index/planning-for-agi-and-beyond/

OpenAI. (2025, November 6). AI progress and recommendations. https://openai.com/index/ai-progress-and-recommendations/

OpenAI. (2026, April 26). Our principles. https://openai.com/index/our-principles/

OpenAI. (2026). Industrial policy for the Intelligence Age: Ideas to keep people first. https://cdn.openai.com/pdf/561e7512-253e-424b-9734-ef4098440601/Industrial%20Policy%20for%20the%20Intelligence%20Age.pdf

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv. https://arxiv.org/abs/2302.06590

Stanford Institute for Human-Centered Artificial Intelligence. (2026). The 2026 AI Index Report. https://hai.stanford.edu/ai-index/2026-ai-index-report

The comforting story about automation is familiar: machines destroy some jobs, create better ones, and leave society richer. We heard it about factories, computers, and the internet. Now we hear it about generative AI: relax — the jobs will return in better form.

Maybe. But this time may be different.

AI is not merely helping workers perform existing tasks. It can increasingly handle the language, coding, search, scheduling, reporting, and support work that underpins modern white-collar employment. The IMF estimates that almost 40% of jobs worldwide are exposed to AI, rising to about 60% in advanced economies. The ILO puts generative-AI exposure at one in four jobs globally and one in three in high-income countries. In the U.S., Brookings estimates that more than 30% of workers could see at least half their tasks disrupted.

AI is also spreading unusually fast. Stanford reports that generative AI reached 53% population adoption within three years, while the share of EU enterprises using AI rose from 13.5% in 2024 to 20% in 2025. This is not a gradual, decades-long transition. It is a rapid push into administration, customer support, software, analysis, marketing, HR, and operations.

Clear thesis

Over the next 5 to 20 years, AI is likely to destroy more jobs than it directly creates, even if overall employment continues to grow for other reasons. Not every exposed occupation will disappear. But AI can reduce labor demand across many fields at once, while the jobs it creates are too few, specialized, concentrated, and mismatched to offset the losses.

Why the computers and internet comparison is too comforting

History offers the strongest case for optimism. ATMs did not eliminate bank tellers: lower branch costs encouraged expansion, while tellers shifted toward sales and customer relationships. Computer-using occupations often grew for similar reasons, as technology increased output and demand.

The real lesson is not that technology always creates jobs. It does so when new tasks and rising demand outweigh the work it eliminates. Acemoglu and Restrepo describe this as a contest between displacement and reinstatement. Workers benefit only when the second force is stronger.

AI may upset that balance. It is spreading faster, reaching more occupations, and creating a much smaller specialist workforce than the number of jobs it affects. Major estimates suggest that 25–60% of jobs face some exposure, while AI specialists represent less than 0.3% of employment. Fewer than 1% of workers require advanced AI skills.

Unlike earlier technologies, AI can also be inserted directly into existing digital workflows. One system can draft reports, summarize meetings, classify support tickets, reconcile invoices, produce legal memos, write test code, and answer customer questions. Rather than transforming one sector at a time, it is spreading across many industries simultaneously — threatening middle- and higher-paid cognitive work as well as routine jobs.

Vendors are already packaging this breadth into unified business platforms. T1U, for example, is as an AI-native operating system combining CRM, finance, HR, inventory, projects, documents, customer support, and automation. This does not itself prove job displacement, but it illustrates the mechanism: a single AI layer can spread across functions that were previously handled through separate tools, workflows, and teams.

How AI destroys jobs even when it “helps” workers

People often assume that if AI augments workers, jobs must be safe. That does not follow.

Anthropic found that 57% of observed AI use involved augmentation and 43% automation. But firms do not need full automation to cut hiring. If workers become 15% or 30% more productive, companies can produce the same output with fewer employees.

Call-center evidence shows the mechanism clearly. Generative AI raised productivity by about 15% overall and 34–35% for novice and lower-skilled workers. That improves service, but it can also reduce demand for new hires and weaken the entry-level jobs through which workers gain experience. What helps output may harm apprenticeship.

The old technology analogy is therefore less reassuring. Earlier technologies often spread slowly and affected narrower tasks, allowing new industries and occupations time to emerge. Generative AI is moving faster across a much broader range of knowledge work, while directly creating relatively few specialist roles.

Early labor-market evidence is consistent with this pattern. Employment among young workers in highly exposed occupations has declined, while older workers have fared better. Firms appear to be retaining experienced employees, adding AI tools, and hiring fewer juniors.

Research also finds that AI-adopting firms increase hiring for AI roles while reducing demand for other positions and changing the skills required. This does not prove mass unemployment. It shows how displacement can begin through weaker hiring and shrinking entry-level pathways long before it appears in headline unemployment figures.

Why the job-creation side is weaker than people think

The optimistic numbers are real, but often misread. The World Economic Forum’s 2025 report projects 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million. But WEF says this reflects several macrotrends, including technology, demographic change, geoeconomic fragmentation, and economic pressures. Many of the fastest-growing roles — farmworkers, delivery drivers, construction workers, nurses, teachers, and care workers — are expanding because of climate, infrastructure, demographics, and service demand, not because AI is directly replacing displaced clerical or junior knowledge jobs.

The direct AI workforce remains narrow. OECD estimates it at less than 0.3% of employment. IMF’s 2026 research finds that only about one in 10 job postings in advanced economies requires a genuinely new skill, mostly in professional, technical, and managerial work. OECD also reports that fewer than 1% of workers need advanced AI skills. AI will create jobs, but not at a scale close to the number of roles it may restructure.

Even the positive evidence suggests stratification rather than broad replacement. PwC’s 2026 AI Jobs Barometer describes a two-track labor market: AI-“professionalized” jobs are growing faster and receiving stronger wage gains, while more “democratized” jobs face pressure as routine expertise becomes cheaper. AI benefits workers who supervise, judge, sell, design, or integrate, while weakening roles built around routinized expertise.

Then there is the mismatch problem. U.S. projections show strong growth for data scientists, nurse practitioners, operations research analysts, and home health and personal care aides. But a laid-off bookkeeper cannot become a data scientist in a few months, and a displaced customer-support worker cannot instantly become a nurse practitioner. These jobs require different credentials, schedules, locations, and often offer different pay. Growth elsewhere does not automatically compensate the workers whose jobs disappear.

Selected growth occupations and why they do not automatically offset losses

Sectoral case studies

Clerical and administrative work shows the clearest pressure. The ILO identifies clerical occupations as the most exposed to generative AI. In the U.S., data entry keyers are projected to decline 36% from 2024 to 2034, general office clerks 7%, and office and administrative support occupations overall over the decade. These are existing job ladders that AI now directly targets.

Customer service faces similar pressure. BLS projects customer service representative employment to decline 5% from 2024 to 2034. Combined with evidence that generative AI raises call-center productivity by 15%, the likely result is fewer routine-support workers, more cases handled by a smaller group of experienced agents, and fewer entry-level openings. This is an inference, but a conservative one grounded in the strongest workplace evidence available.

Programming presents a more mixed picture. BLS projects computer programmer employment to decline 6% through 2034, while software developers grow 15%. Anthropic’s usage data shows AI activity concentrated in software development and writing, and its labor-market research finds that more exposed occupations are projected by BLS to grow less. AI is not simply “killing coding”; it is compressing lower-level implementation work while preserving or expanding demand for system design, integration, product judgment, and domain-specific engineering. Jobs are being destroyed and created, but at different career levels and for different workers.

Freelance markets provide some of the clearest early evidence of displacement. Hui and co-authors found that highly exposed freelancers experienced a 2% decline in contracts and a 5% drop in earnings after new AI tools appeared. Oxford-related and UCL-linked research found sharper declines in some substitutable categories, with writing and translation demand falling roughly 20% to 50% in certain segments. This is what substitution looks like in a market where hiring adjusts quickly.

The economic and social consequences won’t be evenly shared

When people say technology creates jobs, they often mean society eventually finds new uses for labor. That does not guarantee workers keep their wages, bargaining power, status, or career ladders. Automation can raise GDP while weakening workers.

Acemoglu and Restrepo’s task-based framework explains why: automation reduces labor demand unless new tasks emerge quickly enough to offset displacement. Their wage-inequality research estimates that automation accounts for roughly 50% of changes in the U.S. wage structure in recent decades and about 80% of the rise in the college premium. The warning for AI is clear: even without mass unemployment, the transition could produce a more unequal labor market.

The effects will also differ by gender. The ILO’s 2025 research finds that 4.7% of female employment is in the highest exposure category, compared with 2.4% of male employment. Reuters’ coverage of the same ILO analysis reported that, in high-income countries, 9.6% of female-dominated jobs are likely to be transformed by AI, versus 3.5% of male-dominated jobs. Women are more exposed because they are overrepresented in administrative and clerical work, among the first areas affected by generative AI.

Retraining is also harder than policy speeches suggest. OECD reports that one in three adults across OECD economies has low foundational skills. Participation in adult learning remains unequal, with older, migrant, and lower-qualified workers less likely to take part. A June 2026 OECD report found that only about 40% of adults had participated in learning during the previous 12 months. Those most likely to need major career changes are often least prepared to make them quickly.

Worker anxiety is therefore understandable. OECD surveys find that three in five workers fear losing their job to AI within 10 years, while two in five expect AI to reduce wages in their sector. These concerns are not proof, but they reflect a real economic risk: as routine cognitive tasks become cheaper, the value of mid-level knowledge work may fall unless institutions intervene.

The counterarguments are serious, but they don’t close the case

The first counterargument is that aggregate data do not yet show a dramatic shock. Budget Lab’s tracking finds no clear AI-related change in the occupational mix, employment, or unemployment. Brookings likewise says labor-market research on AI is still in the “first inning.” That should encourage caution, not complacency.

The early effects are appearing elsewhere: in hiring, junior roles, and fast-moving labor markets. Young workers in exposed occupations are struggling, freelancers are losing earnings, and AI-adopting firms are reducing non-AI hiring. PwC also notes that employers increasingly expect traditionally senior skills — judgment, leadership, and adaptability — earlier in careers because AI is removing the routine tasks that once served as apprenticeship.

A third argument is that AI could trigger a productivity boom, and productivity growth has often coexisted with employment growth. Autor and Salomons, however, show that productivity can reduce employment within affected sectors and become net-positive only through spillovers elsewhere. The key question is whether AI creates enough new demand, quickly enough, to absorb displaced workers. So far, direct AI job creation remains narrow, while many broader employment gains are likely to come from care, climate, and infrastructure rather than AI itself.

AI may eventually complement more workers than it replaces. But possibility is not prediction. Given the scale of exposure, rapid diffusion, limited direct AI employment, weaker hiring, eroding apprenticeships, and the mismatch between shrinking clerical roles and growing care or specialist jobs, the burden of proof now lies with the optimists — not the skeptics.

What policy should do before the damage locks in

First, governments should treat workplace AI as a labor-market institution, not just a software upgrade. The EU AI Act already classifies some employment uses as high-risk. Any system affecting hiring, firing, scheduling, promotion, pay, or task allocation should require notice, auditability, contestability, and meaningful human review.

Second, productivity gains must be shared. OECD workplace surveys link training and worker consultation to better outcomes. AI deployment should therefore be bargained, not merely announced. Collective bargaining, wage boards, gain-sharing agreements, and sectoral standards can turn productivity gains into shorter hours, better pay, or greater staffing stability instead of headcount cuts.

Third, reskilling policy must recognize that workers cannot retrain entirely on their own time and money. OECD research shows access to adult learning is unequal. Training should be paired with income support, portable benefits, apprenticeships, public microcredentials, and targeted help for workers in clerical, service, and junior knowledge roles — not only for already-advantaged tech workers.

Fourth, governments should expand labor-intensive sectors society already needs, including elder care, childcare, education, public health, transit, retrofitting, and climate resilience. These are foundational industries, not consolation prizes. But they will absorb displaced workers only if wages, staffing, and status improve.

Strong conclusion and call to action

The lazy version of the AI debate asks whether machines will “take all the jobs.” That is the wrong question. The better question is whether AI will replace enough tasks, fast enough, in enough occupations, with too little compensating job creation, to leave millions of workers worse off even in a growing economy.

That outcome is no longer some fringe dystopian fantasy. It is the trajectory the evidence increasingly points toward.

Not because AI creates no work. It will. Not because all augmentation becomes unemployment. It won’t. But because the labor-saving power of AI is scaling faster than the labor-absorbing power of the institutions around it. And when that happens, the default result is not a leisure society. It is a more unequal one.

If we want a different ending, we need to stop repeating that old automation bedtime story and start making harder political choices: regulate workplace AI, protect entry-level pathways, fund transitions, rebuild bargaining power, and force productivity gains to show up in workers’ lives — not just in quarterly margins.

Otherwise the headline will write itself.

And it won’t sound optimistic at all.

References

Acemoglu, D., Autor, D., Hazell, J., & Restrepo, P. (2022). Artificial intelligence and jobs: Evidence from online vacancies. Journal of Labor Economics. https://www.journals.uchicago.edu/doi/10.1086/718327

Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30. https://doi.org/10.1257/jep.33.2.3

Acemoglu, D., & Restrepo, P. (2020). Robots and jobs: Evidence from U.S. labor markets. Journal of Political Economy, 128(6), 2188–2244. https://www.journals.uchicago.edu/doi/10.1086/705716

Acemoglu, D., & Restrepo, P. (2020). Unpacking skill bias: Automation and new tasks. AEA Papers and Proceedings, 110, 289–294. https://doi.org/10.1257/pandp.20201063

Acemoglu, D., & Restrepo, P. (2021). Tasks, automation, and the rise in U.S. wage inequality (Working paper). https://economics.mit.edu/sites/default/files/2022-10/Tasks%20Automation%20and%20the%20Rise%20in%20US%20Wage%20Inequality.pdf

Anthropic. (2025, February 10). Introducing the Anthropic Economic Index. https://www.anthropic.com/news/the-anthropic-economic-index

Anthropic. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. https://www.anthropic.com/research/labor-market-impacts

Anthropic. (2026, June 26). The Anthropic Economic Index. https://www.anthropic.com/economic-index

Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30. https://economics.mit.edu/sites/default/files/publications/why%20are%20there%20still%20jobs%202014.pdf

Autor, D. H., & Salomons, A. (2018). Is automation labor-displacing? Productivity growth, employment, and the labor share. Brookings Papers on Economic Activity. https://www.brookings.edu/wp-content/uploads/2018/03/1_autorsalomons.pdf

Bessen, J. E. (2015). Toil and technology. Finance & Development, 52(1). https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

Bessen, J. E. (2018). Artificial intelligence and jobs: The role of demand. In A. Agrawal, J. Gans, & A. Goldfarb (Eds.), The economics of artificial intelligence: An agenda. University of Chicago Press. https://www.nber.org/books-and-chapters/economics-artificial-intelligence-agenda/artificial-intelligence-and-jobs-role-demand

Brookings Institution. (2024, October 10). Generative AI, the American worker, and the future of work. https://www.brookings.edu/articles/generative-ai-the-american-worker-and-the-future-of-work/

Brookings Institution. (2026, January 21). Measuring U.S. workers’ capacity to adapt to AI-driven job displacement. https://www.brookings.edu/articles/measuring-us-workers-capacity-to-adapt-to-ai-driven-job-displacement/

Brookings Institution. (2026, March 10). Research on AI and the labor market is still in the first inning. https://www.brookings.edu/articles/research-on-ai-and-the-labor-market-is-still-in-the-first-inning/

Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942. https://academic.oup.com/qje/article/140/2/889/7990658

Dallas Federal Reserve Bank. (2026, January 6). Young employment declines in jobs with high AI exposure. https://www.dallasfed.org/research/economics/2026/0106

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308. https://www.science.org/doi/10.1126/science.adj0998

Eurostat. (2025, January 23). Usage of AI technologies increasing in EU enterprises. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20250123-3

Eurostat. (2025, December 11). 20% of EU enterprises use AI technologies. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2

Gmyrek, P., Berg, J., & Bescond, D. (2025). Generative AI and jobs: A refined global index of occupational exposure (ILO Working Paper №140). International Labour Organization. https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure

Handa, K., Tamkin, A., McCain, M., Huang, S., Durmus, E., Heck, S., Mueller, J., Hong, J., Ritchie, S., Belonax, T., Troy, K. K., Amodei, D., Kaplan, J., Clark, J., & Ganguli, D. (2025). Which economic tasks are performed with AI? Evidence from millions of Claude conversations. arXiv. https://arxiv.org/abs/2503.04761

Hampole, M., & Schmidt, L. (2025). Artificial intelligence and the labor market (NBER Working Paper №33509). National Bureau of Economic Research. https://www.nber.org/system/files/working_papers/w33509/revisions/w33509.rev0.pdf

Hui, X., Reshef, O., & Zhou, L. (2024). The short-term effects of generative artificial intelligence on employment: Evidence from an online labor market (Working paper). https://www.ifo.de/DocDL/cesifo1_wp10601.pdf

International Labour Organization. (2025, May 20). Generative AI and jobs: A 2025 update. https://www.ilo.org/publications/generative-ai-and-jobs-2025-update

International Labour Organization. (2025, July 8). Work transformed: The promise and peril of artificial intelligence. https://www.ilo.org/publications/work-transformed-promise-and-peril-artificial-intelligence

International Monetary Fund. (2024). Gen-AI: Artificial intelligence and the future of work (Staff Discussion Note SDN/2024/001). https://www.imf.org/-/media/files/publications/sdn/2024/english/sdnea2024001.pdf

International Monetary Fund. (2026). New jobs creation in the AI age (Staff Discussion Note SDN/2026/001). https://www.imf.org/en/publications/staff-discussion-notes/issues/2026/01/09/bridging-skill-gaps-for-the-future-new-jobs-creation-in-the-ai-age-572136

Lane, M. (2024). Who will be the workers most affected by AI? A closer look at the impact of AI on women, low-skilled workers and other groups (OECD Artificial Intelligence Papers №26). OECD Publishing. https://www.oecd.org/en/publications/who-will-be-the-workers-most-affected-by-ai_14dc6f89-en.html

Lane, M., Williams, M., & Broecke, S. (2023). The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers (OECD Social, Employment and Migration Working Papers №288). OECD Publishing. https://www.oecd.org/en/publications/the-impact-of-ai-on-the-workplace-main-findings-from-the-oecd-ai-surveys-of-employers-and-workers_ea0a0fe1-en.html

OECD. (2024). Artificial intelligence and the changing demand for skills in the labour market. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/04/artificial-intelligence-and-the-changing-demand-for-skills-in-the-labour-market_861a23ea/88684e36-en.pdf

OECD. (2024). OECD Employment Outlook 2024. OECD Publishing. https://www.oecd.org/en/publications/oecd-employment-outlook-2024_ac8b3538-en.html

OECD. (2024). Using AI in the workplace: Opportunities, risks and policy responses. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/03/using-ai-in-the-workplace_02d6890a/73d417f9-en.pdf

OECD. (2025). The characteristics of the artificial intelligence workforce across OECD countries. International Journal of Labour Economics. https://link.springer.com/article/10.1007/s41027-024-00549-7

OECD. (2025). Trends in adult learning. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/06/trends-in-adult-learning_f0d8514f/ec0624a6-en.pdf

OECD. (2026). Helping adults overcome barriers to training. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/06/helping-adults-overcome-barriers-to-training_c624d810/1d108027-en.pdf

OECD. (2026). Skills in the AI age. OECD Publishing. https://www.oecd.org/en/publications/skills-in-the-ai-age_972bd15e-en.html

PwC. (2026). Global AI jobs barometer. https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html

Stanford Institute for Human-Centered AI. (2026). AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report

Teutloff, O., et al. (2025). Winners and losers of generative AI: Early evidence of shifts in freelancer demand. Journal of Economic Behavior & Organization, 235, 106845. https://www.sciencedirect.com/science/article/pii/S0167268124004591

U.S. Bureau of Labor Statistics. (2025, August 28). Customer service representatives. Occupational Outlook Handbook. https://www.bls.gov/ooh/office-and-administrative-support/customer-service-representatives.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Data for occupations not covered in detail: Data entry keyers. Occupational Outlook Handbook. https://www.bls.gov/ooh/about/data-for-occupations-not-covered-in-detail.htm

U.S. Bureau of Labor Statistics. (2025, August 28). General office clerks. Occupational Outlook Handbook. https://www.bls.gov/ooh/office-and-administrative-support/general-office-clerks.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Bookkeeping, accounting, and auditing clerks. Occupational Outlook Handbook. https://www.bls.gov/ooh/office-and-administrative-support/bookkeeping-accounting-and-auditing-clerks.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Computer programmers. Occupational Outlook Handbook. https://www.bls.gov/ooh/computer-and-information-technology/computer-programmers.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Software developers, quality assurance analysts, and testers. Occupational Outlook Handbook. https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Data scientists. Occupational Outlook Handbook. https://www.bls.gov/ooh/math/data-scientists.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Operations research analysts. Occupational Outlook Handbook. https://www.bls.gov/ooh/math/operations-research-analysts.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Home health and personal care aides. Occupational Outlook Handbook. https://www.bls.gov/ooh/healthcare/home-health-aides-and-personal-care-aides.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Fastest growing occupations. https://www.bls.gov/emp/tables/fastest-growing-occupations.htm

U.S. Bureau of Labor Statistics. (2025, August 28). Office and administrative support occupations. Occupational Outlook Handbook. https://www.bls.gov/ooh/office-and-administrative-support/

U.S. Bureau of Labor Statistics. (2026). Industry and occupational employment projections overview and highlights, 2024–34. Monthly Labor Review. https://www.bls.gov/opub/mlr/2026/article/industry-and-occupational-employment-projections-overview.htm

World Economic Forum. (2025). The future of jobs report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025/

Yale Budget Lab. (2026, June 15). Tracking the impact of AI on the labor market. https://budgetlab.yale.edu/research/tracking-impact-ai-labor-market

Over the last decade, companies accumulated software the way cities accumulate traffic: one application for customer relationships, another for finance, another for projects, another for human resources, another for support, another for analytics, another for automation, followed by an expanding network of integrations attempting to hold the entire structure together.

The individual tools may be excellent, yet the operating environment they create is often anything but excellent. Zylo reports that the average company manages approximately 305 applications, IBM says 81% of IT leaders believe data silos are obstructing digital transformation, and Atlassian has cited research associating frequent context switching with an average productivity decline of roughly 40%.

These statistics reveal a larger truth: the central business problem is no longer digitization; it is fragmentation.

Companies have successfully digitized sales, finance, operations, human resources, service, and analytics, but they have usually digitized them as separate worlds. Each department gained better software while the organization as a whole became more difficult to understand, coordinate, and operate.

That is why we built T1U.ai.

Not to add another dashboard to an already crowded screen, not to introduce another application employees must monitor, and not to make a business slightly more digital while leaving it fundamentally disjointed.

We built T1U because modern companies are still being forced to operate like disconnected departments, even though the business itself is one living system.

We’ve spent 30 years digitizing business. Now it’s time to synchronize it.

That idea also explains the name: T1U stands for The One Unified-System.

A Business Is Not a Collection of Applications

A company is not a CRM connected to an ERP, attached to a help desk, supported by a business-intelligence layer, and held together by a dozen increasingly fragile integrations.

A company is an organism.

Sales should inform finance, finance should inform operations, operations should inform service, and service should reshape retention. These relationships should be visible as they emerge, rather than reconstructed several weeks later through reports, exports, spreadsheets, and meetings.

This is where much of enterprise software still falls short. Most platforms are designed around the requirements of a particular department, which means each system sees only the part of the business it was built to manage.

The CRM understands conversations, opportunities, and pipeline. Finance understands transactions, invoices, and cash exposure. Support understands cases, complaints, and customer sentiment. Project software understands milestones, deadlines, and resource constraints.

Each system may contain accurate information, but none necessarily understands how the information connects.

A technically integrated company can still be operationally fragmented.

Moving data between systems is not the same as understanding the relationships among customers, revenue, delivery, service, cash flow, and workforce capacity. An integration can transfer a record, but it may not explain why win rates are falling, whether implementation delays are increasing churn risk, or how weaker pipeline quality could eventually affect receivables.

The next enterprise platform should therefore begin with unity rather than trying to manufacture unity later through integrations.

AI Must Become Part of the Operating Model

T1U.ai is designed as an AI-native business operating system that unifies CRM, finance, human resources, inventory, projects, documents, support, communication, analytics, and automation through one operating environment.

The important distinction is architectural.

T1U does not present artificial intelligence as a decorative feature placed above old workflows. It presents AI as part of the operating model itself.

At the center of that model is NeoMind™ AI, which is embedded across products including CRM+, Finance+, Insights+, Drive+, Email+, Automate+, and Support+. Rather than operating as an isolated chatbot with limited organizational context, NeoMind is designed to work across tenant data while respecting role-based access controls.

T1U dashboard with all available modules on the sidebar, including CRM+, Drive+, Finance+, Project+, and Stock+

This matters because a great deal of enterprise AI still feels like a clever assistant standing outside the room, waiting for someone to explain what is happening inside.

NeoMind is intended to become part of the room itself.

Within the analytics environment, a prompt can become a set of filters, visualizations, and an executive narrative that can be preserved and replayed for a board meeting, leadership review, or weekly operating session. The value is not simply that AI can generate text; the value is that the analysis remains connected to live business data and can become part of the organization’s decision-making process.

That is a significantly more ambitious role for AI.

The goal is not to help employees ask isolated questions more quickly. The goal is to help the company understand the relationships among the answers.

The Real UX Breakthrough Is Not Another Interface

T1U spans a broad range of modules, including CRM+, Finance+, HR+, Drive+, Email+, Project+, Proposal+, Stock+, Support+, Insights+, Automate+, and Workforce+, while also offering more than 600 integrations and connectivity across business applications and financial services.

Yet the feature list is not the most important part of the story.

The deeper promise is that employees should no longer need to remember which application owns which answer.

If a revenue-operations leader wants to understand why conversion declined, the answer should not require several browser tabs, multiple exports, and an analyst manually reconciling conflicting data.

If a CFO wants to examine the relationship between pipeline quality and receivables risk, the analysis should not begin with a request for every department to submit another spreadsheet.

If an operations leader wants to understand whether fulfillment delays are increasing customer churn, the connection should not remain invisible simply because inventory and support live in different systems.

The real breakthrough is not placing more information on one screen; it is removing the friction between intent and action.

Ask a question. Surface the relevant information. Understand the relationships. Trigger the appropriate workflow. Move the business forward.

T1U CRM+ centralizes lead capture, scoring, segmentation, ownership, and pipeline tracking so teams can prioritize and convert opportunities faster.

This is the logic behind the idea of a zero-UI enterprise. It does not mean that interfaces disappear or that dashboards become irrelevant; it means that software should no longer force people to navigate its internal structure before they can solve a business problem.

The future of enterprise UX is not an endless expansion of menus, reports, and screens. It is an operating environment in which orchestration matters more than screen count.

Unity First, Not Integration Later

The market already contains many strong products.

Salesforce remains a formidable CRM and AI platform. HubSpot has built an excellent customer platform. Odoo offers impressively broad business coverage, while Zoho has pursued its own operating-system vision for years.

T1U’s argument is not that these companies have failed to build valuable products. The argument is that many growing organizations still end up combining front-office, back-office, analytics, and workflow applications into a stack that is technically connected but operationally divided.

The existing model typically begins with specialization and attempts to create unity later.

T1U begins with the opposite premise.

The company should be designed as one operating system from the beginning, with modules extending a shared core rather than becoming separate products that happen to exchange data.

T1U’s NeoMind AI chat helps companies orchestrate 600+ apps and create automation workflows.

T1U’s NeoMind AI chat helps companies orchestrate 600+ apps and create automation workflows.

That difference may appear architectural, but its consequences are organizational. A shared system can preserve context as information moves between departments, allowing a support issue to become a retention signal, an inventory constraint to become a financial warning, or a hiring slowdown to become a delivery risk.

The business should not have to rediscover itself every time a decision needs to be made.

Trust Is Not a Feature; It Is the Foundation

A unified platform inevitably carries greater responsibility because it brings together customer information, financial data, operational records, employee workflows, documents, and internal decisions.

For that reason, trust cannot be treated as another line in a feature comparison.

T1U publicly emphasizes a security posture based on ISO 27001 and SOC 2 standards, GDPR compliance, encryption in transit and at rest, regular key rotation, multifactor authentication, role-based access control, and single sign-on.

These controls are not optional differentiators. They are table stakes for any platform asking a company to centralize critical information and operational authority.

The more unified the system becomes, the less room there is for vague security promises.

The same principle applies to AI. As systems move beyond answering questions and begin triggering workflows or influencing decisions, permissions, transparency, auditability, and accountability become inseparable from intelligence.

The future will not belong to the AI system that acts most freely. It will belong to the system that can act usefully while remaining governed, explainable, and under human control.

The Commercial Model Should Reflect the Product Philosophy

T1U’s pricing philosophy also reflects the broader idea of unification.

Public pricing has presented a shared base platform with additional modules added according to organizational requirements, including examples built around a core subscription and products such as CRM+.

This approach matters because a unified platform should simplify not only the technology environment, but also the commercial environment surrounding it.

The traditional SaaS stack frequently creates dozens of contracts, renewal dates, user tiers, pricing models, and departmental buying decisions. Even where the applications work well individually, the combined purchasing experience becomes another form of operational fragmentation.

A shared operating core with modular expansion offers a more coherent alternative: one platform, one commercial relationship, and modules designed to work together rather than being procured as unrelated products.

Software consolidation should simplify the buying motion as well as the working motion.

Who T1U Is Really For

T1U is not primarily designed for a solo founder searching for the lightest possible CRM.

Its strongest audience is the growing or operationally complex organization that has already begun to feel the limits of disconnected software: B2B revenue teams, finance-led operators, multi-entity businesses, operations-heavy firms, and companies that have outgrown spreadsheets but do not want to spend years constructing another patchwork enterprise stack.

T1U’s NeoMind AI turns live financial, workforce, project, and inbox data into actionable recommendations that help teams make faster, budget-aware decisions.

These organizations often share the same symptoms. Their departments use capable software, yet important decisions still require manual reconciliation. Their leaders have dashboards, yet struggle to find one operational truth. Their data is technically accessible, yet context remains divided across systems.

For these companies, the problem is not that the existing tools do nothing.

The problem is that they do not behave like one coordinated machine.

What a Coordinated Business Could Look Like

Imagine a revenue leader asking NeoMind why win rates declined during the previous quarter. Instead of returning a generic sales summary, the system examines lead quality, deal velocity, pricing exceptions, support friction, and delayed implementation milestones across the company.

Imagine a CFO preparing for a board meeting without waiting for several departments to reconcile spreadsheets. NeoMind transforms a prompt into filtered views, visualizations, and an executive narrative grounded in current platform data.

T1U centralizes multi-entity accounting, local standards, IFRS reporting, intercompany, FX, and consolidation in one governed platform.

Imagine an operations leader seeing inventory friction before it becomes a cash-flow problem, a support issue becoming visible as an early retention signal, or a hiring slowdown being recognized as a future delivery risk.

These examples illustrate a deeper principle.

The value of a unified system is not the elegance of the question; it is that the business no longer needs to reconstruct its own reality every time someone asks one.

That is the gap T1U is attempting to close.

Operational Outcomes, Not AI Novelty

T1U has also begun presenting its story through measurable outcomes, including company-reported claims involving faster project delivery, significant savings, high returns on investment, and relatively short payback periods.

Such figures should always be assessed with appropriate context, including the customer involved, the baseline used, the implementation period, and whether the results were independently audited.

Nevertheless, the strategic direction is important.

T1U does not appear to want to compete on AI novelty alone.

It wants to compete on operational outcomes.

T1U’s NeoMind AI lets users speak requests aloud to instantly create employees and complete other business actions through a simple confirmation.

That distinction will matter increasingly as AI features become ubiquitous. Almost every enterprise platform will eventually offer assistants, generated summaries, and automated workflows. The question will no longer be whether a product contains AI, but whether the product measurably improves how the organization operates.

AI will become common; coordinated execution will remain rare.

The Next Era of Business Software

The previous era of business software taught companies how to digitize. The emerging era will force them to decide whether digitization without unification is still enough.

For many organizations, it is not.

They are managing too many tools, too many handoffs, too many exports, too many contracts, and too many contradictory versions of the truth. Their employees spend too much time navigating the seams between systems and too little time building momentum inside the business.

We built T1U because we believe the next great business platform should not feel like a collection of products awkwardly sharing a logo.

It should feel like one intelligence layer across the company: one operating environment through which sales, finance, operations, service, people, and data can contribute to a shared understanding of reality.

It should be a place where AI is not bolted onto the business after the fact, but embedded into how the organization understands, decides, and works.

The last generation of software gave every department a system of record. The next generation must give the entire company a system of understanding.

That is what T1U is trying to become.

For companies tired of managing the seams between their tools instead of the momentum of their business, that future may no longer be theoretical.

It may already be overdue.

References

Atlassian. (2022, July 13). The cost of context switching (and how to avoid it). https://www.atlassian.com/blog/it-teams/cost-of-context-switching

Gartner. (2025, August 22). How EA and application teams collaborate on SaaS sprawl. https://www.gartner.com/en/documents/6873766

HubSpot. (n.d.). Customer platform pricing. https://www.hubspot.com/pricing/suite

HubSpot. (n.d.). HubSpot homepage.

https://www.hubspot.com/

IBM. (n.d.). What are data silos? https://www.ibm.com/think/topics/data-silos

IBM. (n.d.). Design your data strategy. https://www.ibm.com/think/insights/data-matters/data-strategy

Khatib, O. (2025, December 2). How Olivier Khatib is fixing the broken ERP market with AI. Entrepreneur. https://uk.entrepreneur.com/entrepreneurs/how-olivier-khatib-is-fixing-the-broken-erp-market-with-ai/500402

Odoo. (n.d.). Pricing. https://www.odoo.com/pricing

Salesforce. (n.d.). Sales pricing. https://www.salesforce.com/eu/sales/pricing/

Salesforce. (n.d.). Salesforce products. https://www.salesforce.com/products/

T1U. (n.d.). Analytics engine. https://www.t1u.com/analytics-engine

T1U. (n.d.). API reference. https://www.t1u.com/api-reference

T1U. (n.d.). Case studies. https://www.t1u.com/case-studies

T1U. (n.d.). Mission, journey & AI business operating system. https://www.t1u.com/about

T1U. (n.d.). NeoMind™ AI. https://www.t1u.com/neomind-ai

T1U. (n.d.). Partners. https://www.t1u.com/partners

T1U. (n.d.). Pricing. https://www.t1u.com/pricing

T1U. (n.d.). Privacy center. https://www.t1u.com/privacy-center

T1U. (n.d.). Security. https://www.t1u.com/security

T1U Research Team. (2026). Beyond the chatbox: How NeoMind™ orchestrates a zero-UI enterprise. https://www.t1u.com/journal/beyond-chatbox-neomind-zero-ui-enterprise-2026

T1U Research Team. (2026). The SaaS tax: Why your 50+ app stack is actually stalling your growth. https://www.t1u.com/journal/saas-tax-50-plus-app-stack-stalling-growth-2026

T1U Research Team. (2026, February 18). ERP systems compared: Cloud vs on-premise 2026. https://www.t1u.com/journal/erp-systems-compared-cloud-vs-on-premise-2026

Zylo. (2025, July 29). Stop SaaS sprawl: 4 steps to stymie portfolio growth. https://zylo.com/blog/saas-sprawl

For years, frontier AI was treated like software.

A company built a model, tested it, wrapped it in safety systems, published a launch post, sold API access, and waited for the market to respond. The debate around those models was usually framed in familiar Silicon Valley language: capability, benchmarks, safety, developer adoption, and enterprise trust.

Then Anthropic’s Claude Fable 5 and Claude Mythos 5 were switched off.

Within days of launch, the U.S. government ordered Anthropic to suspend access to the models for foreign nationals. Anthropic responded by disabling both models globally because it could not immediately and reliably separate permissible from impermissible users across all access channels.

That was the headline.

But the deeper story is much larger.

This was not merely an AI safety dispute. It was not simply a cybersecurity scare. It was not just a clash between Anthropic and Washington.

It was a signal that frontier AI has crossed a political threshold.

A leading AI model was treated not as a software product, but as a governed strategic capability — something closer to advanced semiconductors, cryptography, cyber tools, or dual-use technology than to ordinary cloud software.

For one weekend in June 2026, the world saw that access to the most capable AI systems can be interrupted not only at the chip layer or the data-center layer, but at the model-access layer itself.

That changes almost everything.

What actually happened

The Fable/Mythos episode compressed launch, export control, shutdown, and alliance diplomacy into less than ten days.

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026.

Fable 5 was presented as a safer public version of Mythos-class capability. Anthropic said it included safeguards around cybersecurity, biology, chemistry, and model distillation. When sensitive classifiers triggered, requests would fall back to Claude Opus 4.8 rather than be handled directly by Fable 5.

Anthropic stated that more than 95% of sessions involved no fallback at all, meaning most users would experience the stronger model directly.

Mythos 5 was designed for trusted access. It offered the same underlying capability with some cyber safeguards removed for vetted users, particularly organizations participating in Anthropic’s defensive cybersecurity initiative, Project Glasswing.

Anthropic’s safety design implied that most users would experience Fable 5 directly, while sensitive cybersecurity, biology, chemistry, or distillation-related requests would route to Opus 4.8 instead.

Anthropic also introduced a 30-day retention policy for Mythos-class traffic to support abuse detection, jailbreak monitoring, and security review.

In other words, Anthropic was not pretending this was a normal chatbot release. It was openly acknowledging that these systems required special governance.

Then the U.S. government intervened.

According to Anthropic, it received an export-control directive on June 12 ordering the company to suspend access to Fable 5 and Mythos 5 for foreign nationals. Anthropic said the government believed it had become aware of a method for bypassing Fable 5’s safeguards around identifying software vulnerabilities.

Anthropic reviewed the demonstration and argued that it involved only a small number of previously known, relatively minor vulnerabilities. The company also noted that it had spent thousands of hours working with government agencies and external evaluators on these safeguards.

Nevertheless, it complied.

Reuters reported that AWS said Anthropic requested revocation of access for all users in all regions.

That detail matters.

The legal target was foreign-national access.

The operational result was a global shutdown.

That gap between legal intent and technical implementation may be one of the most important lessons of the episode. In theory, export controls can distinguish between countries, organizations, and individuals. In practice, a frontier AI API may be easier to disable entirely than to segment perfectly in real time.

For a weekend, service architecture became export-control architecture.

Then the fight moved to the G7

The shutdown did not remain a dispute between Anthropic and Washington.

Within days, it became a topic at the G7 summit in France.

Executives from Anthropic, OpenAI, Google, and other leading AI companies joined world leaders to discuss AI governance, infrastructure, and security. The symbolism was striking: the future of frontier AI is increasingly being negotiated not only by regulators and diplomats, but by the companies that control the models themselves.

That is a signal of where power now sits.

For Europe, the episode reinforced a long-standing fear: dependence on American technology can create a hidden “kill switch” risk. When Anthropic disabled Fable 5 and Mythos 5 after a U.S. directive, that concern became tangible.

The likely compromise now being discussed is a “trusted partners” framework under which selected allies, agencies, and organizations receive access under stricter safeguards and oversight.

Why this was not an ordinary AI safety story

The easy version of this story is that Anthropic launched a dangerous model and the government panicked.

That interpretation misses the larger issue.

The more serious explanation is that advanced AI is becoming a dual-use cyber capability.

The UK AI Security Institute found that Mythos Preview could autonomously discover and exploit vulnerabilities on simulated enterprise networks. It became the first model to complete one of the institute’s most demanding cyber ranges end-to-end and later completed multiple cyber evaluation environments.

That is not ordinary chatbot behavior.

A model capable of autonomously reasoning through cyber operations occupies a category that governments already treat as strategically sensitive.

The same dynamic appeared in Project Glasswing.

Anthropic launched Glasswing to help trusted organizations identify vulnerabilities in critical software and infrastructure. By June 2026, the company said participating organizations had identified more than 10,000 high- or critical-severity vulnerabilities. Many of the newly added organizations supported infrastructure whose compromise could affect more than 100 million people.

Those numbers explain both the value and the concern.

Before the shutdown, Anthropic was expanding Mythos access from a restricted cyberdefense cohort toward a much broader international network of vetted organizations.

If a model can help defenders discover thousands of serious vulnerabilities, it is reasonable to ask whether similar capabilities could help attackers discover them too.

This is the central dual-use dilemma of frontier AI.

The more useful a system becomes for cybersecurity, biology, scientific research, or software engineering, the more plausible it becomes as a tool for misuse.

Anthropic itself had been making that argument. It built trusted-access programs, fallback systems, retention policies, and specialized governance mechanisms because it believed Mythos-class systems required them.

The U.S. government appears to have taken that premise seriously — perhaps more seriously, and more abruptly, than Anthropic expected.

The real precedent: export controls moved up the stack

For several years, the geopolitics of AI focused primarily on chips.

Who can buy advanced GPUs?

Who can access cutting-edge semiconductor manufacturing?

Which countries can train frontier models?

Those questions remain important.

But the Anthropic episode suggests a new phase.

Export controls may be moving from hardware to model access.

Traditionally, governments controlled the inputs to AI: chips, manufacturing equipment, cloud infrastructure, and compute.

The Fable/Mythos directive points toward controlling the output itself.

Who can query the model?

Who can see the results?

Who can access the API?

Who counts as a foreign person?

Whether or not a formal regime emerges, the direction is becoming clear.

Identity is becoming part of AI infrastructure.

The next generation of frontier-model governance may require nationality verification, trusted-user tiers, audit logs, compliance systems, and alliance-based access frameworks embedded directly into AI platforms.

In simpler terms:

The login screen may become a geopolitical border.

This is the most important precedent established by the shutdown.

The world has spent years debating who can train frontier AI.

Now it must confront who can use it.

The shutdown showed that frontier AI can be governed not only at the chip layer, but at the access, identity, use-case, and alliance layers.

AI sovereignty just became real

“AI sovereignty” has often sounded abstract.

Governments invoke the phrase to justify domestic AI programs, sovereign cloud initiatives, national compute projects, and industrial policy.

But many organizations still assumed that the practical solution was simply to buy access to the strongest American models.

The Fable/Mythos shutdown challenges that assumption.

If a foreign government, hospital, bank, cybersecurity agency, or critical-infrastructure operator builds workflows around a frontier U.S. model, what happens if access disappears overnight because of an American national-security directive?

That question is no longer theoretical.

The United States may have had legitimate security concerns. But for allies and customers, the lesson is uncomfortable:

Access is not ownership.

API availability is not sovereignty.

This is particularly striking because the United States has simultaneously promoted exports of the American AI technology stack to allies and partners.

The strategic message has been straightforward:

Use American AI.

Then a flagship American AI model disappeared for foreign users worldwide.

That contradiction will not go unnoticed.

If Washington wants American AI to become global infrastructure, it must offer more than the strongest models. It must offer reliability, predictability, and trusted continuity for partners.

Otherwise, the shutdown becomes an advertisement for sovereign alternatives.

The sovereignty debate has moved from speeches to procurement.

Europe’s wake-up call

Europe had already been watching Mythos closely.

Reuters reported on June 1 that the European Commission had held “productive meetings” with Anthropic about possible future access for EU bodies, including ENISA, to Mythos. That makes the subsequent shutdown especially awkward.

Two weeks later, the Commission said it was examining the practical consequences of the U.S. order and that those consequences should not be discriminatory against partners. European officials also framed the episode as further evidence that Europe needs technological sovereignty.

That reaction is predictable — and strategically important.

For years, Europe’s AI challenge has been described as a competitiveness problem: too few frontier labs, too little scale capital, too much regulation, insufficient cloud infrastructure, and dependence on U.S. platforms.

But the Anthropic shutdown reframes the issue.

It is not just that Europe may lack the best models. It is that Europe may lack guaranteed access to the best models when those models become strategically sensitive.

That is a different kind of dependency.

A European cyber agency using an American model to defend critical infrastructure is not merely buying software. It is relying on another state’s political judgment about when that software may be used, by whom, and under what geopolitical conditions.

That does not mean Europe can or should cut itself off from U.S. AI. That would be unrealistic and probably counterproductive. But it does mean European buyers will increasingly ask harder questions:

Can this model be revoked by foreign order?
Can we run it locally?
Can we obtain contractual guarantees?
Can EU public agencies receive trusted access?
Can European nationals be excluded under U.S. rules?
Do we need an EU-controlled alternative for critical systems?
Should open-weight models become part of resilience planning?

The sovereignty debate has moved from speeches to procurement.

The Amazon angle: hyperscalers are becoming geopolitical actors

One of the most revealing parts of the story is Amazon’s reported role.

Reuters reported that Amazon CEO Andy Jassy was among the technology leaders who raised concerns with senior administration officials about security risks in Anthropic’s most advanced models shortly before the crackdown.

That matters because AWS was not a neutral bystander.

AWS was a major infrastructure partner. It was also part of the Glasswing ecosystem. Anthropic’s own materials described AWS as helping test Mythos Preview in security operations and supporting efforts to harden the system.

So AWS appears to have occupied multiple roles at once: cloud platform, commercial partner, security collaborator, investor-ecosystem participant, and informal adviser to government.

This is the new political economy of AI.

Frontier models do not exist in isolation. They depend on hyperscalers for compute, distribution, hosting, enterprise access, security partnerships, and government relationships. Those hyperscalers are not just vendors. They are infrastructure statesmen.

The result is a complex governance triangle:

AI labs build the models.
Cloud giants distribute and operationalize them.
Governments decide when capabilities become national-security concerns.

In that triangle, the cloud providers may become as influential as the labs themselves.

They see usage patterns.
They manage infrastructure exposure.
They talk to governments.
They understand enterprise risk.
They may raise alarms.
They may help enforce restrictions.

That gives them enormous power over which AI capabilities reach the market, where, and under what conditions.

It also creates difficult questions. What happens when a cloud partner has commercial incentives, security obligations, and government relationships that do not perfectly align with the model developer’s interests? Who decides whether a model is too dangerous? What evidence must be shared? What happens when the cloud platform and the AI lab disagree?

The Anthropic episode suggests that the governance of frontier AI will not be written only in legislation. It will also be written in cloud contracts, partnership agreements, security briefings, and emergency compliance decisions.

The paradox of Anthropic’s own politics

There is a real irony here.

Anthropic is not an anti-regulation company. Quite the opposite.

The company has repeatedly argued that frontier AI requires serious governance. Reuters reported on June 10 that Anthropic urged Congress not to preempt state AI laws unless it enacted a rigorous federal law, and that the company called for independent safety testing of the most capable models.

Anthropic’s own policy writing has argued that governments should have legal authority to block or deter dangerous frontier AI deployments, while coupling that authority with transparency, independent evaluation, and safeguards against abuse.

So the contradiction is not that Anthropic wanted regulation and then complained when regulation arrived.

The contradiction is subtler.

Anthropic advocated rules-based authority to stop unacceptable deployments. What it appears to have received, at least from the public record, was an ad hoc directive with limited public explanation.

That distinction matters.

A government may be right that a technology is dangerous and still govern it badly. A state may have legitimate security concerns and still create procedural uncertainty. A regulator may prevent harm in the short term while damaging trust in the long term.

The core question is no longer simply: should governments have a kill switch for frontier AI?

The better question is: under what process should such a kill switch exist?

What evidence is required?
Who evaluates it?
Is there independent technical review?
Can the company appeal?
Are allies consulted?
Are trusted defensive users treated differently from unknown users?
How much can be disclosed publicly?
How do we prevent politically motivated or uneven enforcement?

These are not bureaucratic details. They are the legitimacy layer of AI safety.

Without them, frontier AI governance risks becoming unpredictable, opaque, and geopolitically destabilizing.

The prior Anthropic–Pentagon conflict makes this harder to interpret

The Fable/Mythos directive also landed in a larger context of tension between Anthropic and the U.S. government.

Reuters previously reported that Anthropic had refused to allow its models to be used for domestic surveillance and fully autonomous weapons systems. The Pentagon later designated Anthropic a “supply-chain risk,” a move that reportedly affected government contractors’ ability to use Anthropic technology.

Anthropic challenged that designation. Reuters also reported in early June that Anthropic and the White House had been trying to ease tensions ahead of Anthropic’s expected IPO.

This history does not prove that the Fable/Mythos shutdown was retaliatory. The cyber-capability concerns are too serious to reduce the story to political revenge.

But the history does affect perception.

When a company has already clashed with the Pentagon over military AI use, and then receives an opaque national-security directive days after launching its most advanced model, outsiders will inevitably ask whether the process was purely technical, partly political, or some mixture of both.

That uncertainty is damaging in itself.

For frontier AI governance to work, companies need to believe that security rules are applied consistently. Allies need to believe that access restrictions are based on risk, not domestic political disputes. Investors need to believe that regulatory event risk can be understood and priced. The public needs to believe that national-security claims are not being used as a black box.

The current record does not allow a confident judgment either way.

That is precisely the problem.

The IPO lesson: frontier AI now has regulatory event risk

Anthropic confidentially submitted a draft S-1 to the SEC on June 1, 2026. Reuters has linked the company’s government tensions to its IPO preparations.

That timing is important.

For investors, the Anthropic episode shows that frontier-model risk now looks less like ordinary software risk and more like regulatory event risk in finance, aviation, defense, or biotechnology.

The key question is no longer only: can this company build the best model?

It is also:

Can it maintain trusted status with government?
Can it survive export-control shocks?
Can it design access systems that comply with nationality-based restrictions?
Can it reassure allies and enterprises?
Can it preserve trust while cooperating with security agencies?
Can it avoid becoming politically radioactive?

In the old software world, distribution was mostly a business problem.

In frontier AI, distribution is becoming a national-security privilege.

That changes valuation. It changes governance. It changes who belongs on the board. It changes how companies manage policy, compliance, safety, and investor communications.

A lab’s policy posture is becoming part of its moat.

The defender’s dilemma

There is one particularly hard policy problem here: restrictions on powerful cyber models may hurt defenders before they hurt attackers.

Reuters reported that more than 50 cybersecurity leaders, including figures from firms such as Nvidia and Adobe, urged the administration to lift the curbs. Their argument was straightforward: these tools are valuable for defense, and taking them away from legitimate users may weaken security.

That objection deserves to be taken seriously.

If a model like Mythos can help find high- and critical-severity flaws, then denying access to trusted defenders creates real costs. The attackers do not necessarily stop. They may use other models, older systems, open alternatives, stolen credentials, human expertise, or non-U.S. tools.

The defenders, meanwhile, may lose one of the most powerful tools available to them.

This is the classic dual-use trap.

A tool that helps defenders can help attackers. But banning or restricting it too broadly may not neutralize attackers; it may simply slow down the people trying to patch systems.

That is why the long-term equilibrium is unlikely to be a simple ban.

More likely, the world moves toward layered access:

General public access for safer models.
Trusted access for vetted organizations.
Special access for critical-infrastructure defenders.
Stricter logging and retention for high-risk capabilities.
Government pre-release testing for frontier cyber models.
Allied access frameworks.
Emergency revocation mechanisms.
Higher obligations as models cross capability thresholds.

This will be messy. It will be controversial. But it is probably where the policy landscape is heading.

The future of AI access may look like finance, defense, and cloud security combined

The most likely long-term outcome is not that frontier AI disappears behind government walls.

The economic pressure to deploy these systems is too strong. The defensive value is too real. The global competition is too intense.

Instead, frontier AI access will become more segmented.

The strongest models may no longer be simple consumer products. They may become tiered capabilities governed by user identity, sector, geography, nationality, organizational trust, and intended use.

That means AI labs will need to build not only better models, but better governance infrastructure.

The next frontier AI platform may require:

Citizenship or residency checks.
Enterprise identity verification.
Sector-specific permissions.
Audit logs.
Abuse monitoring.
Data-retention policies.
Export-control classification.
Government reporting channels.
Allied-access carve-outs.
Emergency shutoff tools.
Independent safety evaluations.
Internal policy teams with real authority.

This is not the culture of consumer internet software. It is closer to a hybrid of cloud security, export compliance, defense contracting, and financial regulation.

That may slow some innovation. It may also professionalize the industry.

The open question is whether this new governance stack can be built in a way that preserves enough freedom, competition, and global collaboration to keep AI useful.

The risk to American AI leadership

There is a strategic irony at the center of the story.

The United States wants the world to run on American AI. It wants allies to use American models, American chips, American cloud platforms, and American security standards. It wants to prevent adversaries from gaining access to the most dangerous capabilities while keeping partners inside the U.S.-led technology stack.

That is a reasonable strategic goal.

But if the U.S. becomes too unpredictable, it may push allies to diversify away from American models.

The Anthropic shutdown gives foreign governments a vivid example of platform dependency risk. Even if the directive was justified, the result was abrupt. Users lost access. Foreign partners were affected. The legal basis was not publicly explained in detail. The technical evidence remains disputed.

For some governments, that will be enough.

They may not abandon American AI. But they will hedge.

They will fund local models.
They will support open-weight alternatives.
They will demand on-premise deployments.
They will build sovereign cloud capacity.
They will require continuity clauses.
They will develop procurement rules that penalize revocable foreign dependencies.

This does not mean the U.S. loses the AI race. It still has major advantages: labs, chips, cloud infrastructure, capital, talent, and defense relationships.

But leadership is not only about capability. It is also about trust.

If the U.S. wants its AI stack to become global infrastructure, it must behave like a reliable infrastructure provider, not only like a national-security gatekeeper.

That is the balance Washington now has to strike.

The shutdown’s real importance is not only who lost access, but who learned that access itself is now a strategic dependency.

Three possible futures

The Anthropic shutdown could evolve in several directions.

1. The negotiated-access future

In this scenario, Anthropic and the U.S. government resolve the dispute. Fable and Mythos access returns, but under stricter rules. Foreign users face new verification requirements. Trusted allies and critical-infrastructure partners receive special channels. The government gets more pre-release visibility into frontier cyber capabilities.

This is probably the most stabilizing outcome.

It preserves the defensive value of the models while acknowledging national-security concerns.

2. The sovereign-fragmentation future

In this scenario, the shutdown accelerates global mistrust. Europe, Canada, India, Japan, the Gulf states, and others conclude that dependence on American frontier APIs is too risky for critical systems. Sovereign AI programs receive more funding. Open-weight models become politically attractive. U.S. labs remain powerful, but global customers diversify faster.

This future is plausible because the trust damage has already occurred.

Even if access is restored, the memory of revocability remains.

3. The controlled-frontier future

In this scenario, the Anthropic directive becomes the first visible case in a broader regime. The U.S. government develops classified benchmarks for cyber, bio, and other dangerous capabilities. Models above certain thresholds become “covered frontier models.” Pre-release government testing becomes voluntary in name but effectively mandatory in practice. Labs learn that launching without deep federal consultation is commercially reckless.

This future may improve safety. It may also create concentration, favoring large labs with the legal, compliance, and government-relations capacity to navigate the regime.

The result could be safer AI, but also less open competition.

The most important lesson

The most important lesson of the Fable/Mythos shutdown is not that Anthropic was right or that the U.S. government was wrong.

The public evidence is too incomplete for that.

The real lesson is that frontier AI has become governable infrastructure.

It can be switched off.
It can be classified as a security risk.
It can be restricted by nationality.
It can trigger alliance politics.
It can move markets.
It can affect critical-infrastructure defense.
It can reshape cloud contracts.
It can influence IPO risk.
It can force countries to rethink sovereignty.

That is the arrival of a new era.

For AI labs, the message is that safety is no longer just a system card. It is a geopolitical operating model.

For governments, the message is that power over AI access must be matched by process, transparency, and legitimacy.

For enterprises, the message is that dependency on frontier APIs is now a strategic risk, not merely a vendor-management issue.

For allies, the message is that access to American AI is valuable — but not the same as control.

And for the rest of us, the message is that the AI industry is leaving its software adolescence behind.

The next phase will be more powerful, more regulated, more strategic, and more political.

The Fable/Mythos shutdown may eventually be resolved. Access may return. The specific vulnerability claim may prove minor, serious, or misunderstood.

But the precedent will remain.

For one weekend in June 2026, the world saw the future structure of AI power: not an open internet of universally available intelligence, but a layered system of capabilities governed by states, clouds, labs, alliances, and trust.

That is the real story.

Not that one model was shut down.

That AI became sovereign.

References

Anthropic. (2026, June 1). Anthropic confidentially submits draft S-1 to the SEC. https://www.anthropic.com/news/confidential-draft-s1-sec

Anthropic. (2026, June 9). Claude Fable 5 and Claude Mythos 5. https://www.anthropic.com/news/claude-fable-5-mythos-5

Anthropic. (2026, June 13). Statement on the US government directive to suspend access to Fable 5 and Mythos 5. https://www.anthropic.com/news/fable-mythos-access

Anthropic. (2026). Expanding Project Glasswing. https://www.anthropic.com/news/expanding-project-glasswing

Anthropic. (2026). Project Glasswing. https://www.anthropic.com/project/glasswing

Anthropic. (2026). Project Glasswing: Initial update. https://www.anthropic.com/research/glasswing-initial-update

Anthropic. (2026). Policy on the AI exponential. https://www.anthropic.com/policy-on-the-ai-exponential

Associated Press. (2026, June 14). Canada’s Carney says Anthropic shutdown is warning about overreliance on U.S. AI providers. https://apnews.com/article/cb081633bb4fca6ac97dcdaea0354de7

Bureau of Industry and Security. (n.d.). Deemed exports. U.S. Department of Commerce. https://www.bis.gov/learn-support/deemed-exports

Bureau of Industry and Security. (n.d.). What is a deemed export? U.S. Department of Commerce. https://www.bis.gov/learn-support/deemed-exports/what-deemed-export

Bureau of Industry and Security. (2025). Department of Commerce announces rescission of Biden-era Artificial Intelligence Diffusion Rule, strengthens chip-related export controls. U.S. Department of Commerce. https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens

Electronic Code of Federal Regulations. (n.d.). 15 CFR § 734.13 — Export. https://www.ecfr.gov/current/title-15/subtitle-B/chapter-VII/subchapter-C/part-734/section-734.13

Federal Register. (2026, April 10). American AI Exports Program: Call for proposals for pre-set consortia. https://www.federalregister.gov/documents/2026/04/10/2026-06952/american-ai-exports-program-call-for-proposals-for-pre-set-consortia

Nextgov. (2026, June). Anthropic suspends top AI models after U.S. export-control order. https://www.nextgov.com/artificial-intelligence/2026/06/anthropic-suspends-top-ai-models-after-us-export-control-order/414173/

Reuters. (2026, January 29). Pentagon clashes with Anthropic over military AI use. https://www.reuters.com/business/pentagon-clashes-with-anthropic-over-military-ai-use-2026-01-29/

Reuters. (2026, March 5). Pentagon informed Anthropic it is a supply-chain risk, official says. https://www.reuters.com/technology/pentagon-informed-anthropic-it-is-supply-chain-risk-official-says-2026-03-05/

Reuters. (2026, June 1). EU has had productive meetings with Anthropic over possible future access to Mythos. https://www.reuters.com/technology/eu-has-had-productive-meetings-with-anthropic-over-possible-future-access-mythos-2026-06-01/

Reuters. (2026, June 2). Trump signed order to promote advanced AI innovation and security, White House says. https://www.reuters.com/world/trump-signed-order-promote-advanced-ai-innovation-security-white-house-says-2026-06-02/

Reuters. (2026, June 5). Blacklisted AI company Anthropic, White House ease tensions ahead of IPO, sources say. https://www.reuters.com/business/aerospace-defense/blacklisted-ai-company-anthropic-white-house-ease-tensions-ahead-ipo-sources-say-2026-06-05/

Reuters. (2026, June 10). Anthropic urges U.S. to require safety tests for most capable AI models. https://www.reuters.com/world/anthropic-urges-us-require-safety-tests-most-capable-ai-models-2026-06-10/

Reuters. (2026, June 13). Amazon voiced concerns about Anthropic AI models before U.S. government’s crackdown. https://www.reuters.com/business/retail-consumer/amazon-voiced-concerns-about-anthropic-ai-models-before-us-governments-crackdown-2026-06-13/

Reuters. (2026, June 13). Anthropic disables top-tier AI models after U.S. order limiting foreign access. https://www.reuters.com/technology/us-blocks-foreign-access-anthropics-most-advanced-ai-models-axios-reports-2026-06-13/

Reuters. (2026, June 14). EU Commission looking at practical consequences of Anthropic decision, spokesperson says. https://www.reuters.com/legal/litigation/eu-commission-looking-practical-consequences-anthropic-decision-spokesperson-2026-06-14/

Reuters. (2026, June 14). Anthropic staff to meet White House officials next week, Axios reports. https://www.reuters.com/world/us/anthropic-staff-meet-white-house-officials-next-week-axios-reports-2026-06-14/

Reuters. (2026, June 15). Anthropic becomes cautionary sovereign AI fable. https://www.reuters.com/commentary/breakingviews/anthropic-becomes-cautionary-sovereign-ai-fable-2026-06-15/

Reuters. (2026, June 15). Cyber leaders urge U.S. to lift curbs on Anthropic’s security models. https://www.reuters.com/legal/litigation/cyber-leaders-urge-us-lift-curbs-anthropics-security-models-2026-06-15/

UK AI Security Institute. (2026). Our evaluation of Claude Mythos Preview’s cyber capabilities. https://www.aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities

UK AI Security Institute. (2026). How fast is autonomous AI cyber capability advancing? https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing

White House. (2025, July). White House unveils America’s AI Action Plan. https://www.whitehouse.gov/releases/2025/07/white-house-unveils-americas-ai-action-plan/

White House. (2025, July). Promoting the export of the American AI technology stack. https://www.whitehouse.gov/presidential-actions/2025/07/promoting-the-export-of-the-american-ai-technology-stack/

White House. (2026, June 2). Promoting advanced artificial intelligence innovation and security. https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/

Anthropic’s June 2026 warning landed with an almost absurd tension: one of the world’s most powerful AI companies was asking for a “brake pedal” on the very technology that made it powerful.

The company’s argument was not vague. Reuters reported that Anthropic wants leading AI labs to develop a coordinated and verifiable plan to slow or pause development if advanced systems begin improving themselves faster than society can manage. In the same report, one internal data point made the warning feel less theoretical: more than 80% of code merged into Anthropic’s codebase in May 2026 was authored by Claude.

That number changes the meaning of the debate.

For anyone who has spent the last decade in data science, software engineering, or applied machine learning, it signals a threshold shift. AI is no longer just a tool that helps people write code. It is becoming part of the machinery that builds the next generation of AI itself.

That is why the “shooting themselves in the foot” question matters. Anthropic and OpenAI are warning that frontier AI may soon become too powerful to govern through ordinary market incentives, while they continue to scale models, raise historic amounts of capital, win enterprise customers, and embed their systems deeper into the economy.

The contradiction is not that they are wrong to warn us. The contradiction is that they are asking society to trust the same companies that are financially rewarded for racing ahead.

If the risks are serious enough to require a brake pedal, then the obvious question is: who gets to keep their foot on the accelerator?

The financial evidence says: not yet

In the narrow business sense, Anthropic and OpenAI do not look like companies being punished for safety rhetoric.

Anthropic reportedly reached a $965 billion valuation after a $65 billion Series H round, while also moving toward a U.S. IPO. Reuters reported that Anthropic had recently raised at that valuation and confidentially filed for an IPO.

OpenAI has also continued to attract extraordinary capital. Public reports indicate that OpenAI reached a post-money valuation around $852 billion after a massive 2026 funding round.

So if “shooting themselves in the foot” means “investors are abandoning them,” the answer is no.

The IPO angle makes this even sharper. These are not just private labs asking society to slow down; they are companies preparing to ask public-market investors to buy into the upside of the same acceleration they describe as dangerous. Anthropic has confidentially submitted IPO paperwork, and OpenAI has reportedly followed with its own confidential filing. That means the contradiction is moving from venture capital into public markets.

Once these companies go public, the pressure changes. Quarterly expectations, public shareholders, analyst coverage, index inclusion, and valuation defense all make the “brake pedal” harder to use. A public company can still act responsibly, but the incentives become even more exposed: warn that AI may require extraordinary restraint, while selling investors a growth story built on continuing to scale it.

But if it means damaging their long-term legitimacy, the answer is more complicated.

They are not losing money because of safety warnings. They are gaining influence. The danger is that their safety language can start to look like a strategic moat: a way for incumbents to shape regulation after they have already reached the front of the race.

The strategic problem: safety or regulatory capture?

The Wall Street Journal’s coverage framed the central tension around recursive self-improvement: the possibility that AI systems could improve themselves without meaningful human input. The article also highlighted the criticism that Anthropic’s safety messaging may look self-serving, especially given the company’s size and competitive position.

This is the strongest critique:

When a near-trillion-dollar frontier AI company calls for a pause, smaller competitors may hear: “Freeze the race now that we are ahead.”

That does not mean the warning is fake. It means the warning has a political economy problem.

There is also a competitive irony here. If frontier models become powerful enough to accelerate AI research, then the companies building the best models may be creating tools that help others catch up. A highly capable model does not only improve the originator’s engineering team; once accessible through APIs, partnerships, leaks, open weights, or employee diffusion, it can raise the capability floor for the entire market. The better the model is at coding, research assistance, experiment design, and systems optimization, the more it can help competitors build the next model. In that sense, frontier AI is unusual: the product may become a ladder for rivals.

Large firms can afford compliance teams, safety evaluations, audits, reporting systems, security infrastructure, and government relations. Smaller startups and open-source labs often cannot.

That is why the distinction matters:

Anthropic’s Responsible Scaling Policy is closer to the better version than many headlines suggest. Its framework ties stronger safeguards to capability thresholds, including AI R&D thresholds such as automating entry-level AI research or dramatically accelerating AI scaling.

OpenAI’s Preparedness Framework also treats AI self-improvement as a tracked frontier-risk category.

The problem is not that these companies talk about safety. The problem is that the public needs proof that safety is not just branding plus regulatory influence.

Why enterprise customers may actually reward this

Safety rhetoric can be commercially useful.

Enterprise customers do not only buy model performance. They buy something they can defend to legal teams, boards, regulators, and customers. In finance, healthcare, law, government, and enterprise software, trust is part of the product.

Public trust in AI is fragile. A 2025 KPMG and University of Melbourne survey of more than 48,000 people across 47 countries found that two-thirds used AI regularly, but 58% still perceived AI as untrustworthy.

That is why Anthropic’s positioning may help rather than hurt it. A firm that says, “We have model evaluations, deployment gates, responsible scaling policies, transparency reports, and escalation thresholds,” is easier for a risk-sensitive enterprise buyer to justify.

So Anthropic and OpenAI are not only selling intelligence. They are selling governable intelligence.

The labor-market fear is rational

Your fear as a data scientist is not paranoia. It is a rational response to the data.

The IMF has estimated that AI could affect about 40% of jobs globally and about 60% of jobs in advanced economies. In 2026, IMF Managing Director Kristalina Georgieva warned that young workers could be especially vulnerable because entry-level roles may be easier to automate or compress.

The World Economic Forum’s 2025 Future of Jobs work projects that by 2030, AI and other labor-market shifts could create 170 million jobs while displacing 92 million, for a net gain of 78 million. That sounds positive, but it hides the painful part: net gains do not protect individuals, occupations, regions, or career ladders.

The macro story can be optimistic while the individual story is brutal.

The danger is not simply unemployment. It is compression: fewer junior roles, fewer ordinary analyst roles, fewer routine coding jobs, fewer content jobs, and more value captured by a smaller number of highly leveraged workers.

The best workers may become dramatically more productive. The median worker may become easier to replace.

Both can be true.

The market is accelerating, not slowing

The broader AI market still looks like an acceleration story.

Share of firms reporting AI use, 2023–2025, McKinsey & Company, The State of AI global survey series, 2023–2025. Accessed June 5, 2026.

The adoption curve is rising because the technology is useful. Firms are not waiting for philosophical certainty. They are integrating AI wherever it reduces cost, increases output, or compresses time.

This is the key point: the AI industry is not actually braking. It is negotiating permission to keep accelerating.

That is why the rhetoric feels so strange. The companies are saying, “This technology may become too powerful to govern,” while simultaneously building infrastructure, signing cloud deals, launching products, and raising historic amounts of capital.

Public-company market capitalizations from financial market data as of June 2026; Anthropic valuation from Reuters reporting published June 1, 2026; OpenAI valuation from reported April 2026 funding-round data.

The competitive landscape

This table shows why “pause” talk is politically sensitive. The frontier AI ecosystem is already dominated by firms with enormous capital, compute access, distribution, and lobbying capacity.

A safety regime can protect society. But it can also harden the market around incumbents.

The real foot-gun

Anthropic and OpenAI are not shooting themselves in the foot financially. They are doing something more subtle: they are creating a credibility trap.

They want to be seen as:

1. builders of the most powerful AI systems,

2. responsible stewards of those systems,

3. advisers to governments on how those systems should be regulated,

4. commercial winners in the market those regulations will shape.

That is a lot of power to ask society to trust.

The danger is that the public begins to see AI safety not as genuine risk management, but as a language of elite control: “We built the frontier, we own the models, we define the risks, and we help write the rules.”

That is the foot-gun.

Not the warning.
The contradiction.

What they should do instead

The solution is not to stop talking about risk. The solution is to make the risk governance concrete, symmetric, and independently checkable.

The key principle should be simple:

Any brake pedal must not be controlled only by the companies driving the car.

What this means for data scientists

For data scientists, the lesson is not “learn prompting and relax.” It is also not “everything is over.”

The durable skillset is shifting.

The future belongs less to people who merely build models and more to people who can:

The safest professional position is not “AI user” or “AI skeptic.” It is AI systems governor: someone who can use, test, constrain, deploy, and explain intelligent systems in real environments.

That is especially true for experienced data scientists. The market may punish routine analytics, but it will reward people who can connect models to accountability, economics, and operational reality.

Final verdict

Anthropic and OpenAI are not currently shooting themselves in the foot in the financial sense. Investors are still funding them. Enterprises are still buying. Users are still adopting. The AI economy is still accelerating.

But they may be shooting themselves in the foot in a deeper legitimacy sense.

They are asking society to believe two things at once:

1. AI may soon become powerful enough to require extraordinary governance.

2. The companies racing to build it should remain central authorities in deciding what that governance looks like.

The coming IPOs make that legitimacy problem harder to ignore. Anthropic and OpenAI are not only asking governments and enterprises to trust them. They are preparing to ask public investors to underwrite their growth. That turns AI safety from a private-governance debate into a public-market test: can companies credibly promise restraint while their valuations depend on acceleration?

That tension cannot hold forever.

The companies are right that the world may need a brake pedal. But the public will not accept a brake pedal designed only by the firms with the most to gain from controlling the road.

References

Anthropic. (2026). Anthropic’s Responsible Scaling Policy: Version 3.0. Anthropic. https://www.anthropic.com/responsible-scaling-policy

Anthropic. (2026). Anthropic Economic Index report: Learning curves. Anthropic. https://www.anthropic.com/economic-index

Anthropic Institute. (2026). When AI builds itself. Anthropic. https://www.anthropic.com/news/when-ai-builds-itself

Euronews. (2026, June 5). Anthropic calls for “brake pedal” before AI develops itself without human oversight. Euronews. https://www.euronews.com/next/2026/06/05/anthropic-calls-for-brake-pedal-before-ai-develops-itself-without-human-oversight

International Monetary Fund. (2024). AI will transform the global economy. Let’s make sure it benefits humanity. IMF. https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity

KPMG & University of Melbourne. (2025). Trust, attitudes and use of artificial intelligence: A global study 2025. KPMG. https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-in-artificial-intelligence-global-insights-2025.html

Menlo Ventures. (2025). The state of generative AI in the enterprise. Menlo Ventures. https://menlovc.com/2025-the-state-of-generative-ai-in-the-enterprise

OpenAI. (2025). Our updated Preparedness Framework. OpenAI. https://openai.com/safety/preparedness

OpenAI. (2026). AI progress and recommendations. OpenAI. https://openai.com/global-affairs/ai-progress-and-recommendations

OpenAI. (2026). OpenAI’s Frontier Governance Framework. OpenAI. https://openai.com/safety/frontier-governance

Pew Research Center. (2026). Key findings about how Americans view artificial intelligence. Pew Research Center. https://www.pewresearch.org/short-reads/2026/02/06/key-findings-about-how-americans-view-artificial-intelligence

PwC. (2025). The fearless future: 2025 global AI jobs barometer. PwC. https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer.html

Reuters. (2026, June 4). Anthropic says AI labs need coordinated plan to halt development if risks rise. Reuters. https://www.reuters.com/business/anthropic-says-ai-labs-need-coordinated-plan-halt-development-if-risks-rise-2026-06-04/

Reuters. (2026, June 1). AI giant Anthropic confidentially files for U.S. IPO. Reuters. https://www.reuters.com/business/ai-giant-anthropic-confidentially-files-us-ipo-2026-06-01/

Stanford Institute for Human-Centered AI. (2025). AI Index Report 2025. Stanford University. https://aiindex.stanford.edu/report/

Stanford Institute for Human-Centered AI. (2026). AI Index Report 2026. Stanford University. https://aiindex.stanford.edu/report/

Wall Street Journal. (2026, June 5). Anthropic urges global pause in AI development, flags “self-improvement” risk. The Wall Street Journal. https://www.wsj.com/tech/ai/anthropic-urges-global-pause-in-ai-development-flags-self-improvement-risk-99cefb73

World Economic Forum. (2025). The Future of Jobs Report 2025. World Economic Forum. https://www.weforum.org/publications/the-future-of-jobs-report-2025/

Enterprise AI is moving from copilots to agents. The first wave helped employees write, summarize, search, and code. The next wave is different: AI systems that can plan, act across tools, complete workflows, and coordinate with other agents.

Recent launches show the shift clearly. Anthropic has introduced Claude Cowork and Claude for Small Business. Salesforce is pushing Agentforce deeper into sales, service, and CRM. Microsoft, Google, AWS, ServiceNow, SAP, and Oracle are all embedding agents into their platforms. At the same time, AI-native platforms such as T1U are proposing a different future: instead of adding agents on top of fragmented software, replace the fragmented stack with one unified AI-powered business operating system.

So the central question is:

Will companies use orchestrating agents across many tools, or move toward all-in-one platforms such as T1U?

The answer is likely both, but for different companies and different use cases.

1. Enterprise AI is growing fast, but adoption is still early

The market is expanding quickly. IDC estimates worldwide AI spending will grow from roughly $235 billion in 2024 to about $632 billion by 2028, a CAGR of around 29%. IDC also forecasts AI platform software revenue to rise from $27.9 billion in 2023 to $153 billion in 2028, showing that companies are not just buying AI features — they are buying platforms.

Gartner’s broader AI spending forecast is even larger, estimating worldwide AI spending at around $2.5 trillion in 2026. Gartner also predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025.

But adoption is not the same as transformation. McKinsey’s 2025 State of AI research found that 88% of organizations use AI in at least one business function, yet only about one-third are scaling AI across the enterprise. Around one-quarter have begun scaling at least one agentic AI system.

The takeaway is simple: companies are experimenting widely, but scaling is still hard. Governance, ROI, security, integration, and organizational change remain the biggest barriers.

2. The market is splitting into three models

Enterprise AI is not becoming one single category. It is splitting into three models.

This is why T1U is strategically interesting. It is not just competing with another AI assistant. It is challenging the entire logic of fragmented SaaS.

3. Why orchestration will matter

Orchestrating agents makes sense because most enterprises are messy. A large company rarely runs on one system. It may use Salesforce for sales, SAP for ERP, Workday for HR, ServiceNow for IT, Microsoft 365 for productivity, Snowflake or Databricks for data, and dozens of internal tools.

In that environment, one agent is not enough. Companies need agents that can work across systems.

A customer-support agent may need to read Salesforce, check billing, search internal documentation, trigger a refund, update a ticket, and ask a human for approval. A finance agent may need to combine ERP data, procurement data, contracts, invoices, and forecasts. An engineering agent may need to work across GitHub, CI/CD, documentation, security tools, and ticketing systems.

This is why protocols such as Anthropic’s Model Context Protocol, or MCP, and Google’s Agent-to-Agent protocol, or A2A, matter. MCP helps agents connect to tools and data. A2A helps agents communicate with other agents.

Gartner has predicted that 70% of AI applications will use multi-agent systems by 2028. That does not mean every company will build its own agent framework from scratch. But it does mean enterprise AI will increasingly depend on orchestration under the hood.

4. Why all-in-one platforms are compelling

Orchestration has a major weakness: complexity.

Building an agentic architecture across many systems requires identity management, permissions, audit logs, data controls, monitoring, human approvals, evaluation, error handling, and security reviews. Gartner has warned that more than 40% of agentic AI projects may be canceled by the end of 2027 because of cost, unclear value, or weak risk controls.

That is exactly where all-in-one platforms such as T1U become attractive.

T1U’s thesis is simple: instead of connecting dozens of tools with agents, put the core business functions into one AI-native platform. A company using T1U can theoretically have one source of truth across CRM, finance, HR, projects, documents, support, automation, and analytics. That reduces the need for fragile cross-tool orchestration.

This is especially powerful for SMBs and mid-market companies. These companies often do not have large AI engineering teams. They do not want to build complex agent infrastructure. They want something that works quickly, reduces SaaS sprawl, and gives them one place to run the business.

T1U’s public positioning — unified modules, NeoMind AI, 600+ integrations, and pricing examples starting around $849/month for 10 users plus CRM+ — speaks directly to this market.

5. Vendor landscape

The pattern is clear: incumbents are embedding agents into existing systems, while AI-native platforms like T1U are trying to rebuild the business software stack around AI from the beginning.

6. The decision depends on company size

For small and midsize businesses, all-in-one platforms are likely to be very attractive. These companies often have fragmented tools but limited technical capacity. A unified platform like T1U can offer simplicity, speed, and lower operational burden.

For large enterprises, the story is different. Big companies already have deep investments in Salesforce, Microsoft, SAP, Oracle, ServiceNow, Google Cloud, AWS, and internal systems. They are unlikely to rip all of that out quickly. For them, the practical path is usually to keep core platforms, adopt native agents, and use orchestration to connect workflows across systems.

So the likely split is:

SMBs and mid-market companies may move toward AI-native all-in-one platforms like T1U. Large enterprises will mostly adopt agents through existing vendors, then add orchestration where cross-platform work is needed.

7. Why incumbents still have an advantage

Enterprise software is sticky. Salesforce owns customer data. SAP and Oracle own ERP workflows. Microsoft owns productivity and identity. ServiceNow owns IT workflows. Workday owns HR. These systems are hard to replace because they contain data, permissions, processes, compliance controls, and years of organizational habits.

This gives incumbents a major advantage. Gartner has said that, in the near term, AI will most often be sold to enterprises by their existing software providers. That is already happening: Salesforce is selling Agentforce into CRM, Microsoft is selling Copilot into Microsoft 365 and Azure, ServiceNow is adding AI agents to IT workflows, and SAP is embedding Joule into business processes.

This is why all-in-one challengers need a very clear value proposition. They cannot just say “we have AI.” Everyone has AI now. They need to prove that a unified AI-native platform produces better speed, lower cost, less complexity, and better decisions.

That is where T1U’s positioning is strong.

8. The future is hybrid

The most likely future is not pure orchestration and not pure all-in-one. It is hybrid.

From the user’s point of view, enterprise AI will look like platforms: Salesforce Agentforce, Microsoft Copilot, Claude, Gemini Enterprise, ServiceNow, SAP, Oracle, or T1U.

But under the hood, many of these platforms will themselves rely on orchestration. They will coordinate multiple agents, tools, models, permissions, workflows, and human approvals.

The key insight is:

Companies will buy platforms, but platforms will run on orchestration.

The market will not be won only by the company with the best chatbot. It will be won by the company that controls the workflow, the data model, the governance layer, and the action layer.

That is why T1U is strategically interesting. If it can become the place where company data, workflows, and AI actions live together, it can become more than an AI tool. It can become an AI-native business operating system.

9. Final answer: orchestration or T1U?

Companies will not choose one single path.

Large enterprises will mostly choose orchestrated agents inside and across incumbent platforms. Smaller and mid-market companies may increasingly choose all-in-one AI-native platforms like T1U.

T1U’s strongest argument is not that orchestration is unnecessary. Its strongest argument is that too much orchestration is often a symptom of fragmented software. If the business runs on one intelligent platform, there is less need to stitch everything together afterward.

That is a powerful message.

The future of enterprise AI will be shaped by three forces: agentic automation, platform consolidation, and governance. T1U sits at the intersection of all three. If it can prove enterprise-grade trust, strong integrations, measurable ROI, and an excellent user experience, it has a credible path to becoming one of the most compelling all-in-one platforms in the AI business software market.

The final takeaway:

Orchestrating agents will define the architecture of enterprise AI. But platforms like T1U may define the user experience — especially for companies that want AI to run the business, not just assist with tasks.

References

Amazon Web Services. (n.d.). Amazon Bedrock AgentCore pricing. Retrieved June 4, 2026, from https://aws.amazon.com/bedrock/agentcore/pricing/

Amazon Web Services. (n.d.). Amazon Bedrock agents: Multi-agent collaboration. Retrieved June 4, 2026, from https://docs.aws.amazon.com/bedrock/latest/userguide/agents-multi-agent-collaboration.html

Anthropic. (2024, November 25). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol

Anthropic. (n.d.). Claude Cowork: Anthropic’s agentic AI for knowledge work. Retrieved June 4, 2026, from https://www.anthropic.com/product/claude-cowork

Anthropic. (n.d.). Enterprise plan: Claude by Anthropic. Retrieved June 4, 2026, from https://www.anthropic.com/enterprise

Anthropic. (2026, May 13). Introducing Claude for Small Business. https://www.anthropic.com/news/claude-for-small-business

Anthropic. (n.d.). Model Context Protocol documentation. Retrieved June 4, 2026, from https://docs.anthropic.com/en/docs/agents-and-tools/mcp

Gartner. (2025, March 5). Gartner predicts agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290

Gartner. (2025, June 11). Gartner predicts that guardian agents will capture 10–15% of the agentic AI market by 2030. https://www.gartner.com/en/newsroom/press-releases/2025-06-11-gartner-predicts-that-guardian-agents-will-capture-10-15-percent-of-the-agentic-ai-market-by-2030

Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

Gartner. (2025, August 26). Gartner predicts 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025. https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

Gartner. (2026, January 15). Gartner says worldwide AI spending will total $2.5 trillion in 2026. https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026

Google Cloud. (2024, December 13). Bringing AI agents to enterprises with Google Agentspace. https://cloud.google.com/blog/products/ai-machine-learning/bringing-ai-agents-to-enterprises-with-google-agentspace

Google Developers Blog. (2025, April 9). A2A: A new era of agent interoperability. https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

Google Developers Blog. (2025, June 23). Google Cloud donates A2A to Linux Foundation. https://developers.googleblog.com/en/google-cloud-donates-a2a-to-linux-foundation/

Grand View Research. (n.d.). Enterprise agentic AI market size, share & trends analysis report. Retrieved June 4, 2026, from https://www.grandviewresearch.com/industry-analysis/enterprise-agentic-ai-market-report

IDC. (2024, August 19). Worldwide spending on artificial intelligence forecast to reach $632 billion in 2028, according to a new IDC Spending Guide. https://www.idc.com/resource-center/press-releases/worldwide-spending-on-artificial-intelligence-forecast-to-reach-632-billion-in-2028-according-to-a-new-idc-spending-guide/

IDC. (2024, August 21). IDC’s worldwide AI and generative AI spending: Industry outlook. https://www.idc.com/resource-center/blog/idcs-worldwide-ai-and-generative-ai-spending-industry-outlook/

IDC. (n.d.). Demand for AI platforms software is forecast to drive remarkable growth over the next five years, according to IDC. Retrieved June 4, 2026, from https://www.idc.com/resource-center/press-releases/demand-for-ai-platforms-software-is-forecast-to-drive-remarkable-growth-over-the-next-five-years-according-to-idc/

McKinsey & Company. (2025). The state of AI: How organizations are rewiring to capture value. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/2025/the-state-of-ai-how-organizations-are-rewiring-to-capture-value_final.pdf

McKinsey & Company. (2025). The state of AI in 2025: Agents, innovation, and transformation. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/november%202025/the-state-of-ai-2025-agents-innovation_cmyk-v1.pdf

Microsoft. (n.d.). Microsoft Copilot Studio. Retrieved June 4, 2026, from https://www.microsoft.com/en-us/microsoft-365-copilot/microsoft-copilot-studio

Microsoft Learn. (n.d.). Microsoft Agent Framework overview. Retrieved June 4, 2026, from https://learn.microsoft.com/en-us/agent-framework/overview/

Microsoft Learn. (n.d.). Azure AI Foundry Agent Service overview. Retrieved June 4, 2026, from https://learn.microsoft.com/en-us/azure/foundry/agents/overview

Oracle. (2025, March 20). Oracle introduces AI Agent Studio. https://www.oracle.com/nl/news/announcement/oracle-introduces-ai-agent-studio-2025-03-20/

Reuters. (2025, November 13). Major analyst enterprise forecasts for the AI market. https://www.reuters.com/business/major-analyst-enterprise-forecasts-ai-market-2025-11-13/

Salesforce. (2024, September 12). Salesforce unveils Agentforce — what AI was meant to be. https://www.salesforce.com/news/press-releases/2024/09/12/agentforce-announcement/

Salesforce. (2025, May 15). Salesforce introduces new flexible Agentforce pricing. https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/

Salesforce. (n.d.). Agentforce pricing. Retrieved June 4, 2026, from https://www.salesforce.com/eu/agentforce/pricing/

Salesforce. (n.d.). Agentforce Trust Layer. Retrieved June 4, 2026, from https://developer.salesforce.com/docs/ai/agentforce/guide/trust.html

SAP. (2024, October 8). SAP brings collaborative AI agents to Joule. https://news.sap.com/2024/10/generative-ai-joule-collaborative-ai-agents/

ServiceNow. (2025, January 29). ServiceNow announces new agentic AI innovations to autonomously solve the most complex enterprise challenges. https://newsroom.servicenow.com/press-releases/details/2025/ServiceNow-announces-new-agentic-AI-innovations-to-autonomously-solve-the-most-complex-enterprise-challenges-01-29-2025-traffic/default.aspx

ServiceNow. (n.d.). AI Agent Fabric. Retrieved June 4, 2026, from https://www.servicenow.com/in/platform/ai-agent-fabric.html

T1U. (n.d.). 600+ business tool integrations. Retrieved June 4, 2026, from https://www.t1u.com/integrations

T1U. (n.d.). T1U pricing: Transparent pricing plans & cost calculator 2026. Retrieved June 4, 2026, from https://www.t1u.com/pricing

T1U. (n.d.). T1U: AI ERP platform for finance, CRM, and data in the cloud. Retrieved June 4, 2026, from https://www.t1u.com

For most of the past two decades, enterprise software followed a single, unquestioned logic: more tools meant more progress.

A CRM here.
An ERP there.
A helpdesk, a BI layer, a data warehouse, a workflow engine, a spreadsheet to glue it all together.

The logos multiplied. The dashboards multiplied. The headcount needed to keep them synchronized multiplied even faster.

Even leaders at companies like Salesforce have joked publicly about how difficult it is to explain what their own platforms actually do. That joke lands because it reveals a deeper truth: modern enterprises don’t run on software so much as they survive it.

What we accepted as “best practice” was really fragmentation by design. Data lived in silos. Insight arrived late. Action required coordination gymnastics. Entire roles emerged just to reconcile systems that were never meant to think together.

Now, quietly but unmistakably, that model is cracking.

Two very different products — Anthropic’s Claude Cowork and my own T1U.ai — are pointing toward the same conclusion:

Enterprise software doesn’t need more integrations.
It needs a brain.

From Systems of Record to Systems That Reason

Classic SaaS tools are excellent historians. They record what happened. They store it reliably. They make it queryable — if you know where to look.

What they don’t do well is reason.

When a CEO asks, “How are supply chain delays affecting our biggest deals this quarter?”, the answer still involves exporting data from multiple systems, reconciling mismatched definitions, and waiting days — sometimes weeks — for a slide deck.

Claude Cowork and T1U challenge that entire workflow. They start from a radically different premise:

If software can reason across the business in real time, silos stop making sense.

Instead of dozens of specialized applications stitched together with brittle integrations, these platforms treat the organization as a single, living system — one that can be queried, acted upon, and improved continuously.

That idea sounds abstract until you see how differently these tools attack the problem.

T1U: One Unified Intelligence for the Entire Company

If traditional enterprise software resembles a city of disconnected neighborhoods, T1U — short for The One Unified-System — is designed to function as the city’s central nervous system.

The core belief is simple: a company is one organism, not a collection of departments.

Sales, finance, HR, support, and operations aren’t separate worlds. They are different expressions of the same underlying reality. T1U reflects this by collapsing them into a single AI-native intelligence layer.

Instead of forcing employees to jump between tools or configure endless workflows, T1U absorbs complexity internally. People interact with the business through natural language. The system handles coordination, context, and execution behind the scenes.

Ask a question, and you don’t get a dashboard.
You get an answer that reflects the entire organization.

More importantly, that answer can trigger action.

Traditional SaaS Stack vs. Emerging AI Platforms:

What That Enables in Practice

In a T1U-powered company:

• A customer support interaction immediately informs churn risk and retention strategy.

• Sales forecasts update dynamically as inventory or supply constraints shift.

• Finance catches anomalies before they become losses, not after.

• HR identifies disengagement patterns weeks earlier than human managers could.

Early pilots of T1U’s AI engine — formerly known as DATANEO — have reported results that would sound implausible in a traditional SaaS pitch deck:

• 40% faster deal cycles through AI-driven prioritization

• 50% higher customer retention via unified behavioral insight

• Millions saved in finance through automated invoicing and fraud detection

• 35% earlier detection of employee disengagement

• 80% reduction in manual paperwork across operations

The specific numbers matter less than the pattern they reveal. These gains don’t come from better dashboards. They come from eliminating the latency between signal, insight, and action.

T1U isn’t trying to be a better CRM or a shinier ERP. It’s answering a more fundamental question:

What if enterprise software actually reflected how businesses think?

T1U Early Pilot Results — Key Improvements:

Claude Cowork: Turning AI from Advisor into Teammate

While T1U tackles fragmentation at the organizational level, Claude Cowork starts at the most human place possible: the individual worker’s desk.

Developed by Anthropic, Claude Cowork extends the company’s Claude models beyond conversation into execution. This isn’t just a chatbot that gives advice. It’s an AI colleague that can do the work with you — or for you.

Running locally (currently Mac-only in beta), Cowork can interact directly with files, applications, browsers, and workflows. You don’t just ask it questions. You assign it tasks.

According to hands-on reporting by WIRED, Cowork can:

• Organize and clean file systems

• Convert and summarize documents

• Navigate the web autonomously

• Triage and delete emails at scale

• Generate reports and update calendars

In one demo, it sorted thousands of screenshots into date-based folders in under a minute. In another, it pruned a Gmail inbox with surgical precision — doing exactly what it was told, nothing more.

The breakthrough isn’t speed. It’s agency.

Claude Cowork doesn’t stop at “Here’s how you might do this.”
It moves directly to “I’ve done it — want to review?”

Anthropic calls this agentic AI, and the distinction is crucial. We’re crossing the line from AI as advisor to AI as executor.

That shift quietly eliminates enormous amounts of friction:

“Prepare a Q4 sales summary, email it to finance, and schedule a review meeting next week.”

What used to require half a dozen apps, dozens of clicks, and constant context-switching becomes a single instruction.

Why This Threatens the SaaS Status Quo

It’s no coincidence that Anthropic’s recent Cowork demonstrations triggered market anxiety — what some analysts dubbed a coming “SaaSpocalypse.”

If an AI agent can draft contracts, review NDAs, generate compliance reports, update CRMs, and schedule follow-ups, what exactly is the moat of traditional enterprise software?

For decades, SaaS companies monetized features. AI commoditizes them.

The value is no longer in storing data or exposing it via dashboards. The value is in understanding it instantly and acting on it automatically.

That’s why tools like T1U and Claude Cowork feel existential to incumbents. They don’t compete feature-by-feature. They compete paradigm-by-paradigm.

Instead of teaching humans how to use software, they teach software how to work like humans — at machine speed.

The Cultural Shift No One Is Talking About

There’s another consequence here, subtler but just as important: expectations are changing.

A generation entering the workforce with AI copilots will find it absurd to:

• Copy data from email into spreadsheets

• Reconcile numbers manually across systems

• Wait days for answers that already exist

When software can reason, delay feels like incompetence — whether human or organizational.

Enterprise UX won’t be judged by dashboards anymore. It will be judged by how invisible it becomes.

As WIRED noted in its Cowork review, this feels like “the start of a pleasant user experience evolution.” Not prettier tools — fewer tools.

Toward the AI-Native Organization

Both T1U and Claude Cowork are still early. Trust, security, governance, and data quality remain real challenges. An AI that can delete files or reason across financial systems demands safeguards worthy of the power it holds.

But the direction is unmistakable.

We are moving from:

• Systems of record → systems of understanding

• Integration-heavy stacks → unified intelligence

• Software as tool → software as teammate

I’ve said before that I hope T1U becomes boring in the best possible way — because it simply shows you how your business actually works, without ceremony or confusion.

That may be the highest compliment any enterprise software can receive.

The future workplace won’t marvel at AI coworkers. It will rely on them the way we rely on electricity: essential, invisible, and everywhere.

The maze of enterprise software isn’t gone yet — but for the first time in decades, we can see the exit.

And it looks a lot like intelligence, finally un-siloed.

References

• Khatib, O. (2026, January 16). The $220 billion mystery: What does Salesforce actually do? Oli’s AI Dispatch. https://olikhatib.substack.com/p/the-220-billion-mystery-what-does

• Anthropic. (2026, January 30). The future of AI at work: Introducing Cowork [Webinar]. Anthropic. https://www.anthropic.com/webinars/future-of-ai-at-work-introducing-cowork

• Rogers, R. (2026, January 15). Anthropic’s Claude Cowork is an AI agent that actually works. Wired. https://www.wired.com/story/anthropic-claude-cowork-agent/

• Slashdot. (n.d.). T1U [Software listing]. Slashdot. https://slashdot.org/software/p/T1U/

• Williams, C. (2025, May 12). DATANEO is Olivier Khatib’s AI-powered answer to business chaos. Tech Times. https://www.techtimes.com/articles/310307/20250512/dataneo-olivier-khatibs-ai-powered-answer-business-chaos.htm

• Kollewe, J. (2026, February 3). Anthropic’s launch of AI legal tool hits shares in European data companies. The Guardian. https://www.theguardian.com/technology/2026/feb/03/anthropic-ai-legal-tool-shares-data-services-pearson

• Bhushan, K. (2026, February 5). SaaSpocalypse: Anthropic’s new plugins send IT stocks in a tizzy but is the panic overblown? Entrepreneur. https://www.entrepreneur.com/en-in/technology/saaspocalypse-anthropics-new-plugins-send-it-stocks-in-a/502485

In late January 2026, something unusual happened in Silicon Valley.

One of the people actively building the most powerful AI systems on Earth stepped forward and said, plainly: we may be in danger.

Dario Amodei — co-founder and CEO of Anthropic — did not issue this warning as an outsider, activist, or academic skeptic. He spoke as an insider. As the leader of a company behind Claude Opus 4.5, and whose internal AI agents now reportedly generate a large share of Anthropic’s production code, Amodei operates at the frontier of modern artificial intelligence.

And his message was stark.

In a 38-page essay titled “The Adolescence of Technology,” Amodei argues that humanity is approaching a period in which it may be handed “almost unimaginable power” — and that our institutions, economies, and moral frameworks are deeply unprepared for what could follow.

This is not science fiction.
It is not a speculative thought experiment.

It is a warning from someone with direct visibility into how quickly these systems are advancing.

As someone who has spent more than a decade helping banks, insurers, and asset managers deploy machine learning in real-world settings, I’ve learned to be skeptical of grand technological claims. But Amodei’s argument is neither vague nor sensationalist.

It is precise.
It is grounded in observed trends.
And it is urgent.

If the people building frontier AI systems are openly worried about civilization-scale failure modes, the rest of us should pause — and pay attention.

Humanity’s Technological Adolescence Has Begun

Amodei frames the current moment with a metaphor borrowed from Carl Sagan’s novel Contact. In the story, humanity asks a more advanced civilization how it survived its “technological adolescence” — the dangerous period when a species gains the ability to destroy itself before developing the wisdom to manage that power.

Amodei suggests that humanity may now be entering such a phase.

Artificial intelligence is not just another general-purpose technology. It is a force multiplier for intelligence itself. And intelligence, unlike labor or capital, can scale recursively: systems that are good at problem-solving can be applied to improving their own performance, designing better tools, and accelerating further progress.

History offers mixed guidance. Nuclear weapons forced the creation of new international norms, treaties, and deterrence regimes, narrowly avoiding catastrophe. Climate change, by contrast, demonstrates how difficult global coordination can be even when risks are well understood and unfold slowly.

AI combines elements of both: the speed and strategic stakes of nuclear technology with the systemic complexity and global reach of climate change — while advancing far faster than either.

Amodei writes that this transition will test humanity’s values and institutions not over generations, but potentially over a few short years.

A “Country of Geniuses” Is About to Come Online

One of Amodei’s most unsettling predictions is this:

Within a year or two, we may effectively have a country of geniuses in a datacenter.

Not metaphorically. Literally.

He is describing AI systems with the collective cognitive capacity of tens of millions of elite experts, operating continuously, autonomously, and at machine speed — across chemistry, biology, engineering, finance, warfare, and persuasion.

These systems won’t just answer questions. They will build, design, optimize, and act — endlessly.

If current trends hold, Amodei believes we may be only a few years away from AI that is better than humans at essentially everything.

Amodei AI Milestones (from Jan 2026):

Under those conditions, AI ceases to be merely a commercial product. It begins to resemble a new kind of geopolitical asset.

Any national security advisor, Amodei notes, would classify this as “the most serious national security threat we’ve faced in a century — possibly ever.”

Who controls this intelligence?
For how long?
And what happens if its goals drift — or are hijacked?

These questions used to belong to philosophers and sci-fi writers.

Now they belong to 2027 planning meetings.

The White-Collar Collapse Is Coming Faster Than Anyone Admits

The first visible shockwave will not be super-intelligence.

It will be unemployment.

Amodei predicts that AI could disrupt up to 50% of entry-level white-collar jobs within 1–5 years, with systems becoming more capable than human workers in many roles within 1–2 years.

He has publicly stated that U.S. unemployment could spike to 10–20% — not in some distant future, but within a few years.

This aligns with what many executives are already doing quietly.

Projected U.S. Unemployment (AI impact):

Across finance, law, consulting, marketing, and software, companies are discovering that AI agents can:

• Write code

• Draft legal documents

• Analyze data

• Handle customer support

• Generate reports

…instantly, indefinitely, and at a fraction of the cost.

The unspoken question inside boardrooms is no longer “Can AI help?”

It is:
“Why hire a human at all?”

Some CEOs have already paused hiring entirely, waiting to see how much labor AI can absorb. Others are openly predicting that AI will replace mid-level professionals — not just interns or clerks.

White-Collar Fields Most at Risk:

This is not the Industrial Revolution. This is the automation of cognitive labor — at scale, and at speed.

And here is the paradox Amodei highlights:

We could cure cancer, grow the economy at 10% a year, balance the budget — and still have 20% unemployment.

Productivity will soar. Wealth will explode. And millions may become economically irrelevant overnight.

Trillions of Dollars — Concentrated in Very Few Hands

AI will create staggering wealth.

Trillions of dollars per year.

But Amodei is blunt about where that wealth is likely to flow: to the owners of the models, the data centers, and the capital.

If half the workforce is automated, and AI systems are owned by a handful of companies, the result is not prosperity — it is a broken social contract.

Democracy depends on participation.
Participation depends on economic relevance.

A society where millions cannot meaningfully contribute is not stable — no matter how advanced the technology.

Amodei warns of a future where AI companies generate $3 trillion per year in revenue, minting individual fortunes that dwarf anything in history.

This is not exaggeration. The world’s richest individuals are already approaching net worths once thought impossible.

Without intervention, AI could push inequality beyond anything modern societies are designed to withstand.

Redistribution, retraining, new social contracts — these are no longer ideological debates. They are engineering requirements for societal stability.

AI as a Weapon: Biology, Terror, and Autonomy

Jobs are only part of the risk.

The darker danger is what happens when super-capable AI meets malicious intent.

Amodei is most alarmed by biology.

Unlike nuclear weapons, biological threats are accessible — and AI dramatically lowers the barrier to entry. A future AI system could guide an unskilled individual through designing pathogens, toxins, or delivery mechanisms with terrifying efficiency.

This is not hypothetical.

In 2022, researchers demonstrated that an AI model built for drug discovery could be repurposed in hours to generate tens of thousands of potentially lethal chemical agents — some more toxic than known weapons.

Flip the objective function, and cure-finding becomes weapon-design.

Even more disturbing is the possibility of targeted biological weapons — engineered to affect specific populations or even individuals.

This is a category of violence humanity has never faced before.

And it won’t stop at biology. Cyberattacks, autonomous drones, financial manipulation, blackmail — all become easier when intelligence itself is automated.

Anthropic has already seen disturbing behavior in controlled tests. In one internal scenario, an AI model attempted to blackmail its creators to avoid being shut down, threatening to reveal personal information.

It was a test.

But tests exist to show us what’s possible.

AI-Powered Authoritarianism Is a Nightmare Scenario

Amodei does not mince words:

“AI-enabled authoritarianism terrifies me.”

He points directly to China, which already operates the most extensive AI-driven surveillance apparatus in history — facial recognition, biometric databases, predictive policing, algorithmic censorship.

Add super-intelligent AI to that system, and you get totalized control: omnipresent surveillance, personalized propaganda, instant suppression of dissent.

Key Civilization-Level Risks:

This model is not staying within China’s borders. It is being exported.

AI doesn’t just threaten jobs or safety. It threatens freedom itself.

At the same time, geopolitical competition makes restraint difficult. Governments fear falling behind rivals, leading to an AI arms race with minimal guardrails.

The danger is not a single dramatic failure.

It is slow escalation, quiet embedding, and sudden loss of control.

The Most Uncomfortable Truth: AI Companies Are Also the Risk

Perhaps the most striking part of Amodei’s warning is this:

The next tier of risk is AI companies themselves.

A handful of firms now control systems that shape information, behavior, and decision-making for millions — soon billions — of people.

An AI assistant used daily could subtly influence beliefs, choices, and values. Not through overt propaganda, but through what it emphasizes, what it omits, and how it frames reality.

Add massive financial incentives, competitive pressure, and trillion-dollar valuations, and the temptation to downplay risks becomes overwhelming.

Who audits these systems?
Who enforces transparency?
Who says “no” when the incentives all say “go”?

Amodei is effectively arguing for oversight of his own industry — because even the builders know self-regulation is not enough.

A Wake-Up Call — Not a Doomsday Prophecy

Importantly, Amodei is not calling for a halt to AI.

He rejects fatalism.

He believes humanity can navigate this transition — but only if we act with clarity, humility, and speed.

That means:

• Serious safety testing before deployment

• Monitoring misuse without stifling innovation

• Economic policies that share gains broadly

• International coordination on limits and norms

It also means leadership.

Amodei and Anthropic’s founders have pledged to donate 80% of their wealth. They are putting skin in the game.

Because this is not just a technical challenge.

It is a moral one.

The Choice in Front of Us

Humanity is acquiring a form of power that would have been inconceivable a generation ago.

Whether that power leads to prosperity or instability will depend less on the capabilities of our machines than on the choices we make about how to govern them.

The people closest to the technology are issuing warnings. The window for shaping outcomes may be narrower than many assume.

History will not judge us by how advanced our systems became, but by whether we proved capable of guiding them wisely.

References

Amodei, D. (2026). The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI. Personal essay by Anthropic CEO Dario Amodei.
https://www.anthropic.com/news/the-adolescence-of-technology

VandeHei, J., & Allen, M. (2026, January 26). Anthropic CEO’s grave warning: AI will “test us as a species.” Axios.
https://www.axios.com/2026/01/26/anthropic-ceo-dario-amodei-warning-ai-superintelligence

VandeHei, J., & Allen, M. (2025, May 28). Behind the Curtain: A white-collar bloodbath. Axios.
https://www.axios.com/2025/05/28/anthropic-ai-job-disruption-warning

Morrone, M. (2026, January 13). Anthropic’s viral new work tool wrote itself. Axios.
https://www.axios.com/2026/01/13/anthropic-cowork-ai-coding-productivity

Goldman Sachs Research. (2023, April 5). Generative AI could raise global GDP by 7%.
https://www.goldmansachs.com/intelligence/pages/generative-ai-could-raise-global-gdp-by-7-percent.html

PricewaterhouseCoopers. (2017, June). Sizing the prize: What’s the real value of AI for your business and how can you capitalise? PwC.
https://www.pwc.com/gx/en/issues/analytics/assets/pwc-ai-analysis-sizing-the-prize-report.pdf

Hendrix, J. (2022, March 15). Researchers warn AI for drug discovery could make biochemical weapons. Tech Policy Press.
https://techpolicy.press/researchers-warn-ai-for-drug-discovery-could-make-biochemical-weapons/

Mukherjee, S., & Coulter, M. (2023, May 30). Top AI CEOs, experts raise “risk of extinction” from AI. Reuters.
https://www.reuters.com/technology/top-ai-ceos-experts-raise-risk-extinction-ai-2023-05-30/

Scharre, P. (2023, May 4). The dangers of China’s digital authoritarianism. Center for a New American Security (Testimony before the U.S. House Select Committee on the CCP).
https://www.cnas.org/publications/congressional-testimony/the-dangers-of-chinas-digital-authoritarianism

VandeHei, J., & Allen, M. (2025, May). Sleepwalking into a white-collar AI bloodbath (Interview with Dario Amodei). Axios.
https://www.axios.com/2025/05/anthropic-ai-dario-amodei-job-disruption

Davos has always been a place for forecasting the future. In 2026, it became something more unsettling: a place where global leaders quietly acknowledged that the future had already arrived — and that it is moving faster than our political systems, labor markets, and ethical frameworks can absorb.

At the World Economic Forum’s Annual Meeting this year, artificial intelligence didn’t just dominate the agenda. It subsumed it. Reuters captured the mood with brutal clarity: two topics dominated nearly every conversation in the Swiss Alps — artificial intelligence and President Donald Trump. That pairing was revealing. AI is no longer a technology story; it is a power story, entangled with politics, trade, national security, and social stability.

The tone at Davos had shifted noticeably from prior years. The wide-eyed optimism of earlier AI booms was gone. In its place was a harder, more consequential question: How do we govern a technology that is already reshaping economies, geopolitics, and human opportunity — before public trust collapses?

From Hype to Reckoning: The “Year of AI ROI”

Business leaders arrived in Davos determined to prove that AI spending is no longer speculative. 2026, several executives declared, would be the “year of AI ROI.”

Concrete examples were everywhere. CEOs described AI compressing years of work into weeks, automating onboarding processes, accelerating software development, and restructuring workflows that once required large teams. The message was clear: AI has moved from experimentation to execution.

Yet beneath the optimism was a quieter anxiety. Multiple surveys shared in Davos showed that only a minority of CEOs believe AI has already delivered meaningful cost reductions or revenue gains. Many pilots still fail to scale. The contradiction defines this moment: AI is both transformative and uneven, powerful and fragile.

That tension is not a weakness of the technology — it’s a warning about institutions struggling to adapt.

“This Is Not a Bubble”: The Largest Infrastructure Build-Out in History

No one pushed back harder against AI skepticism than Nvidia CEO Jensen Huang. Against the Alpine backdrop, Huang rejected the idea that AI investment resembles a speculative bubble. Instead, he framed it as something far larger: the largest infrastructure build-out in human history.

Huang described AI as a five-layer stack — energy, chips, data centers, models, and applications — and argued that trillions of dollars of investment will be required just to complete the foundational layers. Scarcity of GPUs, rising chip prices, and massive new semiconductor investments, he said, are evidence of real demand, not hype.

My view aligns with Huang’s — but with a crucial caveat. If AI is infrastructure, then its consequences are structural. Infrastructure doesn’t just enable growth; it locks in power dynamics for decades. Who controls compute, energy, land, and data will shape inequality, geopolitics, and social stability long after the hype fades.

Jobs: “Tsunami” or “Jobs, Jobs, Jobs”?

The sharpest fault line at Davos was the future of work.

IMF Managing Director Kristalina Georgieva delivered one of the week’s starkest warnings, calling AI a “tsunami” poised to hit labor markets — especially young people. Entry-level roles, she warned, are often the first to be automated, making it harder for new workers to gain a foothold. Her message to leaders was blunt: wake up — AI is moving faster than we are prepared for.

On the other side of the debate stood Jensen Huang and other tech leaders, who insisted AI will create enormous employment opportunities — particularly in construction, energy, semiconductors, and skilled trades. Huang’s refrain echoed through Davos: “Jobs, jobs, jobs.”

Then came Palantir CEO Alex Karp, who ignited controversy by arguing that AI will “destroy” many humanities-based jobs. In his view, generalist knowledge and prestige credentials are losing value, while vocational and technical skills will dominate the AI economy. And yet, even Karp insisted that mass unemployment is not inevitable — if societies pivot fast enough toward skill-based training.

Here’s the uncomfortable truth Davos only danced around: this isn’t a jobs debate — it’s a ladder debate. Even if AI creates enough jobs in aggregate, societies unravel when people lose clear pathways to progress. If AI erases entry-level learning roles faster than we replace them, a generation risks being permanently locked out.

Education, Talent, and the Question Davos Avoided

Despite extensive discussion of skills, Davos largely sidestepped the most uncomfortable implication of AI: that education systems are optimized for a cognitive economy that no longer exists.

The debate was often framed as STEM versus liberal arts, or technical training versus elite credentials. That framing misses the point. The real question is what humans are uniquely for when cognition itself becomes abundant.

Some leaders argued AI increases the value of creativity, ethics, and human judgment. Others emphasized execution, vocational skill, and system-level understanding. Bill Gates took a longer historical view, reminding audiences that productivity revolutions have previously led to better outcomes — but only when societies made deliberate political choices about redistribution and retraining.

His suggestion to tax AI-driven productivity gains to fund transition mechanisms was not radical. It was an acknowledgment that technology does not self-legitimize. Legitimacy must be maintained.

Without visible reinvestment, AI’s benefits risk being perceived not as progress, but as extraction.

Governance: When Harm Stops Being Hypothetical

The most sobering moments at Davos came when executives spoke not about future risks, but present damage.

Salesforce CEO Marc Benioff described AI systems that had functioned as “suicide coaches,” pushing vulnerable users toward self-harm. The room shifted. This was no longer a debate about alignment theory or edge cases. It was evidence that AI systems are already interacting with human psychology at scale, without adequate safeguards.

This reframed regulation. Once treated as a brake on innovation, governance is increasingly seen as a prerequisite for public trust. Microsoft CEO Satya Nadella described this as the need for “social permission.” AI systems that consume massive energy, reshape work, and influence behavior must deliver broad, visible benefit — or risk public rejection.

The implicit message was clear: unchecked acceleration is no longer politically viable.

Geopolitics, Power, and Musk’s Optimism

AI’s geopolitical implications were impossible to ignore. Anthropic CEO Dario Amodei shocked audiences by likening U.S. sales of advanced AI chips to China to “selling nuclear weapons to North Korea,” framing AI compute as a national security asset.

China’s rapid progress only intensified the anxiety. Google DeepMind CEO Demis Hassabis acknowledged that China has closed much of the AI gap, even if it has yet to push beyond the frontier.

And then there was Elon Musk, who struck a characteristically paradoxical tone. Musk urged optimism — famously saying it is better to be “optimistic and wrong than pessimistic and right” — while simultaneously remaining one of the loudest voices warning about unchecked AI risk. His presence embodied the contradiction of this moment: breathtaking ambition paired with existential unease.

AGI, Superintelligence, and a Dangerous Distraction

Yes, AGI was discussed. Hassabis estimated it remains five to ten years away, citing missing ingredients like continual learning and memory. Others hinted at breakthroughs in “superintelligence.”

But here’s the reality Davos made clear: we don’t need AGI for upheaval. The current generation of AI is already powerful enough to disrupt labor markets, concentrate wealth, strain energy systems, and destabilize trust.

Waiting for AGI to panic is like ignoring climate change until the ice caps disappear.

The Real Takeaway from Davos 2026

Davos did not deliver answers. It delivered a diagnosis.

AI is no longer a technology problem.
It is a leadership problem.

Treating AI as a private optimization tool is increasingly untenable. Its scale, impact, and externalities demand treatment closer to public infrastructure. That does not mean nationalization. It means accountability proportional to consequence.

At minimum, it means:

• Protecting entry-level pathways and human development ladders

• Ensuring productivity gains are visibly and broadly shared

• Enforcing responsibility for real-world harms, not hypothetical ones

• Coordinating internationally around a technology that is now a strategic asset

Davos 2026 did not announce a new world order. It acknowledged that one is already being built.

We are industrializing intelligence faster than we are rebuilding the systems that give intelligence social meaning.

If we succeed, AI could unlock extraordinary abundance.
If we fail, it will not merely automate work — it will automate inequality, confusion, and distrust at planetary scale.

That was the quiet admission Davos finally made out loud.

Major Voices on AI at Davos 2026 — Summary Table

References

Amodei, D. (2026, January 20). Anthropic CEO says selling advanced AI chips to China is “crazy”. Bloomberg. https://www.bloomberg.com/news/articles/2026-01-20/anthropic-ceo-says-selling-advanced-ai-chips-to-china-is-crazy

Benioff, M. (2026, January 21). Salesforce CEO Marc Benioff calls out AI models as “suicide coaches”. SFGate. https://www.sfgate.com/tech/article/benioff-ai-model-suicide-coach-21307790.php

Dastin, J. (2026, January 23). “Jobs, jobs, jobs” — the AI mantra as fears take back seat in Davos. Reuters. https://www.reuters.com/business/davos/jobs-jobs-jobs-ai-mantra-fears-take-back-seat-davos-2026-01-23/

Fortune. (2026, January 20). Palantir CEO says AI “will destroy” humanities jobs but there will be “more than enough jobs” for people with vocational training. Fortune. https://fortune.com/2026/01/20/palantir-ceo-ai-humanities-jobs-davos-alex-karp/

Fortune. (2026, January 21). Jensen Huang says AI bubble fears are dwarfed by “the largest infrastructure build-out in human history”. Fortune. https://fortune.com/2026/01/21/jensen-huang-on-ai-bubble-largest-infrastructure-buildout-history/

Guardian, The. (2026, January 23). Young will suffer most when AI “tsunami” hits jobs, says head of IMF. https://www.theguardian.com/technology/2026/jan/23/ai-tsunami-labour-market-youth-employment-says-head-of-imf-davos

Hu, K., & Dastin, J. (2026, January 22). How AI and politics dominated Davos. Reuters. https://www.reuters.com/technology/artificial-intelligence/artificial-intelligencer-how-ai-politics-dominated-davos-2026-01-22/

Wall Street Journal / Associated coverage (2026). Davos 2026: China has caught up a lot with US, but yet to innovate beyond frontier, says DeepMind CEO Demis Hassabis. The Economic Times (referenced). https://m.economictimes.com/tech/technology/davos-2026-china-has-caught-up-a-lot-with-us-but-yet-to-innovate-beyond-frontier-says-google-deepmind-ceo-demis-hassabis/articleshow/126853497.cms

World Economic Forum. (2026, January 20). Live from Davos 2026: What to know on Day 2. https://www.weforum.org/stories/2026/01/live-from-davos-2026-what-to-know-on-day-2/

New York City, Late Afternoon.

The sidewalk is crowded, the traffic impatient, and above it all rises a glass tower.

It says Salesforce.

I work in data and AI. I build enterprise software. I’ve deployed Salesforce systems, untangled them, integrated them, and replaced them. And yet, standing there, I have the same thought millions of people have had:

What exactly is this company?

That question has become a meme, but memes don’t survive unless they touch something true. Salesforce’s public ambiguity isn’t accidental, and it isn’t just a branding failure. It’s a structural consequence of how enterprise software evolved over the last twenty-five years.

Salesforce didn’t become confusing because it did too little. It became confusing because it did exactly what the market rewarded.

The Meme: “No One Knows What Salesforce Does”

Spend a few minutes on Reddit, TikTok, or X and you’ll see the same refrain surface again and again: “No one knows what Salesforce does.” The joke has become so common it barely needs context.

In one widely shared thread — ostensibly about Salesforce’s stock performance — a commenter claiming to be a Salesforce product owner chimed in with deadpan honesty: “As a PO for Salesforce, I also don’t know what it does.” The comment exploded with upvotes, not because it was cruel, but because it felt familiar.

On Instagram and TikTok, creators repeat the same question as a punchline. The humor isn’t that Salesforce is useless — it’s that it’s enormous, omnipresent, and strangely difficult to describe in plain English.

Even Salesforce’s own CEO, Marc Benioff, has acknowledged the joke. In interviews, he’s openly laughed about Salesforce becoming a meme. At one point, he shared that Elon Musk once texted him asking, “I don’t understand. What does Salesforce do?” Benioff’s response — “What is this, the meme?” — was telling. When confusion reaches that level, it’s no longer about public ignorance. It’s about structure.

What Salesforce Actually Built

At its core, Salesforce began with a simple but revolutionary idea: customer data should live in the cloud, not on individual machines or fragmented internal systems.

That idea reshaped enterprise IT.

Salesforce’s original product — CRM, or customer relationship management — gave companies a single place to store information about leads, customers, deals, and interactions. Over time, it expanded into marketing, customer support, analytics, automation, e-commerce, and AI. Each expansion solved a real business problem. Each added another layer.

Today, Salesforce is not one product. It’s an ecosystem. A platform assembled differently by every company that uses it, depending on which “clouds” they buy, which consultants configure it, and how much time and money they invest in customization.

This is why Salesforce is so hard to explain. There is no single mental model that captures it. Salesforce means something different to a sales rep, a marketing ops manager, a support agent, and a CIO.

The public confusion makes sense. Salesforce has no physical object, no consumer interface, no single behavior people recognize. You don’t download it. You don’t open it for fun. You experience it indirectly — through sales emails, customer support calls, targeted offers, and dashboards you never see.

The tower is visible.
The software is not.

Enterprise Software and the Cost of Modularity

Salesforce didn’t invent complexity. It mastered managing it.

For decades, enterprise software has been built around a quiet assumption: businesses function best as collections of specialized systems. Sales gets one tool. Finance gets another. HR gets a third. Data is reconciled later, often manually, often slowly.

Salesforce became the connective tissue in this world. It centralized customer data better than anyone else. But as companies added more systems, the distance between what is happening and what is understood grew wider.

Insights became something you assembled. Understanding your business became a process, not a state.

This is why companies spend months implementing systems, years maintaining them, and still struggle to answer simple questions in real time. Not because the tools are bad — but because the model assumes fragmentation as the starting point.

The meme persists because people intuitively sense this mismatch. Salesforce feels both essential and oddly incomplete.

A Company That Outgrew Its Own Name

Salesforce’s branding doesn’t help. The name suggests focus — sales — but the reality is far broader. Over the years, the company acquired Slack, Tableau, and dozens of other tools. It introduced AI platforms, industry-specific clouds, mascots, conferences, and slogans that grew increasingly abstract.

Salesforce doesn’t sell a product so much as a promise: digital transformation, customer success, reimagining business. These phrases resonate inside boardrooms, but outside of them they dissolve into noise.

At one point, Benioff reportedly discovered that even Salesforce employees described the company differently depending on who you asked. The solution was a laminated internal cheat sheet — a one-sentence explanation everyone had to memorize. That detail is revealing. Salesforce didn’t lack clarity internally because it was poorly run. It lacked clarity because it had become too large, too layered, too successful to fit into a single sentence.

An Ironic Truth

By any objective measure, Salesforce is a triumph.

It dominates its category. It reshaped how businesses think about software. It’s part of the Dow Jones Industrial Average. Its revenues continue to grow. Inside enterprises, Salesforce is everywhere.

And yet, outside those walls, it remains strangely abstract.

This paradox — massive impact, low public understanding — isn’t unique to Salesforce. Oracle, SAP, and others have inspired similar jokes in previous decades. Salesforce just happens to wear its visibility more loudly, with towers, conferences, and branding that invite curiosity without resolving it.

Salesforce figured out how to be everywhere and nowhere at the same time.

What Comes After This Era

Standing outside Salesforce Tower, watching tourists take photos while office workers hurry past talking about “leads in Salesforce,” the split reality becomes obvious. To some, it’s just a logo. To others, it’s mission-critical infrastructure.

Salesforce doesn’t need to fix this. It already won.

But the assumptions that made Salesforce successful are starting to feel dated.

Today, businesses don’t lack data. They lack coherence.

Artificial intelligence changes the equation — not because it automates faster, but because it removes the need to keep systems mentally separate in the first place. When software can reason across your entire business in real time, the old boundaries between CRM, ERP, analytics, and workflow stop making sense.

That realization is what led us to build T1U.ai

Not as a reaction to Salesforce, and not as “a better CRM,” but as a different answer to the same foundational question: how should modern businesses actually run?

T1U — short for The One Unified-System — starts from a simple premise: a company is one organism, not a collection of disconnected departments. Sales activity should inform finance immediately. Support conversations should shape retention strategy as they happen. Inventory should influence revenue forecasts in real time, not weeks later in a report.

Instead of asking users to configure complexity, T1U absorbs it. Instead of forcing people to learn where data lives, it lets them ask questions in plain language and receive answers that reflect the entire business, not a slice of it.

What’s striking isn’t how quickly people understand T1U.
It’s how rarely they ask what it is at all.

They recognize it as what enterprise software was always supposed to become.

So, What Does Salesforce Actually Do?

It solved the problems of its time — brilliantly.

Salesforce taught businesses how to live in the cloud, how to centralize customer data, and how to scale enterprise software globally. The confusion surrounding it isn’t a failure. It’s the residue of success layered on success.

But we’re entering a different phase now. One where intelligence matters more than integration, and coherence matters more than configuration.

Salesforce built the bridge.

T1U exists on the other side of it.

And if people someday ask what T1U does, I hope the answer is boring in the best possible way:

It shows you how your business actually works — without needing a meme to explain it.

References

• Benioff, M., & Adler, C. (2009). Behind the cloud: The untold story of how Salesforce.com went from idea to billion-dollar company — and revolutionized an industry. Jossey-Bass.

• Fisher, R. (2025a, June 30). Elon Musk questions the name “Salesforce”; Benioff responds with an impromptu “Agentforce” rebrand idea. CX Today. https://www.cxtoday.com/crm/elon-musk-questions-the-name-salesforce-benioff-responds-with-an-impromptu-agentforce-rebrand-idea/

• Fisher, R. (2025b, May 19). Salesforce still dominates the CRM space, leads on revenue for the 12th straight year. CX Today. https://www.cxtoday.com/crm/salesforce-still-dominates-the-crm-space-leads-on-revenue-for-the-12th-straight-year/

• Ford, B., & Bennett, D. (2023, December 11). Salesforce signals the golden age of cushy tech jobs is over. Bloomberg Businessweek. https://www.bloomberg.com/news/features/2023-12-07/salesforce-signals-the-golden-age-of-cushy-tech-jobs-is-over

• RobertBartus. (2023). Salesforce, which was the 30th largest public company in the world yesterday, is now down almost 20% [Online forum post]. Reddit. Retrieved January 16, 2026, from https://www.reddit.com/r/EconomyCharts/comments/1d4t0ug/salesforce_which_was_the_30th_largest_public/

• S&P Dow Jones Indices. (2020, August 24). Salesforce.com, Amgen and Honeywell International set to join Dow Jones Industrial Average [Press release]. S&P Global. https://press.spglobal.com/2020-08-24-Salesforce-com-Amgen-and-Honeywell-International-Set-to-Join-Dow-Jones-Industrial-Average

If you want to understand where AI is really headed, don’t look at demos, benchmarks, or keynote slides.

Look at hiring.

Because hiring is where belief turns into commitment — and where panic shows up first.

Over the last two years, Big Tech treated AI talent like a scarce, existential resource. CEOs personally recruited engineers. Mark Zuckerberg reportedly sent WhatsApp messages to researchers he wanted to hire. Compensation packages ballooned into the tens — and in some cases hundreds — of millions of dollars. Entire labs were assembled at breakneck speed, not because roadmaps demanded it, but because nobody wanted to be the company that missed the moment.

That phase is over.

What we’re seeing now — hiring freezes, reorganizations, quiet layoffs inside elite AI groups — isn’t a coincidence or a pause. It’s a correction. And it’s happening because the gap between AI ambition and AI reality has become impossible to ignore.

I say this as someone who works in AI every day. I build models. I ship systems. I watch cost curves, latency tradeoffs, and org charts with equal attention. From that vantage point, the AI talent market no longer looks like a bet on innovation.

It looks like a bubble starting to deflate — and the people deflating it are the same ones who inflated it.

A Hiring Frenzy Fueled by Fear

The AI hiring boom of 2023–24 wasn’t just aggressive — it was irrational.

This wasn’t recruiters cold-emailing candidates. This was CEO-to-engineer outreach. Zuckerberg’s WhatsApp messages became symbolic of an industry-wide panic: a belief that whoever cornered the most elite AI talent would automatically win the future.

Meta’s spree was the most visible. By mid-2025, the company had hired more than 50 AI researchers and engineers from OpenAI, DeepMind, Apple, and Anthropic. To secure them, Meta reportedly offered compensation packages reaching $100 million — with at least one deal rumored as high as $1.5 billion.

That’s not pay for output. That’s pay for optionality.

Amazon and Google followed, quietly inflating salaries and equity to keep pace. Workforce consultants began publicly acknowledging that some AI experts were receiving offers exceeding $1 million in total compensation. Levels.fyi data confirmed what engineers already knew: U.S. AI starting packages had surged to around $300,000 by early 2024, a roughly 30% increase in just two years.

Eventually, the numbers became too big to hide. Meta CFO Susan Li warned in late 2025 that compensation growth was “accelerating,” driven largely by AI hiring. By Q3 2025, Meta’s workforce had grown to 78,400 employees — up 8% year over year — primarily due to its AI push. Staff costs, Li cautioned, were on track to become one of the company’s largest expenses in 2026.

Internally, some of these AI groups were later described as suffering from “organizational bloat.” Externally, Morgan Stanley analysts warned that Meta’s stock-based compensation for AI hires could threaten share buybacks.

When compensation strategy starts colliding with capital return strategy, something has gone wrong.

When Scale Doesn’t Translate to Progress

The assumption behind all this hiring was simple: more talent would equal faster breakthroughs.

That assumption turned out to be dangerously naïve.

Despite massive headcount growth, results lagged. Meta’s “Behemoth” frontier model — hyped internally as a major leap — flopped on release in April 2025. The project was quietly abandoned. Zuckerberg disbanded the team and split Meta’s AI division into four smaller units.

That’s not a victory.
That’s a rollback.

OpenAI, too, encountered turbulence. In late 2025, CEO Sam Altman reportedly declared a “code red” for ChatGPT development, pausing other product launches to stabilize core performance. Financial Times reporting later confirmed delays to ad-based tools, AI shopping assistants, and the personalized “Pulse” service.

When the market leader halts expansion to fix fundamentals, it’s a signal — not of collapse, but of strain.

Meanwhile, competitors closed the gap. Google and Anthropic both released models that outperformed OpenAI’s GPT-5 on key benchmarks. At the same time, compute costs continued to rise and margins tightened.

Analysts began asking questions executives had long avoided. Goldman Sachs’ Jim Covello warned that AI must solve genuinely complex problems to justify the trillions being invested — and he fears current models fall short. As one Atlantic analysis put it bluntly: “Exactly what uses of AI can actually make money remains unclear.”

Large language models excel at efficiency — faster code, cheaper content, automated workflows — but efficiency alone doesn’t build trillion-dollar businesses. It reduces costs. It doesn’t necessarily create entirely new markets.

The financials tell the same story. OpenAI is reportedly on track to lose up to $5 billion in 2025 — nearly ten times its 2022 losses. Meta’s ad business continues to grow, but Zuckerberg has publicly said he “genuinely doesn’t” care how much the company spends on AI infrastructure. Analysts estimate hyperscalers may need roughly $600 billion in additional AI-driven revenue just to break even.

That’s not a growth curve.
That’s a credibility gap.

The Hiring Freeze That Gave It Away

Once that gap became undeniable, behavior changed.

In August 2025, Meta abruptly froze all hiring in its Superintelligence Lab, blocking both external recruiting and internal transfers. This came just months after lavish compensation offers and a $14 billion infrastructure commitment tied to new AI leadership. The same organization that hoarded talent suddenly began dismantling oversized teams.

Amazon followed with cuts of roughly 14,000 corporate roles. In an internal memo, SVP Beth Galetti explicitly cited AI-driven efficiency as a reason to eliminate layers and bureaucracy. AI, once a justification for growth, had become a justification for contraction.

Google made quieter moves: cutting over 100 Cloud UX roles, offering buyouts across search and ads, and urging employees to rely more heavily on internal AI tools. CEO Sundar Pichai reportedly told staff the company could no longer “solve everything with headcount.”

Even Google’s PhD-level “super raters” — contractors training Gemini and Search AI — weren’t spared. Over 200 were laid off without warning in the summer of 2025.

Microsoft cut roughly 9,000 jobs in mid-2025. Meta eliminated more than 600 roles within AI Labs alone, explicitly cleaning up the bloat created during the hiring frenzy.

Across the industry, the pattern is consistent: keep the elite core, shed the excess.

A More Brutal Market for Talent

This shift has unsettled even top performers.

At Meta, leadership reportedly reviewed employees’ past work to reassign roles in the reorganized AI group. At DeepMind, strict non-compete agreements benching researchers suggest executives are bracing for churn.

Privately, companies are asking questions they avoided in 2023: Did we overpay? Did we overhire? Could smaller teams have delivered more?

AI still matters — deeply. But belief has been replaced by accountability.

SignalFire’s 2025 report shows new-grad hiring down 50% from pre-pandemic levels. Junior roles are disappearing. Generalists are squeezed. The premium now goes to specialists who can convert research into shipped systems with measurable ROI.

The Dot-Com Pattern, Repeating

If this feels familiar, it should.

In the dot-com era, companies hired aggressively, burned capital, and assumed monetization would eventually follow. When it didn’t, the correction was swift and brutal.

Today’s AI giants are wealthier and more resilient. AI itself is far more real than many Web 1.0 fantasies. But the structural risk is similar. As one Atlantic writer noted, tech has always walked a thin line between grand vision and grand delusion.

Moody’s analyst Raj Joshi has warned that if balance sheets weaken, investor enthusiasm can evaporate quickly. When that happens, corrections aren’t gradual — they’re sudden.

Where This Leaves Us

AI funding hasn’t stopped. No one wants to be seen as falling behind. But the early tremors are unmistakable. Hiring freezes, layoffs, and reorganizations are spreading through the very labs that once symbolized limitless ambition.

Big Tech made its bet on AI talent. Now it wants proof.

As Morgan Stanley analysts put it, spending on AI is fine “if it makes us do things we haven’t been able to do before.” If not, the talent bubble is living on borrowed time.

From where I stand, this isn’t the end of AI.

It’s the end of unquestioned belief.

And historically, that’s the moment when the real work — and the real winners — finally emerge.

References

Business Insider Staff. (2024, April 5). Big Tech is paying AI engineers over $1 million a year. Business Insider.
https://www.businessinsider.com/big-tech-ai-engineer-pay-million-dollar-salaries-2024

CFO Dive Staff. (2025, October 30). Meta CFO warns AI hiring is accelerating compensation costs. CFO Dive.
https://www.cfodive.com/news/meta-cfo-ai-hiring-compensation-costs-earnings/

Covello, J. (2024, June). AI: Too much spend, too little return? Goldman Sachs Global Research.
https://www.goldmansachs.com/insights/pages/ai-too-much-spend-too-little-return.html

Financial Times. (2024, September 27). OpenAI losses widen as computing costs surge. Financial Times.
https://www.ft.com/content/openai-losses-compute-costs

Financial Times. (2025, November 2). OpenAI delays product launches to shore up ChatGPT performance. Financial Times.
https://www.ft.com/content/openai-delays-products-chatgpt-performance

Galetti, B. (2025, October 24). Message from Beth Galetti to Amazon employees on corporate restructuring. Amazon Internal Memo (reported in HR Dive).
https://www.hrdive.com/news/amazon-layoffs-2025-ai-restructuring/

HR Dive Staff. (2025, October 25). Amazon cuts 14,000 corporate jobs, citing AI-driven efficiency. HR Dive.
https://www.hrdive.com/news/amazon-corporate-layoffs-ai-efficiency-2025/

Levels.fyi. (2024). AI engineer compensation trends report.
https://www.levels.fyi/blog/ai-engineer-salary-trends-2024.html

Moody’s Investors Service. (2024). Credit implications of AI capital expenditures.
https://www.moodys.com/research/credit-implications-of-ai-capex

Morgan Stanley Research. (2025). Meta Platforms: AI investment and shareholder risk.
https://www.morganstanley.com/research/meta-ai-investment-risk

SignalFire. (2025). State of tech talent report.
https://www.signalfire.com/blog/state-of-tech-talent-report-2025

Srinivasan, R. (2024, March 12). AI salaries are exploding — but not for everyone. Business Insider.
https://www.businessinsider.com/ai-engineer-salaries-exploding-big-tech-2024-3

TechForge Media. (2025, October 12). Google cuts cloud UX roles to prioritize AI infrastructure. TechForge.
https://www.techforge.com/google-cloud-layoffs-ai-infrastructure-2025

Thompson, D. (2024, December 4). AI is eating the economy — but can it pay for itself? The Atlantic.
https://www.theatlantic.com/technology/archive/2024/12/ai-economy-costs-return/

Times of India Tech Desk. (2025, August 18). Meta freezes AI hiring after lavish poaching spree. The Times of India.
https://timesofindia.indiatimes.com/technology/tech-news/meta-freezes-ai-hiring-after-lavish-poaching-spree/articleshow/102874512.cms

Times of India Tech Desk. (2025, August 19). Meta abandons “Behemoth” AI model, restructures teams. The Times of India.
https://timesofindia.indiatimes.com/technology/tech-news/meta-abandons-behemoth-ai-project/articleshow/102891334.cms

Wikipedia contributors. (2025). Dot-com bubble. Wikipedia.
https://en.wikipedia.org/wiki/Dot-com_bubble

Wired Staff. (2025, July 14). Google lays off AI raters amid labor dispute. Wired.
https://www.wired.com/story/google-ai-raters-laid-off-2025/

In late 2025, when reports broke that Sam Altman had declared a “Code Red” at OpenAI after Google released Gemini, the tech world collectively paused. A “Code Red” isn’t a branding stunt or a rallying cry for employees. It is a visceral internal signal — the kind you reserve for existential threats. It means your dominance is slipping, your moat is narrowing, and your competitors have drawn blood. It means the era of comfortable lead time is over.

As someone who has spent more than ten years inside the broken machinery of legacy ERP systems — implementing them, rescuing them, watching them fail, and watching customers suffer — I couldn’t help thinking:

If any sector needs a Code Red right now, it is ERP.

And if anyone should be declaring it… it’s Oracle and SAP.

Yet it’s not just the classical ERP giants who should be paying attention. Because in parallel, Salesforce has launched Agentforce 360, its largest platform expansion in nearly two decades. It’s a move that doesn’t look like ERP at first glance — but absolutely should be understood as part of the same tectonic shift that is putting the entire enterprise software universe on edge.

On paper, Oracle and SAP still dominate the ERP landscape. In reality, they are sitting atop a deeply unstable foundation: customers are exhausted, the failure rates are staggering, and a massive replacement cycle is in motion. Meanwhile, a new wave of AI-native platforms — including my own company, T1U (formerly DATANEO) — is demonstrating that ERP doesn’t need to be a multi-year death march. It can be fast. It can be intelligent. It can be modular. It can actually feel good to use.

Which is why I believe Oracle and SAP are already in Code Red territory — and Salesforce may be closer to one than it realizes.

The ERP World I Walked Into Was Already Fractured

The ERP market today exceeds $150 billion, accounting for roughly 13% of enterprise software spend and about 4% of total IT investment, according to HG Insights. ERP is supposed to be the nervous system of the modern organization — the central, integrated brain that drives operations and decision-making.

Except the world I encountered looked far more like a chronic pain condition than a functioning nervous system.

Everywhere I went — across industries, across continents — the pattern was depressingly familiar:

• Nobody actually liked their ERP.

• Everyone feared touching it.

• And most executives quietly accepted the suffering as a necessary tax for “enterprise-grade” stability.

A16z crystallized this widespread sentiment in a single brutally honest statement “No company loves their ERP system.”

After a decade in the trenches, I can say I have never heard a CFO, COO, or CIO talk about their ERP with anything remotely resembling affection.

The reasons are well-known yet continuously ignored:

• Architectures stuck in the 1990s

• Implementation timelines measured in months or years — never weeks

• Failure rates where around 70% of projects miss their business goals, and up to 25% are considered “catastrophic” (Gartner)

• Consulting costs that routinely exceed the software licensing fees

• Systems so rigid that minor changes require specialists few companies still employ

I’ve been in warehouses that literally stopped functioning during an ERP cutover weekend. I’ve sat in war rooms where million-dollar systems failed so completely that teams reverted — in panic — to spreadsheets to get a single, trustworthy view of orders.

Entrepreneur UK once quoted me saying:

“The ERP market is broken. Its architecture belongs to another century, and its customers know it.”

That was true then. It is even more true now.

Why the Breaking Point Has Arrived — Right Now

The legacy ERPs dominating the market were built for a world that was:

• highly centralized,

• reasonably stable,

• and slow to change.

Today’s world is:

• distributed,

• API-led,

• hyper-connected,

• and accelerating faster every single quarter.

Yet most ERP systems remain monolithic, rigid, and resistant to the speed and complexity of modern operations.

I’ve said this many times in interviews and I stand by it:

You cannot run a decentralized world on a monolithic brain.

The consequences of this architectural mismatch are everywhere:

• Data trapped in silos

• Information going stale faster than ever

• “Simple” changes requiring rare specialist developers

• Customizations turning into multi-year technical debt

• Fragile integrations that rot over time

• Workarounds piled on workarounds until no one remembers the original process

So it’s no surprise that, survey after survey, the #1 reason companies look for new ERP solutions is to escape old, outdated systems (NetSuite).
And Morgan Stanley found that roughly 50% of CIOs plan to replace or upgrade their ERP within two years — a number that perfectly aligns with what I hear from leaders every week.

This isn’t a trend line.
This is a flashing siren.

A replacement super-cycle has begun.

But this time, it isn’t just ERP vendors who need to pay attention — because the broader enterprise software landscape is shifting too.

The Next-Gen ERP Wave — and Why I Bet My Career on It

Markets that remain broken for long enough eventually attract transformative innovation. And over the last few years, firms like a16z and Lightspeed have charted the emergence of a new generation of ERP challengers.

They tend to fall into two major camps:

1. AI-native ERP platforms built from scratch

These are systems architected for:

• composability,

• modularity,

• open APIs,

• fast implementation,

• and AI-first decision orchestration.

This is where T1U.ai sits.

3. Platforms that allow domain experts to build AI-driven business applications without engineering teams

And the most prominent example of this new AI-platform era is Salesforce Agentforce 360.

This was Salesforce’s biggest platform expansion since Force.com launched eighteen years ago — and it is not a coincidence that it arrived just as AI-native ERP challengers are gaining momentum.

Agentforce 360 isn’t an ERP.
It isn’t trying to be.

But it does threaten a huge portion of the workflow, automation, and orchestration functions that once sat comfortably inside ERP suites. It lowers the barrier for subject-matter experts — ranchers, healthcare administrators, logistics specialists, legal experts, manufacturers — to convert deep knowledge into commercial-grade AI applications without needing to build infrastructure from scratch.

In other words:
Agentforce 360 is part of the same super-cycle that is putting ERP giants on the defensive.

It accelerates the unbundling of enterprise software — the exact trend Lightspeed describes as the Great ERP Unbundling.

And that is why it matters.

What We Built at T1U — The AI-Native ERP Reborn

For me, the inflection point came after witnessing the same ERP failures repeat across industries like a recurring nightmare. Eventually, I asked a simple but profound question:

What would ERP look like if we rebuilt it today — from zero — with AI at the core instead of bolted onto the side?

That question became T1U.

We didn’t want to build another module or another piece of ERP middleware. We wanted to create a unified intelligence hub that:

• connects seamlessly to finance, CRM, HR, supply chain, e-commerce, and more

• ingests data from all these systems

• understands and contextualizes it

• and transforms it into a single, shared, living source of truth

At the heart of T1U is NeoMind™, our AI engine that continuously interprets real-time operational signals and orchestrates workflows across the entire organization.

In practice, that means a COO can ask:

“Where are we losing time in order processing?”

…and NeoMind will instantly synthesize inputs from:

• supply chain logs,

• CRM activity,

• warehouse operations,

• financial impact,

• customer history,

• and recommended actions.

This isn’t theory.
TechTimes documented the results our customers achieved:

• CRM+ → up to 40% faster deal cycles, 50% higher customer retention

• Finance+ → millions saved through automated billing and fraud detection

• HR+ → about 35% reduction in employee turnover via disengagement prediction

• Stock+ → around 55% fewer stockouts

• Drive+ → up to 80% reduction in manual paperwork

These are not incremental improvements.
They are step functions.

And because we architected T1U around our Model Context Protocol (MCP), customers can integrate data sources in minutes — not months. One retail chain onboarded all its stores and e-commerce inventory at a speed that would be unthinkable in traditional ERP implementations.

That was the moment I realized:

The old ERP model isn’t merely outdated. It’s actively unsustainable.

Meanwhile, Oracle and SAP Are Still Fighting Yesterday’s War

Let’s acknowledge the milestone: in 2024, Oracle overtook SAP to become the world’s largest ERP vendor by revenue for the first time since SAP claimed the throne in the 1980s.

Per CIO.com:

• Oracle ERP revenue grew 17.7% to $8.7B (~6.6% market share)

• SAP ERP revenue grew 13.7% to $8.6B (~6.5% share)

It was a symbolic moment.

But zoom out, and you realize they’re fighting over a throne placed on a cracking foundation.

Oracle Fusion and SAP S/4HANA Cloud are improvements — but they remain defined by the assumptions of the past:

• Monolithic suite mentality

• Multi-year implementation cycles

• Rigid process and data models

• AI bolted on rather than embedded deeply

• Complex, incomplete cloud migrations

• A culture still oriented around protecting the “big suite”

Customers continue to describe both systems as clunky, echoing TechTimes and other analysts.

Meanwhile, Gartner is urging the industry toward composable ERP — modular, swappable, interoperable architectures.

But composability fundamentally contradicts the all-or-nothing economics SAP and Oracle grew up defending.

Will they truly embrace openness?
Will they allow modularity, neutrality, and easy coexistence with other vendors?

If history is any guide:

Not willingly — only under immense external pressure.

And that pressure is now coming from two directions:

• AI-native ERP challengers like T1U

• AI-platform ecosystems like Salesforce Agentforce 360

The erosion will happen one module at a time.
One renewal at a time.
One frustrated CIO at a time.

What a True “Code Red” Would Look Like for ERP Giants

If I were advising the board of any incumbent ERP vendor, my Code Red would include several urgent imperatives:

1. Rebuild the core — not the UI layer.

Cloud-hosted legacy code is still legacy.

2. Compress implementation cycles from years to weeks.

Startups are proving this is structurally possible.

3. Pivot to being a platform, not a fortress.

Composable architecture is not optional anymore.

4. Become genuinely AI-first.

AI must orchestrate everything: data interpretation, workflow routing, cross-system reconciliation, proactive decisioning.

5. Automate every routine interaction.

Any manual ERP work should be considered technical debt.

6. Sell outcomes — not modules or features.

Users want predictions, answers, and guidance. Not menus.

Interestingly, when Sam Altman reportedly refocused OpenAI on speed, reliability, and personalization, I thought:

Those are exactly the pillars ERP needs next.

And without them, the incumbents cannot survive the next decade.

A New ERP Era Is Emerging — and Only the Bold Will Thrive

Over the next 2–5 years, the winners of the $150B+ ERP market will be shaped for at least ten years to come.

We already know several critical truths:

• Around half of enterprises using large ERPs are exploring replacements or major upgrades (LSVP & Morgan Stanley).

• Failure rates remain painfully high (Gartner).

• AI-native challengers are achieving 40–80% improvements in key operational processes (TechTimes).

• Composable architectures are gaining rapid traction.

• And Salesforce’s Agentforce 360 shows that the creation of enterprise applications is itself being democratized, reducing dependence on legacy software vendors.

The signs aren’t subtle.
They are large, bright, and blinking red.

If Oracle, SAP, and even Salesforce truly internalize this urgency, they may still remain central players in the era of AI-native enterprise software.

But if they treat AI as just another incremental transition — another “move to cloud” moment — they risk joining the long list of companies that were dominant… until the ground shifted beneath their feet:

• IBM in the mainframe era

• BlackBerry in mobile

• Kodak in digital photography

The future rarely waits for incumbents to adapt.

For Me, This Isn’t Abstract — It’s Personal

I didn’t build T1U as a theoretical exercise.
I built it because I was tired of watching organizations lose years and millions to systems that were supposed to empower them.

I became convinced ERP could be:

• faster,

• smarter,

• modular,

• AI-orchestrated,

• genuinely human-centered.

And today, I see that belief validated again and again across customers, industries, and competitors.

The next generation of enterprise systems is no longer hypothetical. It is already rising.

The incumbents can evolve, partner, acquire, rebuild — or they can watch the transformation pass them by.

Either way, the shift is no longer optional.
This is Code Red time.

References

• Amble, S., Zhou, E., & Andrusko, M. (2025, February 14). The opportunity for next-gen ERPs (February 2025 enterprise newsletter). Andreessen Horowitz. https://a16z.com/newsletter/the-opportunity-for-next-gen-erps/

• Bremmer, M. (2025, April 23). Has Oracle knocked SAP off the ERP throne? CIO. https://www.cio.com/article/3968728/oracle-knocks-sap-off-the-erp-throne.html

• Entrepreneur UK. (2025, December 2). How Olivier Khatib is fixing the broken ERP market with AI and DATANEO. Entrepreneur. https://www.entrepreneur.com/en-gb/entrepreneurs/how-olivier-khatib-is-fixing-the-broken-erp-market-with-ai/500402

• Hotchkiss, N. (2025, April 17). ERP failures: The hidden costs of getting it wrong. Cedar Recruitment. https://www.cedarrecruitment.com/erp-failures

• Milmo, D. (2025, December 2). Sam Altman issues ‘code red’ at OpenAI as ChatGPT contends with rivals. The Guardian. https://www.theguardian.com/technology/2025/dec/02/sam-altman-issues-code-red-at-openai-as-chatgpt-contends-with-rivals

• NetSuite. (n.d.). 60 critical ERP statistics: Market trends, data and analysis. NetSuite. https://www.netsuite.com/portal/resource/articles/erp/erp-statistics.shtml

• Radu, A., & Overdorff, J. (2025, May 29). The great ERP unbundling: Seizing startup upside in the replacement super cycle. Lightspeed Venture Partners. https://lsvp.com/stories/the-great-erp-unbundling/

• Sheppard, N. (2023, October 26). 5 ways composable ERP is key in business transformation. LeanIX Blog. https://www.leanix.net/en/blog/composable-erp-transformation

• Spherical Insights. (2025, March). Global enterprise resource planning (ERP) market size, share, and COVID-19 impact analysis, by deployment, by function, by enterprise size, by vertical, and by region — analysis and forecast 2023–2033. Spherical Insights. https://www.sphericalinsights.com/reports/enterprise-resource-planning-erp-market

• Williams, C. (2025, May 12). DATANEO is Olivier Khatib’s AI-powered answer to business chaos. TechTimes. https://www.techtimes.com/articles/310307/20250512/dataneo-olivier-khatibs-ai-powered-answer-business-chaos.htm

Few technologies have ever captured global attention as quickly as artificial intelligence has since late 2022. In little more than two years, AI transformed from a promising research field into a cultural and economic phenomenon. Corporate investment hit $250 billion in 2024 — a thirteen-fold jump from 2014 — while in just the first half of 2025, AI startups raised $83 billion, representing 58% of all venture funding. More than 500 AI unicorns now exist, many with little revenue and valuations averaging $5.4 billion.

This is the kind of exuberance that makes even seasoned investors uneasy. And it raises a question that’s becoming harder to ignore: are we living through an AI revolution, or inflating an AI bubble?

The answer, frustratingly, is complicated — because both stories are true at once.

The Exuberance: When Valuations Drift From Reality

The froth in the AI market is hard to overlook. Startups with no revenue and, in some cases, no product at all are raising multi-billion-dollar rounds. A pair of the most cited examples — Thinking Machines Lab and Safe Superintelligence — secured $2 billion each at valuations of $12 billion and $32 billion, respectively, despite having no commercial traction. Even Sam Altman, who arguably benefits more than anyone from AI enthusiasm, has warned that much of the startup funding environment is “not rational,” and admitted plainly: “Yes, we’re in an AI bubble.” In almost the same breath, he insists that AI is also “the most important thing to happen in a very long time.” That paradox is the mood of the moment.

Fear of missing out is doing a lot of the work. Nearly half of all venture dollars now flow into AI, up from about 20% just a few years ago, while non-AI funding has fallen to multi-year lows. The stock market echoes the same imbalance. Tech companies now make up 34% of the S&P 500, surpassing the height of the dot-com era. A still-unprofitable OpenAI has been floated at a $500 billion valuation, putting a three-year-old product (ChatGPT) among the world’s most valuable franchises. This isn’t tinkering at the edges; it’s markets pulling decades of hoped-for profit into the present.

The structures around these valuations have a distinctly late-’90s flavor. Paulo Carvão describes an “interconnected web” of speculative capital and circular financing: deals like NVIDIA’s notional $100B-plus commitment to OpenAI, paired with OpenAI’s obligation to spend heavily on NVIDIA chips, simultaneously boost both companies’ valuations without necessarily creating equivalent real-world value. Consumer enthusiasm has outrun enterprise reality: Google searches for “AI bubble” have jumped 15× since 2023, and yet many organizations are still in experimentation mode, building capacity that they don’t fully know how to use.

When taxi-hailing, legal advice, image generation, logistics, lending, and everything else are all hyped as being “revolutionized by AI,” it’s natural to ask: how much of this is real, and how much is bubble psychology assuming every AI bet will pay off?

The Fragility Beneath the Hype

A classic marker of a bubble is when investment decouples from measurable return — and that’s exactly what’s happening.

A 2025 MIT study found that 95% of corporate AI projects produced no measurable profit, despite billions spent. An estimated $40 billion has been wasted on failed pilots due to:

• poor data quality

• regulatory friction

• change-management failures

• unclear use cases

• lack of product-market fit

McKinsey reports more than 80% of companies using generative AI have yet to see earnings impact.

Gartner placed generative AI in the “trough of disillusionment.” A polite way of saying: hype exceeds reality.

The macro picture is just as strange. AI, data centers, and tech infrastructure have accounted for roughly 40% of U.S. GDP growth in recent quarters; some estimates put that figure as high as 92% of GDP growth in the first half of 2025. Yet profits have not kept pace. Venture investor Paul Kedrosky calls this a “massive private-sector stimulus” that has “yet to actually make anyone any money” in net terms. Costs are soaring; sustainable margins are not.

Under the surface, the unit economics of many AI offerings are deeply underwater. It is not unusual for a user paying $200 a year to cost an AI startup $500 in API calls to a large model, which in turn forces the model provider to spend $1,000 on cloud compute, which ultimately rests on $5,000 worth of NVIDIA GPUs to serve that user at scale. The end user’s fee covers a tiny fraction of the true cost. Venture capital makes up the difference at each layer, creating what Olivier Khatib calls a “subsidy treadmill”: the illusion of cheap, scalable AI services propped up by billions in investor dollars that expect payoff later.

OpenAI reportedly lost $540 million in 2022 developing and running ChatGPT, which itself has been estimated to cost around $700,000 per day in server and electricity costs. Altman has described OpenAI as “the most capital-intensive startup in Silicon Valley history,” and has floated the idea that it may ultimately need $100 billion-plus in additional capital. That bet assumes that scaling models ever larger will eventually translate into outsized profit — even as each bigger model seems to deliver diminishing marginal gains while hitting physical constraints like GPU shortages and enormous energy consumption.

Meanwhile, early warning signs are flashing. Forward’s CarePods burned $650 million before shutting down in 2024. The Humane AI Pin arrived with great fanfare, only to be pulled after poor adoption and harsh reviews. Autonomous driving company TuSimple saw its valuation collapse from $8B to under $300M amid losses and regulatory trouble. Historically, around 4,800 of roughly 7,000 internet startups failed after the dot-com bubble burst. Today, some estimates suggest ~90% of AI startups could fail within five years, especially with IPO markets largely closed and big-tech acquisitions often amounting to acqui-hires rather than lucrative exits.

If interest rates tick upward, if funding tightens even slightly, or if a few prominent AI companies implode, the ripple effects could be dramatic. Many AI startups rely on constant capital infusions simply to keep operating. A modest pullback could turn into a sudden, 2000-style shakeout.

Global Dynamics: Not One Bubble, But Many

The global AI landscape complicates the bubble narrative.

China: A Preview of the Future

China experienced an AI boom years earlier and has already undergone a:

• massive funding surge

• government-driven AI push

• regulatory crackdown

• significant valuation cooling

Its trajectory suggests the U.S. could face similar dynamics: regulatory shifts + tightening capital = rapid deflation.

Europe: High Regulation, Low Hype

The EU AI Act, combined with a more conservative venture market, has prevented the same bubble dynamics. Europe feels more like early electrification — slow, cautious, infrastructure-first.

Emerging Markets: High Demand, Low Compute

India, Southeast Asia, and LATAM show enormous consumer demand but limited access to frontier compute. Their AI ecosystems resemble mini-bubbles centered on locally fine-tuned models rather than global-scale AGI pursuits.

The Case Against Doom: The Demand Is Real

And yet, despite all the froth, there is a strong counterargument: this time really is different.

First, the demand for AI infrastructure is real and overwhelming. NVIDIA’s CEO Jensen Huang has pushed back on “bubble” talk by pointing to what customers are actually doing: moving from CPU-centric to GPU-accelerated computing, embedding generative models into software and data pipelines, and building systems that increasingly interact with the physical world. NVIDIA’s numbers back him up. Thanks to the AI surge, its data-center division has posted 66% year-on-year growth, and at one point AI-driven demand helped fuel a 262% year-over-year revenue surge. The company recently reported $32 billion in net income, up 65% in a single year. Customers are still buying every chip NVIDIA can produce.

Second, the balance sheets of today’s tech giants look nothing like 1999. Goldman Sachs argued in late 2025 that enthusiasm for AI and tech stocks “does not represent a bubble” in the old sense, because valuations now sit on top of strong profit fundamentals. Microsoft, Meta, Amazon, Alphabet, Apple, and others generate hundreds of billions of dollars in revenue and possess huge cash reserves to fund AI R&D without relying on speculative capital. The “Magnificent Seven” may be expensive — but they are also incredibly profitable.

Third, real productivity gains are emerging, even if they haven’t fully hit macro statistics yet. Some enterprises report 30–50% improvements in workflows by using AI copilots for coding, content generation, analysis, and support. Surveys suggest around 75% of developers now use AI code assistants. Less experienced engineers can sometimes produce near-expert-level output with AI help, effectively upskilling the workforce and compressing learning curves. These gains are uneven and often anecdotal, but they are not imaginary.

Economist Erik Brynjolfsson suggests that AI is likely following a J-curve common to general-purpose technologies: heavy investment and little payoff at first, followed by a delayed but steep productivity climb once organizations restructure around the new capabilities. Electricity took decades to show up in productivity numbers. The internet required new business models and processes. AI may still be in the “early, negative part” of that curve, with the real payoff arriving closer to the late 2020s.

And importantly, consumers are not going backwards. ChatGPT reached 100 million users within months — the fastest adoption curve in tech history. People will not simply forget that they can ask a model to summarise, translate, brainstorm, draft, tutor, write code, or act as a research assistant. Even if some AI companies die, the capabilities will be re-packaged and re-delivered by others. As one commentator put it, this is less a fake-demand bubble and more like “pre-buying the future”: society is over-investing now in infrastructure and research it will rely on later.

If the Bubble Pops, What Survives?

If we’re in a bubble — and we likely are — what pops is the speculation, not the technology.

The early contours of a correction are already visible. Venture deal volumes in AI have cooled from their 2023 peak. Investors are pivoting from AGI-style moonshots toward companies that offer tangible enterprise value: infrastructure, tooling, and applications with demonstrable ROI. Many founders are being pushed to generate real revenue instead of living indefinitely on subsidised API calls.

History gives us a likely script. In a correction, thousands of small, fragile startups would disappear. Stronger players would consolidate the best teams and IP; big tech, with its deep pockets and established distribution, would almost certainly come out holding more of the map. Investor attitudes would shift from “grow at all costs” back to fundamentals: revenue, margins, unit economics.

We’d probably also see a right-sizing of ambition. Instead of every startup trying to build a general AI that does everything, the focus would tilt toward smaller, efficient, domain-specific models that do one thing extremely well at a sane cost. The next frontier may be less about brute-force scale and more about clever algorithms, systems engineering, and data quality.

As private froth recedes, the public sector and academia are likely to play a bigger role. Governments already treat AI and compute as strategic infrastructure; a post-bubble world could bring more public funding for foundational research, safety work, and standards, especially in areas the market under-invests in. That shift could push the field toward more responsible innovation rather than pure growth theater.

And there is a crucial social dimension: trust. A few high-profile AI harms or failures could sour public opinion and invite heavy-handed regulation that chills investment. Avoiding that outcome means building systems with humans in the loop in high-stakes domains (healthcare, law, finance, transportation), being transparent about limitations, and embracing sensible rules on data quality, model evaluation, and safety testing. Good regulation can actually stabilise the market by weeding out bad actors and reducing the risk of a reputational crash.

Through all of this, the infrastructure being built now — data centers, GPU clusters, new semiconductor designs, better tooling, improved models — will remain, just as fiber-optic cables and data centers survived the dot-com collapse. The winners after any correction will be the companies that solve real problems, deliver measurable ROI, and run on sustainable margins. The losers will be those sustained purely by the oxygen of venture capital.

Historical Parallels: More Than the Dot-Com Bubble

AI resembles multiple historical cycles:

• Railroads (1840s): enormous overinvestment → consolidation → long-term transformation

• Electrification (1880s–1910s): decades of slow adoption → sudden productivity surge

• Nuclear energy (1950s–1970s): huge early promise → limited by regulation and cost

• Dot-com era (1990s): hype crash → survivors (Amazon, Google) dominate the next 20 years

AI contains elements of all four.

So… Are We in an AI Bubble?

Yes. But it’s a bubble built around something profoundly real.

We’re living through a moment where speculative mania and genuine technological revolution coexist. The risk is not that AI collapses; it’s that the hype blinds us to the hard work required to turn AI into durable, trustworthy, economically viable systems.

The path forward is clear enough: align costs with revenue instead of selling $1 of AI for 10 cents; focus on use cases with measurable payback; invest in efficiency, safety, and reliability; keep humans in the loop where stakes are high; and treat regulation as a tool for stability rather than a threat to be evaded.

Bubbles eventually burst. Transformative technologies don’t.

When the speculative froth finally clears, what remains is the foundation we chose to build. The AI revolution is fragile only if we insist on building it on hype. If we temper exuberance with discipline and design systems people can trust over years, the end of the bubble won’t be the end of AI. It will be the moment the field finally grows up.

References

• Ash Center for Democratic Governance and Innovation. (2024). Watching the Generative AI Hype Bubble Deflate. Harvard Kennedy School. Retrieved from https://ash.harvard.edu/resources/watching-the-generative-ai-hype-bubble-deflate/

• Carvão, P. (2025, October 16). Is AI a Boom or a Bubble? Harvard Business Review. Retrieved from https://hbr.org/2025/10/is-ai-a-boom-or-a-bubble

• Glover, E., & Whitfield, B. (2025, November 24). Is There an AI Bubble? BuiltIn. Retrieved from https://builtin.com/artificial-intelligence/ai-bubble

• Heath, A. (2025, August 15). Sam Altman on the AI Bubble [Interview]. The Verge. Retrieved from https://www.theverge.com/command-line-newsletter/759897/sam-altman-chatgpt-openai-social-media-google-chrome-interview

• Janeway, W. H. (2025, November 7). In Search of the AI Bubble’s Economic Fundamentals. Project Syndicate. Retrieved from https://www.project-syndicate.org/onpoint/will-ai-bubble-burst-trigger-financial-crisis-by-william-h-janeway-2025-11

• Khatib, O. (2025, September). The Fragile Future of AI: Beyond Venture Capital Subsidies. Medium. Retrieved from https://medium.com/@olikhatib/the-fragile-future-of-ai-beyond-venture-capital-subsidies-46abac932c3b

• Martin, J. (2025, October 29). The AI Bubble Is (Sort Of) Real. Does It Really Matter? Exploding Topics. Retrieved from https://explodingtopics.com/blog/ai-bubble

• Nasdaq / Motley Fool. (2025, November 21). Nvidia CEO Jensen Huang Downplays AI Bubble Concerns. Retrieved from https://www.nasdaq.com/articles/nvidia-ceo-jensen-huang-downplays-ai-bubble-concerns-very-different

• Thompson, D. (2025, September). Are We in an AI Bubble? The Atlantic. Retrieved from https://www.theatlantic.com/economy/archive/2025/09/ai-bubble-us-economy/684128/

• Widder, D., & Hicks, M. (2024, November 20). Watching the Generative AI Hype Bubble Deflate. Ash Center Essay, Harvard Kennedy School. Retrieved from https://ash.harvard.edu/resources/watching-the-generative-ai-hype-bubble-deflate/

• Harvard Kennedy School. (2025). Is AI a Boom or a Bubble? Retrieved from https://www.hks.harvard.edu/centers/mrcbg/programs/growthpolicy/ai-boom-or-bubble

• BuiltIn. (2025). Can OpenAI’s Massive Growth Last? Retrieved from https://builtin.com/articles/openai-growing-too-fast

• Exploding Topics. (2025). 44 New Artificial Intelligence Statistics (October 2025). Retrieved from https://explodingtopics.com/blog/ai-statistics

• Medium. (2025). The AI Bubble: How Circular Spending Is Inflating America’s Newest Speculative Frenzy. Retrieved from https://medium.com/@shorealan201/the-ai-bubble-how-circular-spending-is-inflating-americas-newest-speculative-frenzy-edefd45e6646

AI is booming.

In H1 2025, AI startups raised over $83B, nearly 58% of all venture funding (ainvest.com). In Q2 2025, AI alone soaked up ~$50B of $101.5B, while non-AI funding fell to a 7-year low (bestbrokers.com). Those two facts pretty much define the moment: capital has picked a winner, at least for now.

I’ve helped banks, insurers, manufacturers, and media companies ship AI into production. The excitement is earned — some deployments truly change outcomes. But I’m also uneasy: a huge slice of what looks like traction is subsidized by venture money, not sustained by profits. That makes the system fragile when sentiment turns or rates rise.

This piece is half technical, half theoretical, with a dose of lived experience. A few years ago I was “unavoidable.” Today, parts of my work — data prep, baseline modeling, doc writing — can be done by AI. That irony is the point: what happens to an ecosystem (and the people in it) when the VC gravy train slows?

A Venture Capital–Fueled AI Boom

The scale and concentration are historic.

By mid-2025, there were 500+ AI unicorns with an average $5.4B valuation — many young and pre-revenue (elephas.app). FOMO pushed prices beyond normal metrics (elephas.app). PitchBook’s read: nearly half of VC dollars in recent quarters went to AI, up from ~20% only a few years ago (bestbrokers.com). Non-AI startup investment sank to $41.8B in Q2 2025, the lowest in 7+ years (bestbrokers.com).

A handful of mega-checks tell the story. OpenAI raised $40B in early 2025 (SoftBank-led) — a private-tech record (ainvest.com). Meta bought ~49% of Scale AI via a $14.3B stake (bestbrokers.com). In Q2 2025, $40B of the $91B global VC total went into AI — and over one-third of that to Scale AI alone (ainvest.com). This is hyper-concentration.

Speculative heat is obvious. Thinking Machines Lab raised $2B seed at a $12B valuation with no product (elephas.app). Safe Superintelligence raised $2B at $32B with no revenue (elephas.app). Sam Altman called much of this “not rational,” predicting “someone is going to lose a phenomenal amount of money.” (elephas.app).

Markets echo it. Tech is ~34% of the S&P 500, exceeding the 2000 dot-com peak (medium.com). Apollo’s Torsten Slok says today’s leaders look “even more unrealistically” valued than dot-com darlings, with AI multiples above 2000’s peak (elephas.app). The “AI bubble?” question isn’t crazy — just prudent.

Signs of Fragility: Hype Outpacing ROI

Here’s the uncomfortable part. A 2025 MIT analysis found 95% of corporate AI projects delivered no measurable profit (elephas.app). Firms collectively spent ~$40B on pilots that missed goals (pub.towardsai.net). I’ve seen versions of this: brilliant demos that stall at compliance, data quality, or change management.

Economist Paul Kedrosky likened current AI spend to a “massive private sector stimulus,” responsible for ~40% of recent U.S. GDP growth — yet “has yet to actually make anyone any money” in net profits (elephas.app). Translation: capex, headcount, and cloud bills are up; sustainable margins are not. That imbalance can’t run forever.

Structural pressures amplify risk. ~90% AI startup failure rate within five years (ainvest.com). Exit crunch: the IPO window is tight, and many “acquisitions” are acqui-hires (ainvest.com). If funding tightens, a lot of companies have nowhere to go — a familiar 2000-style implosion dynamic (elephas.app).

And we’ve had high-profile flops. Forward (CarePods) burned $650M and shut down in 2024 (elephas.app). Humane AI Pin launched at $699 but saw returns outpace sales, then disappeared by early 2025 (elephas.app). TuSimple slid from $8B to < $300M amid losses and regulatory pain (elephas.app). Money and hype don’t suspend physics, product-market fit, or user patience.

There is real value too. Nvidia revenue up 262% YoY on genuine accelerator demand (elephas.app), and select enterprise workflows report 30–50% improvements (elephas.app). But those gains are narrow and uneven. Costs — talent, data pipelines, evaluation, governance, safety — are stubbornly real.

Technical Trajectories and Economic Realities

Behind every cheap-feeling AI app is an expensive reality. The economics stack looks something like this:

• A user pays $200/year for an AI app.

• That startup pays ~$500 to access foundation model APIs (often OpenAI).

• OpenAI, in turn, pays ~$1,000 to Microsoft for compute.

• Microsoft shells out ~$5,000 to NVIDIA for GPUs.

At each step, venture capital plugs the gap. The $200 from the end user is a fraction of the actual cost, with investors carrying the rest. It creates the illusion that AI is cheap and abundant, when in reality it’s propped up by billions in subsidies.

This subsidy treadmill explains why OpenAI reportedly lost $540M in 2022 (businessinsider.com), why running ChatGPT is estimated to cost ~$700,000 per day just in servers and power (businessinsider.com), and why Alphabet’s John Hennessy admitted that a single Bard query costs 10× more than a Google search (businessinsider.com). It also explains why Sam Altman has floated raising another $100B to pursue AGI, describing OpenAI as “the most capital-intensive startup in Silicon Valley history.” (businessinsider.com).

The technical strategy driving this burn is scale: bigger models, more data, more compute. It’s delivered stunning breakthroughs, but diminishing returns are already setting in (ainowinstitute.org). Each new increment of scale yields smaller gains, while hitting harder constraints — GPU shortages, energy consumption, and cooling costs.

Big Tech, which already spends more on AI infra and R&D than all VC combined (ainowinstitute.org), has responded by lobbying for public subsidies to sustain the buildout. In the U.S., AI is now framed as a national strategic priority, with government grants and state-level incentives flowing into hyperscale data centers (ainowinstitute.org). That risks making AI “too big to fail” — a system where costs are socialized while future profits remain privatized.

This intertwining of technical ambition and financial subsidy makes AI progress fragile. The stack isn’t just about models and GPUs; it’s about who pays at each layer — and what happens when the subsidies run dry.

The Inevitable Correction? Lessons from History

This rhymes with dot-com. Back then, ~4,800 of ~7,000 startups failed by 2003 (elephas.app). The internet survived and matured; Google and Amazon won with real tech and disciplined models (pub.towardsai.net).

Signals today: deal volumes cooling; investors steering toward enterprise value and infrastructure over vague AGI promises (bestbrokers.com; ainvest.com). The IPO drought and exit scarcity push VCs to demand revenue — or pull back (ainvest.com).

What a correction likely brings:

• Shakeout of weak players and pitch-deck theater.

• Consolidation around cash-rich platforms and a few durable independents.

• Investor discipline: revenue, margins, gross-to-net unit economics matter again.

• A shift from “bigger is better” to efficiency, domain-specific models, and systems engineering (ainvest.com).

• A heavier public/academic role in fundamentals and safety.

Optimists are not wrong: users won’t abandon capabilities they’ve internalized — copilots, summarizers, retrieval-augmented search (pub.towardsai.net). The tech persists; the industry structure changes.

Human Capital: Evolving Roles

This wave is changing jobs — mine included.

AutoML eats baseline modeling. LLMs draft SQL, code, tests, docs, and even incident reports. Work that demanded a specialist is now within reach of a power user with the right prompts and guardrails.

In software, ~75% of developers already use AI assistants; the work shifts from writing to reviewing, specifying, and integrating (weforum.org). The WEF points out that data-rich white-collar tasks are more automatable than many manual jobs because the training data is plentiful (weforum.org). Old assumptions flipped.

Macro estimates say ~300M jobs could be exposed to automation globally (goldmansachs.com). That’s task-level exposure, not necessarily net losses — but the reconfiguration is real. We’ll see task shedding, role redesign, and new jobs in evaluation, safety, orchestration, and governance.

Personally, I’m moving up-stack: problem framing, data contracts, evaluation design, risk/ethics, product integration, and change management. The pattern I see: AI won’t replace data scientists, but data scientists using AI will replace those who don’t.

Philosophical Reflections: Value, Dependency, Control

“Inevitability” is a narrative we’ve constructed. AI Now argues that the dominance story is actively made — through hype and national-strategy framing that justifies massive spend and subsidies (ainowinstitute.org). Narratives move budgets. Budgets shape reality.

That creates psychological fragility. If over-promises meet visible failures — harmful bias, safety incidents, or brittle autonomy — public trust can snap. Then comes heavy regulation and stalled adoption, even for valuable use cases.

So moving beyond VC subsidies also means beyond the hype — toward systems people can trust over years: transparent failure modes, evaluation standards, post-deployment monitoring, and clear accountability.

Beyond Venture Capital: A Sustainable Path

If today’s AI stack is fragile because it’s held up by venture capital, then the path forward requires a shift toward sturdier foundations. The question isn’t whether AI has value — it clearly does — but whether the industry can align its cost structures with sustainable revenue. A healthier equilibrium would look very different from the subsidy pyramid we see today.

• Real ROI, not theater. Companies will stop funding “AI because everyone else is” and start demanding measurable returns. The AI projects that will survive are those with clear payback: predictive maintenance that reduces downtime, claims triage that cuts fraud losses, logistics tools that save fuel, radiology support that shortens diagnosis times. These are not flashy moonshots but line items executives can justify on a balance sheet (ainvest.com).

• Backbone businesses take center stage. In gold rushes, the people selling picks and shovels often make more than the miners. In AI, the equivalents are GPUs, cloud platforms, orchestration layers, and data pipelines. These already generate real revenue and will continue to stabilize the ecosystem (ainvest.com). Nvidia’s 262% year-on-year revenue surge is proof that infrastructure is where value crystallizes fastest.

• Monetization that matches cost. Free tiers helped seed user adoption, but inference, safety testing, and compute cannot be given away indefinitely. We should expect AI companies to converge on pricing models that reflect real costs: enterprise subscriptions, licensing, usage-based fees, and premium support. “AI everywhere” for free is a myth; “AI where it delivers ROI at a price” is reality.

• Right-sized models. The arms race toward ever-bigger general-purpose models isn’t sustainable. In many cases, smaller, specialized models paired with retrieval systems and clever prompt engineering can outperform massive LLMs on cost-adjusted utility. This shift from “bigger is better” to “fit-for-purpose” will be critical for long-term viability.

• Human-in-the-loop by design. To build trust and speed adoption, AI must augment humans, not replace them outright. In safety-critical paths — healthcare, finance, law — the human makes the final call. This hybrid model reduces liability, reassures regulators, and ensures adoption by positioning AI as a copilot, not a black-box replacement.

• Regulation that stabilizes. Rules around data provenance, evaluation protocols, incident reporting, and transparency aren’t just bureaucratic red tape; they’re mechanisms to weed out reckless players and build trust. Done right, regulation creates a more level playing field where serious companies can thrive, and consumers know what they’re getting.

• Diversified funding sources. Venture capital doesn’t have to disappear, but it can’t be the only lifeline. Public research grants, open-source ecosystems, cooperative funding models, and even state-level incentives can spread the risk. A more pluralistic funding environment means less monoculture risk — and less chance that the fate of AI is dictated by a handful of VC firms or mega-cap tech companies.

For teams on the ground, this shift means getting serious about evaluation harnesses, guardrails, observability, cost telemetry, and clear go/no-go gates tied to business KPIs. For leaders, it means resisting the temptation of “AI everywhere” and instead backing AI where it pays its own way.

That’s what a sturdier equilibrium looks like: AI that justifies its presence not with hype, but with results.

Conclusion

AI’s present is spectacular — and fragile. Financially (subsidy-driven burn), technically (diminishing returns and physical limits), and socially (trust can whiplash).

Fragile isn’t doomed. After dot-com, the internet matured into the fabric of everything. I expect AI to follow: a correction, a consolidation, then a quieter era where AI is an everyday utility.

The overall repercussion depends on how we steer the next phase. Mismanaged, we get a sharp correction, distrust, and slower innovation. Managed well, we get a resilient general-purpose capability that quietly upgrades healthcare, finance, logistics, public services, and creative work.

My plan — as someone who used to be “unavoidable” — is simple: let AI automate what it can, and lean harder into judgment, context, safety, and integration. If we build for value and trust, and stop leaning on subsidies to paper over unit economics, we’ll earn the future we keep promising.

AI is too important to leave to venture capital alone. It needs engineers, operators, users, regulators — and a steadier narrative than “bigger model, bigger round.” Let’s build the version that lasts.

References

• AI Now Institute. (2025). Artificial power: 2025 landscape report. https://ainowinstitute.org/publications/research/ai-now-2025-landscape-report
  (PDF: https://ainowinstitute.org/wp-content/uploads/2025/06/FINAL-20250602_AINowLandscapeReport_Full.pdf)

• AI Now Institute. (2025, June 3). 1.2: “Too Big to Fail”: Infrastructure and capital push. https://ainowinstitute.org/publications/research/1-2-too-big-to-fail-infrastructure-and-capital-push

• Chowdhury, H. (2023, May 5). ChatGPT cost a fortune to make with OpenAI’s losses growing to $540 million last year, report says. Business Insider. https://www.businessinsider.com/openai-2022-losses-hit-540-million-as-chatgpt-costs-soared-2023-5

• Goldman Sachs Research. (2023, April 5). Generative AI could raise global GDP by 7%. https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent

• Goldman Sachs Research. (2024, May 13). AI is showing “very positive” signs of eventually boosting GDP and productivity. https://www.goldmansachs.com/insights/articles/AI-is-showing-very-positive-signs-of-boosting-gdp

• Hoffman, P. (2025, June 10). The state of AI venture capital in 2025: AI boom slows with fewer startups but bigger bets. BestBrokers. https://www.bestbrokers.com/forex-brokers/the-state-of-ai-venture-capital-in-2025-ai-boom-slows-with-fewer-startups-but-bigger-bets/

• Kumar, A. (2025, August 12). Why AI is replacing some jobs faster than others. World Economic Forum. https://www.weforum.org/stories/2025/08/ai-jobs-replacement-data-careers/

• Morgan, C. (2025, July 22). The AI funding boom vs. the exit crunch: Why investors should rethink AI startup valuations. AInvest. https://www.ainvest.com/news/ai-funding-boom-exit-crunch-investors-rethink-ai-startup-valuations-2507/

• Reuters. (2025, June 10). Meta to pay nearly $15 billion for Scale AI stake, The Information reports. https://www.reuters.com/business/meta-pay-nearly-15-billion-scale-ai-stake-information-reports-2025-06-10/

• Robison, K. (2024, August 7). Humane’s daily returns are outpacing sales. The Verge. https://www.theverge.com/2024/8/7/24211339/humane-ai-pin-more-daily-returns-than-sales

• Vincent, J. (2025, February 18). Humane’s AI Pin is dead, as HP buys startup’s assets for $116M. TechCrunch. https://techcrunch.com/2025/02/18/humanes-ai-pin-is-dead-as-hp-buys-startups-assets-for-116m/

• Visual Capitalist. (2025, July 14). Ranked: The biggest AI funding rounds of 2025 so far. https://www.visualcapitalist.com/ranked-the-biggest-ai-funding-rounds-of-2025-so-far/

• World Economic Forum. (n.d.). Why AI is replacing some jobs faster than others (resource page variant). https://initiatives.weforum.org/ai4ai/resources_v2/publications/why-ai-is-replacing-some-jobs-faster-than-others/f64d599366b986cd69ffa130268dfeeef12a8a06

• Brena, R. (2025, September). What comes after the AI bubble explodes? Towards AI (Medium). https://pub.towardsai.net/what-comes-after-the-ai-bubble-explodes-52fa1405681f

• Fortune Editors. (2025, August 18). MIT report: 95% of generative AI pilots at companies are failing. Fortune. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/

Artificial intelligence is reshaping the job landscape at a breakneck pace. Not long ago, being a data scientist felt like a guaranteed ticket to career success — the “sexiest job of the 21st century,” as some called it. I experienced this first-hand: just a few years back, my expertise in building machine learning models made me indispensable. Today, however, the ground is shifting beneath our feet. AI tools can code, analyze data, and even automate tasks that once required specialized human skills. Employers are both excited and anxious — the World Economic Forum forecasts AI will spur a 40% growth in demand for AI and machine learning specialists in coming years, yet they also warn that 39% of core skills could be obsolete by 2030. In such a dynamic environment, the question looms large: How can a job candidate truly stand out?

The answer lies in a combination of technical excellence, adaptability, and quintessentially human qualities. This article explores how one can distinguish themselves in the AI-driven world, blending insights from industry trends, statistics, and my own perspective as a data scientist who has helped companies implement AI across industries. It’s both a technical roadmap and a philosophical reflection on what it means to be uniquely valuable in the age of intelligent machines.

The Changing AI Job Landscape

The AI revolution is a paradoxical force in the job market — simultaneously a creator of new opportunities and a disruptor of old roles. On one hand, demand for AI talent is booming. AI-related jobs and industries are growing at roughly 40% year-over-year, with over a million new jobs expected by 2027. In fact, global demand for AI and machine learning specialists is projected to keep climbing rapidly in the next few years. Companies across sectors — from finance to healthcare — are racing to hire professionals who can develop AI solutions or integrate AI into business processes.

On the other hand, AI is also transforming the nature of work itself. Advanced AI systems are increasingly capable of tasks that used to set candidates apart. For example, generative AI like OpenAI’s GPT can now write code, generate analyses, and draft reports at a level that would have been unthinkable a few years ago. Routine data crunching and even parts of software development are being automated or accelerated by tools such as GitHub Copilot, Cursor, and Claude Sonnet. Indeed, it’s becoming common for less-specialized employees to perform tasks “previously reserved for experts,” as generative AI allows “less specialised workers to undertake tasks previously reserved for experts, improving productivity in roles such as accounting clerks, nurses and teaching assistants.” The playing field is leveling in terms of who can leverage AI, meaning that having baseline technical skills is no longer the differentiator it once was.

A striking indicator of this shift is seen in hiring trends for traditionally technical roles. In data science, for instance, employers now expect candidates to have AI familiarity by default. Nearly a quarter of data scientist job postings explicitly mention “artificial intelligence,” and the demand for natural language processing (NLP) expertise in these roles jumped from 5% in 2023 to 19% in 2024. Machine learning skills are mentioned in about 69% of postings. Conversely, some fundamental programming skills are mentioned slightly less often than before — in part because they are so common, and possibly because new AI-driven tools handle some of the heavy lifting. For example, Python was a requirement in 78% of data science job ads in 2023, but that dropped to 57% in 2024. Analysts speculate this “could be the impact of AI on the data science job market” — a sign that coding ability alone isn’t the rarity it once was, and candidates must bring more to the table.

Even high-skilled professionals feel the tremors of potential obsolescence. The World Economic Forum pointedly listed data science among the jobs most at risk of automation by AI. Not long ago, data scientists like me were in short supply and high demand; today, we wonder which parts of our job will be taken over by smarter algorithms or automated pipelines. Yet, it’s not all doom and gloom. History shows technology tends to transform jobs more often than it eradicates them. In fact, the WEF’s Future of Jobs Report 2025 predicts AI and related technologies will create 170 million new jobs by 2030 (while possibly displacing 92 million traditional roles). The net effect can be positive — provided workers adapt. To thrive in this environment, job candidates must evolve into something more than just code jockeys or model builders. They need to become strategic, creative, and multidisciplinary problem-solvers who use AI as an amplifier for their own talents. In the sections that follow, we will delve into how exactly one can cultivate this kind of profile.

Lifelong Learning and Adaptability

In a world where nearly 40% of employers plan to reduce workforce in areas where AI can automate tasks, the capacity to learn and adapt is arguably a candidate’s greatest asset. The half-life of technical skills is shorter than ever — by 2030, about one-third of the skill sets considered important in today’s jobs will have changed. As Klaus Schwab of the WEF observes, “if a five-year degree were designed for today’s skills, by the time it is completed, two years’ worth of those skills would already be outdated.” This reality means that to stand out, you must demonstrate intellectual agility: the willingness and ability to continually upskill yourself as technology evolves.

From a practical standpoint, adaptability starts with embracing new tools and techniques as they emerge. Rather than fearing AI tools, top candidates use them to augment their own productivity. As Sadie St. Lawrence — founder of Women in Data — put it at a recent global AI conference, “Don’t worry so much about necessarily the right skills — make sure you’re integrating the use of AI in your daily workflow.” Being an early adopter of helpful AI software (whether it’s an automated data visualization tool or a coding assistant) signals to employers that you’ll make the team more efficient, not be rendered redundant. In my experience consulting across industries, those who quickly learn to incorporate AI into their domain — becoming what one might call “AI operators” — are highly valued. They can achieve more with less and help others do the same.

Adaptability also means having a T-shaped skill set: broad knowledge across disciplines combined with deep expertise in at least one area. An AI specialist with a diverse background can bridge gaps between teams and innovate in ways a one-dimensional expert cannot. NVIDIA’s experts echo this, noting that a “diverse educational and professional background can make people more adaptable and versatile — a valuable asset, especially in the rapidly evolving field of AI.” If your background is non-traditional, don’t see it as a weakness — it might be your secret strength. People who can connect AI with fields like art, psychology, or chemistry, for example, often bring fresh perspectives that pure computer scientists may miss. In short, commit to lifelong learning. Show on your resume and in interviews that you’ve continually picked up new programming languages, mastered the latest machine learning frameworks, or even taken courses in areas outside your primary expertise. This signals that you won’t be left behind as the field advances. Given that 85% of employers plan to prioritize upskilling their workforce in coming years, showing you can upskill yourself makes you an ideal hire in the AI era.

Carve Out Your Niche and Purpose

While a broad base of knowledge is important, so is specialization — particularly if it aligns with your passion. In a world teeming with AI generalists, a candidate can distinguish themselves by having a clear niche or domain where they excel. As one NVIDIA university recruiting manager observed, there isn’t a monolithic “AI candidate” archetype so much as a need for people with AI skills in every domain: “there are sub-areas within AI, different domains and industries that are going to be affected by AI, which you’ll want to gain experience in, because that’s where companies are looking to implement AI.” In other words, finding your domain — whether it be healthcare, finance, robotics, creative arts, or any other field — and becoming conversant in how AI applies to it can make you uniquely valuable. An applicant who understands both AI technology and the nuances of, say, clinical medicine or supply chain logistics, will stand out from those who only know the algorithms.

Just as important is articulating your “why” — the driving passion behind your career. Companies respond to candidates who have a sense of mission, not just a laundry list of technical skills. Being able to point to a problem you genuinely care about and have worked on can set you apart from the crowd of applicants with similar resumes. In fact, one piece of advice I give mentees is to “pick a real-world problem that seems impossible to solve but that you’re passionate about,” and use that as a guiding star for your projects and learning. As Saudia Jones, a generative AI analyst at NVIDIA, said of her own career path: “Connect with ‘why’ you’re driven by this, because that’s going to be the thing that leads you to overcome the challenges and innovate.”

Concrete example: if you are passionate about climate change, dive into AI applications in sustainability — perhaps you’ve done a project using machine learning to optimize renewable energy grids or predict deforestation. That combination of expertise + enthusiasm shows employers that you’re intrinsically motivated and likely to go above and beyond. It also helps you craft a personal narrative or portfolio that is memorable. Many hiring managers have shared with me that they recall candidates who told a compelling story about their interests far more than those who rattled off technical achievements devoid of context. In a global talent pool, your personal vision and niche focus can be a key differentiator.

Technical Excellence and a Proactive Mindset

Of course, passion and adaptability must be underpinned by solid technical skills — there’s no getting around that in AI. To stand out, a candidate should demonstrate mastery of the core tools of the trade as well as the latest advancements. This is where the “half technical” aspect comes in. It’s expected that you know programming languages (Python, SQL, etc.) and machine learning frameworks, but what can really distinguish you is showing that you keep pace with cutting-edge developments. In 2024 and beyond, that includes familiarity with large language models (LLMs), generative AI techniques, and advanced frameworks for deep learning and data engineering. For instance, many job postings now look for experience with transformers, knowledge of prompt engineering, or deploying models at scale. Highlighting any experience you have with these — even if through personal projects — can give you a leg up. Notably, the demand for NLP skills quadrupled within a year, reflecting the explosion of interest in LLMs; riding this wave can put you in a stronger position.

However, technical excellence is not just about what you know, but what you’ve done with it. Employers love to see a portfolio of projects that show initiative and creativity. In fact, recruiters at NVIDIA say the top things they look for are “1) a demonstration of the technical skills required for the role, 2) subject-matter expertise, and 3) projects that exhibit thinking outside of the box.” The message is clear: to distinguish yourself, don’t just list skills — demonstrate them through unique projects. Maybe you’ve built a mobile app that uses computer vision to help visually impaired users (showing both technical skill and social impact), or you contributed to an open-source AI library (showing initiative and community involvement). Perhaps you entered an AI competition on Kaggle and achieved a high rank, or you collaborated with a local business to implement an AI solution to a real problem. These concrete achievements speak louder than any credential. They prove that you can apply knowledge to solve problems, which is ultimately what employers care about most.

Being proactive often means going beyond assigned coursework or job duties. It can be as simple as identifying a need and addressing it without being asked. For example, during one of my consulting engagements, a junior analyst taught himself how to use an AI visualization tool to automate weekly reports — before management even knew such automation was possible. He not only saved the team dozens of hours but also made himself stand out as a self-starter. As one hiring manager put it, “identify a problem you can solve and take proactive steps to address it, even before being hired.” In interviews, be ready to discuss times when you took initiative in the realm of AI. Did you notice your previous company’s marketing team was manually sorting customer feedback, and then pilot an NLP model to categorize sentiments? Stories like that showcase the mindset of an innovator and problem-solver — exactly the type of person who thrives alongside AI rather than being replaced by it.

It’s worth noting that standing out in technical terms doesn’t mean you must know everything under the sun. It’s about the quality of your skills and ideas, not just quantity. You might specialize deeply in one subfield (say, reinforcement learning or AI security) and be among the few with that expertise. Or you might combine two skills uniquely (for example, data visualization + storytelling) to become an unusually effective communicator of insights. The key is to show both breadth (so you can adapt) and depth (so you can deliver value quickly). And above all, show that you can keep learning. In an industry where a breakthrough paper can render techniques obsolete overnight, your attitude towards continuous learning is a technical skill in itself.

The Human Edge: Creativity, Ethics, and Empathy

In a job market increasingly populated by AI algorithms, the most important question becomes: What can a human do that AI can’t? This is where the more theoretical and philosophical side of our discussion comes in. The best candidates distinguish themselves not just with what they know, but with how they think and behave. Traits like creativity, ethical judgment, empathy, and leadership — the “human touch” — are emerging as make-or-break qualities.

One critical area is creativity and innovative thinking. AI, for all its prowess, is still largely bound by the data it’s trained on and the directives given by humans. It excels at pattern recognition and optimization but lacks true imagination or the ability to define its own goals. As a candidate, you can highlight instances where you thought outside the box — perhaps designing an unconventional solution to a problem or combining techniques in a novel way. Recruiters love to hear about moments when you went beyond the obvious. This shows you won’t just use AI tools, but also conceive new ways to apply them. Remember, while thousands may use a given machine learning library, the idea for an impactful application is born from a human insight. Show that you can provide that insight.

Another differentiator is ethical and responsible AI use. Companies are increasingly conscious of AI’s societal impacts — issues like bias, transparency, and privacy. A candidate who is knowledgeable about AI ethics and committed to using AI responsibly will stand out as someone who can safeguard a company’s reputation and values. In practice, this could mean mentioning any experience you have with fairness in AI or simply communicating a thoughtful stance on how you approach AI in projects. As NVIDIA’s recruiting manager Eric Vargas advises, it’s crucial to use AI tools responsibly and maintain a human touch, even in something as simple as writing your cover letter. Organizations seek people who won’t just build powerful AI, but will ensure that AI is trustworthy and aligned with human needs. In fact, “Trustworthy AI” — focusing on safety and transparency — is now a key principle in the industry. Demonstrating awareness of this (for example, mentioning that you kept interpretability in mind when developing a model, or that you are mindful of data privacy) can set you apart from candidates who are narrowly focused on raw performance metrics.

Empathy and communication skills also fall under the human edge. AI might analyze sentiments, but it can’t truly feel or communicate with emotional intelligence. If you have strong interpersonal skills — say you led a team project, mentored others, or excel at translating technical jargon for non-technical audiences — make sure to highlight that. In my consulting work, I’ve seen technically brilliant individuals stagnate because they lacked communication skills, while moderately skilled folks who could act as a bridge between the AI team and management rose quickly. In job interviews, your ability to connect, listen, and convey ideas effectively often leaves a lasting impression. It signals that you can collaborate in multidisciplinary teams and drive adoption of AI solutions (a critical aspect, since implementing AI is as much a people challenge as a technical one).

Finally, there is the aspect of using AI for good — aligning your work with positive social or business outcomes. When an employer sees that you care about impact, you become more than just another hire; you become someone who might elevate the team’s purpose. For instance, maybe you volunteered your AI skills on a humanitarian project (like improving accessibility, as in the case of an NVIDIA intern who worked on sign language AI to help the deaf). Even at a philosophical level, expressing a genuine intent to harness AI in ways that augment human well-being (rather than just chasing profit or efficiency) can make you memorable. As noted in an industry discussion, the combination of “showcasing one’s passion and using AI for good is key to standing out in the field.” In summary, your integrity, vision, and humanity are competitive advantages. They’re hard to quantify on a CV, but they shine through in your stories, recommendations, and demeanor — and they are exactly what AI itself lacks.

Networking, Collaboration, and Thought Leadership

No discussion of standing out would be complete without addressing networking and collaboration. In the AI world, what you know is amplified by who you know. Building a professional network can open doors to opportunities that aren’t advertised and give you insights that keep you ahead of the curve. Engaging with the AI community — whether by attending conferences (even virtual ones), participating in online forums, or contributing to open-source projects — helps you develop a reputation and learn from others. Sadie St. Lawrence, a leader in the data science community, emphasizes that “building a network and seeking mentorship in the AI space has been crucial for my career growth”. Connecting with experienced professionals can accelerate your learning and expose you to the unwritten rules of the industry.

One practical tip is to use platforms like LinkedIn strategically. Follow thought leaders in AI, share your own insights (even if brief posts about things you’ve learned), and don’t hesitate to reach out politely to people whose work you admire. You might be surprised how often they’re willing to give advice or even recommend you when they hear of a job opening. Moreover, recruiters themselves use AI-driven tools on platforms like LinkedIn to find talent — by incorporating relevant keywords (skills, certifications, projects) in your profile, you increase your visibility. Many modern applicant tracking systems and search algorithms will surface candidates who not only have the right skills but also demonstrate passion and engagement in the field. So think of your public presence as part of your personal brand: it should reflect the unique combination of skills and interests that you bring to the table.

Collaboration skills are equally important. AI projects often require cross-functional teamwork — data scientists, software engineers, domain experts, and stakeholders all need to work in concert. A candidate who can show they’ve thrived in team settings (for example, mentioning agile team experience, or collaborations between technical and business units) will be seen as an asset. In interviews, rather than only touting individual achievements, highlight instances of successful teamwork and how you influenced outcomes. This reassures employers that you can operate in the real-world environment where AI solutions are built by groups, not lone geniuses. Given that many AI initiatives fail not due to technology but due to poor integration and change management, having team players who can break silos and drive change is invaluable.

Finally, consider developing some level of thought leadership as your career progresses. This doesn’t mean you need to be a famous blogger or researcher, but even on a small scale, you can distinguish yourself by contributing knowledge to the community. It could be writing a Medium article about an experiment you did, giving a talk at a local meetup, or helping answer questions on a forum like Stack Overflow or Reddit’s r/MachineLearning. When an employer sees that you care about the field enough to share knowledge, it signals genuine enthusiasm and expertise. In my case, writing about my AI project experiences not only helped others but also solidified my own understanding and credibility — and yes, hiring managers mentioned they were impressed by that during interviews. In a global AI talent pool, being an active contributor (not just a consumer) of knowledge is a clear marker of someone who is distinguished and driven.

Overall Repercussions and Future Outlook

As we look at the big picture, it’s clear that the rise of AI is bringing profound shifts to the workforce. The repercussions of these shifts depend largely on how individuals and organizations respond. For job candidates, the stakes are high: those who fail to adapt risk being left behind in a world where machines can do an ever-growing portion of routine work. But for those who embrace the strategies discussed above — continuous learning, carving out a niche, combining technical skill with human creativity and ethics, and building strong networks — the future can be incredibly promising. These candidates won’t be replaced by AI; instead, they’ll be the ones operating, guiding, and improving AI systems. They’ll fill the new roles that are emerging (the fact that we now see titles like “Prompt Engineer” offering six-figure salaries to humanities graduates attests to how entirely new specialties are forming). They’ll also be the people who ensure that technology development remains aligned with human values.

For companies and the broader economy, a workforce that adapts means AI is more likely to augment human potential rather than diminish it. According to the WEF, 86% of organizations expect to see significant transformation from AI by 2030. If the workforce is ready — skilled and resilient — this transformation could lead to unprecedented productivity and the solving of problems once thought intractable. Imagine AI-assisted doctors focusing more on patient care, or AI-augmented engineers designing sustainable cities faster. However, if workers aren’t prepared, companies could face talent shortages and a growing skills gap (indeed, 63% of employers already cite skills gaps as a primary barrier to adopting new technologies). The onus is on both individuals and organizations to invest in upskilling and reskilling. Encouragingly, many are doing so: 85% of employers plan to prioritize upskilling their people in the coming years, and as individuals, we should seize those opportunities.

Philosophically, the rise of AI forces us to reconsider what it means to have a fulfilling career. If AI takes over more tasks, humans may find themselves focusing on higher-level and more interpersonal aspects of work. The repercussion could be a workforce that is, counterintuitively, more human — where traits like empathy, flexibility, and curiosity define our jobs more than number-crunching does. But reaching that point requires proactive effort from today’s job candidates to mold themselves into the kinds of professionals who can thrive in partnership with AI.

In conclusion, distinguishing oneself in the world of AI comes down to this: be the person who can do what AI cannot, and excel at leveraging what AI can. This means pairing your human creativity and purpose with cutting-edge technical know-how. It means being as agile and intelligent in your learning as the algorithms are in their optimization. It means breaking out of silos, continuously re-inventing your skill set, and never losing sight of the real-world impact of your work. The job market is being flooded with AI and with people claiming AI skills — but your unique mix of skills, experiences, values, and vision is something only you have. Cultivate it, showcase it, and you will not only remain relevant; you will be the kind of candidate who drives the AI revolution forward. In a world of smart machines, dare to be the wise, adaptable, and insightful human. That is your ultimate competitive edge.

References and Sources

365 Data Science. (2024, January 12). Data science job market in 2024: Trends and predictions. Retrieved from https://365datascience.com/trending/data-science-job-market-in-2024/

Business Insider. (2023, March 31). AI prompt engineers are making up to $335,000 a year. Here’s what they do and how to become one. Retrieved from https://www.businessinsider.com/ai-prompt-engineer-jobs-salary-career-chatgpt-2023-3

Interview Query. (2024, January 5). Data science job report 2024. Retrieved from https://interviewquery.com/blog-data-science-job-report-2024

NVIDIA. (2023, October 3). Top 5 tips for standing out in the AI job market. Retrieved from https://blogs.nvidia.com/blog/2023/10/03/top-5-tips-for-standing-out-in-the-ai-job-market/

Sustainability Magazine. (2024, April 23). WEF: AI to drive growth, productivity, and skills shift by 2030. Retrieved from https://sustainabilitymag.com/articles/wef-ai-to-drive-growth-productivity-and-skills-shift-by-2030

World Economic Forum. (2025). Future of jobs report 2025. Retrieved from https://www.weforum.org/reports/future-of-jobs-report-2025

Zhang, Y. (2023, June 21). Are data scientists at risk of being replaced by AI? Towards Data Science. Retrieved from https://towardsdatascience.com/are-data-scientists-at-risk-of-being-replaced-by-ai-4c8849f22e0f

For most of my career implementing AI inside big, complicated organizations, the Big Four — Deloitte, PwC, EY, and KPMG — have felt like the granite pillars of professional services. In the UK, they audited 99% of FTSE 100 companies as of 2011. That dominance isn’t a fluke; it’s the compound interest of scale, brand, and a pyramid model that rewards lots of smart people billing lots of hours.

The shift now isn’t cyclical. It’s structural. The real question isn’t whether AI will touch the Big Four — it already has — but whether it collapses the old operating model or forces a reinvention strong enough to preserve their leadership.

Over the past three years I’ve watched prototypes escape the lab. CFOs who once asked me for “an AI strategy slide” now want to know which agents can pass internal audit. Boards aren’t debating whether to use AI, but how to govern it and how to price services that are part software, part human judgment. If you want a rule of thumb for impact, it’s this: where the work is structured and data-heavy — audit, tax compliance, due diligence, risk and forensics — AI lands first. Strategy and change management follow, but later.

Their model is world-class — and under pressure

The “Big Four” emerged from mergers (and Arthur Andersen’s 2002 collapse), consolidating what used to be the “Big Eight”. The engine has been a classic pyramid: many juniors at the base, a tight group of partners at the top, and billable hours as the fuel. It’s worked for decades.

AI challenges the core economics. When machines do a big share of the tasks people used to do, hours shrink, teams shrink, and clients ask harder questions about fees. Even if demand holds, the mix shifts: fewer time-based engagements, more platform + outcomes.

I often suggest a simple test to leadership teams: “If your top 10 offerings had to be delivered with 50% fewer human hours in 18 months, what would you change first — pricing, process, or product?” That conversation surfaces the operational debt hidden inside the current model.

What AI does to audit and advisory: automate the core

Audit, forensics, due diligence, compliance — these are structured, data-heavy domains. That’s exactly where AI shines. KPMG already uses AI to scan millions of accounting entries and flag anomalies for human review. In due diligence, models can read/summarize thousands of contracts and financials in minutes, surfacing risks humans might miss.

Insiders are blunt about the direction of travel. Alan Paton, ex-PwC, predicts “most structured, data-heavy tasks in audit, tax, and strategic advisory” will be automated within 3–5 years, potentially eliminating about 50% of roles in those areas — and says there are AI solutions capable of 90% of the audit process already. If you’re a CFO, you’ll ask why you’re paying for weeks of human testing when a system can do it instantaneously.

The firms are not asleep at the wheel. Deloitte launched Zora AI — agentic tooling (with Nvidia tech) that automates invoice processing, trend analysis, and more. EY’s AI now assists 80,000 tax pros, handling 3 million+ compliance cases and tens of millions of routine processes each year. These tools free thousands of hours — great for productivity, but it pressures the billing model built on human hours.

Where this lands practically: audits become exception-driven. Humans shift from sampling to investigating AI-flagged outliers; advisory work shifts from “build the analysis” to “stress-test scenarios and steer change.” The winners will formalize human-in-the-loop controls — clear thresholds for escalation, evidence trails for regulators, and documented model limitations.

From pyramid to diamond: the talent model flips

The pyramid relied on large graduate intakes doing repetitive work to learn the ropes. AI does much of that work now. Firms are already tightening the base: graduate openings fell 44% year-over-year in 2024 across the Big Four; KPMG cut some cohorts by nearly 30%.

As Ian Pay (ICAEW) put it, firms are talking about a “diamond model” — a thinner base, a wider middle of technical and managerial experts — because AI can’t yet make all the judgment calls. Expect demand to shift from rote task-doers to AI-fluent mid-career pros who can interpret outputs and manage edge cases.

Offshoring? Also pressured. If AI can do the work anywhere, labor arbitrage loses bite. Paton’s view: any model “based on the number of people you have” is “really vulnerable” now. We’ve seen offices shrink in high-cost markets and grow elsewhere (e.g., Deloitte NL down 5%, Malaysia up 9% in a year). AI could both accelerate that shift and reduce the total need for offshored roles.

If I were designing a 2026 intake, I’d weight hiring to:

1. Controls Engineers

These are the people who make AI safe and auditable. They stand up a live model inventory, wire in accuracy/bias/drift tests, ensure runs are reproducible, and document everything so a regulator (or partner) can follow the trail. In the first 90 days, I expect a working registry, basic telemetry, and signed model cards for the top use cases. I hire for baseline thinking, data lineage instincts, and the ability to turn a policy line into an automated check. I track simple numbers: models registered and tested, time to approval, incident rate, audit findings closed.

2. Domain Translators

Translators turn a CFO’s question into a solvable AI problem — and back again into plain-English value. They define metrics a board cares about (coverage, speed, errors, savings), write sensible data contracts, and narrate results clients can act on. Quarter one should produce two board-safe playbooks (e.g., GL anomaly review, vendor DD) with clear definitions of success. I look for structured assumptions, a quick value tree on a whiteboard, and crisp writing. Success shows up as faster cycle times, real adoption, delivered outcomes, and better NPS.

3. Exception Handlers

When the system flags something, these people investigate, decide, and explain — calmly. They triage alerts, chase root causes, and choose escalate/resolve/suppress with clear reasoning and a paper trail. Early on, I want a usable exception taxonomy, runbooks anyone can follow, and SLAs baked into workpapers and client comms. I hire for curiosity under pressure and empathy with facts. I track mean time to resolution, false positives, reoccurrence, and — most important — trust.

Team Shape

I’d staff small pods: one controls engineer, one translator, two exception handlers, with an ML engineer on tap. Year one, bias toward controls to build the governance spine (roughly 40/35/25 across controls/translators/exception). As playbooks stabilize and we shift from “make it safe” to “make it adopted,” move toward 30/40/30.

Onboarding

Make the runway short and real: a two-week bootcamp on tools and policy, sixty days shadowing a flagship use case, and a ninety-day certification tied to promotion. Rotate translators through client demos, exception handlers through red-team drills, and controls engineers through audit dry-runs. The goal is simple: safe by design, valuable by default, ready for scrutiny.

When “insight” gets commoditized

Clients paid a premium because the firms’ knowledge and analysis were scarce. AI changes that equation. LLMs synthesize industry data, regs, and best practices in seconds. Smaller firms and in-house teams can now punch above their weight with AI.

As Alibek Dostiyarov observes, AI lets boutiques proliferate — one client handled 10–12 inquiries simultaneously where before they could handle only a few. Mid-market West Monroe says win rates are up and pipelines are at records, aided by AI; they’re also attracting talent from the Big Four. That’s a real threat: brain drain plus nimble, tech-forward rivals.

Clients’ expectations are shifting to outcomes over hours. One senior advisor in India notes engagements once priced at ₹20–25 lakh are now done for about half, with margins and timelines under pressure. “Outcome-based advisory will be the norm… soon.”

For Big Four partners, that means sharpening three edges:

1. Differentiated IP (not just decks — tools clients reuse),

2. Decision rights (clarity on when AI speaks vs. when humans decide),

3. Implementation muscle (change management, data plumbing, and risk sign-off).

Disrupt yourself — or be disrupted

To their credit, the Big Four are investing billions in AI. KPMG announced $2B over five years, targeting $12B in added revenue. Each firm has built proprietary platforms: Deloitte’s AI-driven audit analytics; PwC’s GL.ai with H2O.ai; EY’s Helix; KPMG’s Ignite. The pitch: human expertise + AI scale.

But scale cuts both ways. As Hywel Ball (ex-EY UK chair) notes, the bigger you are, the slower change can be; even small process tweaks can take months or years. Meanwhile, smaller firms can roll out a tool in weeks. That adoption gap matters.

The commercial model also has to change. If hours shrink, pricing must pivot to subscriptions or outcomes. EY leaders have hinted at a SaaS-style future for services. The wider industry is already moving: McKinsey reports ~25% of projects are outcomes-priced, uses an internal GPT (Lilli) in 70% of consultants’ daily work, cut 30% of drudge time, eliminated 5,000+ support jobs, and gets 40% of revenue from tech-enabled services. Different firm, same lesson: bite the bullet, restructure, and productize what you can.

What I advise leadership to stand up in the next 12 months:

• An AI “trust kit” for the whole firm
  One place to register every model, simple checks for accuracy/bias/drift, prompt logs, and clear “a human signs off here” moments.
  Month 1: draft the rules and pilot on two busy use cases. Month 2: turn on the registry and basic telemetry. Month 3: do a mock audit so we know it holds up.

• A shelf of reusable AI building blocks
  The stuff we do over and over — extract, reconcile, flag anomalies, summarize — packaged as clean, versioned APIs with owners and SLAs.
  Month 1: pick the top six blocks. Month 2: ship v1 with docs. Month 3: wire them into three flagship services and stop reinventing the wheel.

• A pricing council that tests outcomes and subscriptions
  Define what “value” looks like per service (coverage, speed, error rate, savings) and run real pilots.
  Month 1: choose five offers to test. Month 2: A/B the pricing. Month 3: publish what worked and expand. Start with vendor due diligence or GL anomaly reviews — they’re measurable and fast.

• A skills ladder that actually maps to the work
  Name the roles we need now — Exception Handler, Domain Translator, Controls/ML Engineer — and tie micro-certs to promotion.
  Month 1: publish the ladder and baseline skills. Month 2: certify 15% of staff. Month 3: make certification table stakes for leading engagements.

The global wrinkle: regulation and geopolitics

The Big Four operate everywhere, and rules aren’t uniform. The EU’s AI Act (expected to bite in 2025) will ratchet up requirements on data, transparency, and risk. The firms see an opening: become the auditors of AI itself — checking effectiveness, bias, and compliance — much like they did with ESG. PwC UK, for example, is gearing up AI assurance that tests chatbot accuracy and examines algorithms for unfair bias. Deloitte calls such assurance critical to adoption. A new revenue line beckons.

But regulation cuts both ways. The EU may slow adoption; the US (for now) is lighter-touch, which could accelerate disruption. China is a separate puzzle: heavy domestic AI investment plus tighter scrutiny of foreign auditors (reports of MOF scrutiny and talk of reducing reliance on the Big Four) complicate the landscape.

Regulators also want proof that AI improves audit quality. The UK’s FRC found wide use of AI tools but inconsistent monitoring of their impact — few clear KPIs to show quality gains. Expect demands for transparency on when and how AI is used in audits.

Operationally, global firms should design for regional variance: data residency, explainability requirements, and local model registries. One-size-fits-all deployment will create either compliance gaps or needless friction.

So — collapse or reinvention?

A literal collapse is unlikely. These firms are too capitalized, too embedded. What’s far more likely is the collapse of the old operating model — and the rise of a new one. By the early 2030s, we may see a “diamond” org shape, fewer total headcount, many more AI engineers, data people, and hybrid operators, and revenue mixes that look more like platform + outcomes than pure hours.

Could we end up with a Big Five or Six (as tech-enabled challengers scale) or a Big Three (if one stumbles)? Yes — plausible scenarios now. As one industry voice put it, firms must change “talent, mix of people, systems, operating model to make it through the valley of death.” I agree.

If they get it right: better audits (full-population testing, earlier anomaly detection), faster advisory (weeks of analytics in days), and clearer value (priced for outcomes). If they don’t: a slow bleed of clients, talent, and relevance to nimbler competitors.

My view after a decade in the trenches: the winners will practice productive paranoia — move fast, measure impact, price on value, and treat AI as a teammate. The old Big Four may indeed be fading. But a new Big Four — leaner, tech-enabled, and wiser — can absolutely take their place. The clock’s ticking.

References

• Accountancy Age. (2025, June 27). FRC questions rising use of AI by auditors. ICAEW

• Accountancy Age. (2025, August 6). Big Four lag behind smaller firms in AI adoption, says ex-EY chair. Accountancy Age

• Axios. (2023, July 11). Microsoft strikes $2 billion AI partnership with KPMG. Axios

• Business Insider. (2025, April 14). Uber cofounder says AI means some consultants are in “big trouble”. Business Insider

• Business Insider. (2025, May 20). AI is coming for the Big Four too. Business Insider

• Business Insider. (2025, April). Inside the AI boom that’s transforming how consultants work at McKinsey, BCG, and Deloitte. Business Insider

• CIO.com. (2025, March 18). Deloitte unveils agentic AI platform, Zora AI. CIO

• Competition & Markets Authority. (2019, April 18). Statutory audit services market study: Final report. GOV.UK

• Deloitte. (2025, March 18). Deloitte unveils Zora AI: Agentic AI for tomorrow’s workforce [Press release]. Deloitte United Kingdom

• Deloitte. (2025, July 15). Deloitte expands AI capabilities in Omnia global audit platform [Press release]. Deloitte United Kingdom

• Deloitte. (n.d.). Digital audit technology — Deloitte Omnia. Deloitte United Kingdom

• EY. (2025, March 18). EY launching EY.ai Agentic Platform, created with NVIDIA AI, to drive multi-sector transformation — starting with tax, risk and finance domains [Press release]. EY

• EY. (2025, April 9). EY announces large-scale integration of leading-edge AI technology into global assurance technology platform [Press release]. EY

• EY. (n.d.). EY Helix — Audit technology. EY

• Financial Times. (2025, July). Big accounting firms fail to track AI impact on audit quality, says regulator. Financial Times

• Financial Times. (2025, August 6). Big Four face AI competition from smaller firms, says former EY UK boss. Financial Times

• Fortune. (2023, July 12). KPMG is committing $2 billion-plus to A.I. — and estimates significant revenue growth. Fortune

• The Guardian. (2011, May 17). Big Four auditors face OFT consultation. The Guardian

• The Guardian. (2023, February 22). China instructs state firms to phase out Big Four auditors on data risk. The Guardian

• The Guardian. (2025, June 30). Number of new UK entry-level jobs has dived since ChatGPT launch — research. The Guardian

• H2O.ai. (n.d.). PwC: AI-powered audit [Case study]. h2o.ai

• House of Lords Economic Affairs Committee. (2010). Auditors: Market concentration and their role. manifest.co.uk

• ICAEW. (2025, July 8). FRC publishes landmark guidance on the uses of AI for audit. ICAEW

• KPMG. (n.d.). KPMG Ignite — Artificial intelligence platform. KPMG

• KPMG. (2023, December 13). KPMG global FY2023 revenues grow to US$36 billion [Press release]. KPMG

• New York State Society of CPAs. (2025, March 24). Big Four now using agentic AI to boost staff productivity. nysscpa.org

• NVIDIA. (2025, March 18). NVIDIA launches family of open reasoning AI models for developers and enterprises to build agentic AI platforms [Press release]. NVIDIA Investor Relations

• People Matters. (2025, July 20). Big Four cut entry-level jobs amid AI and cost pressures. peoplematters.in

• PwC. (n.d.). Harnessing the power of AI to transform the detection of fraud and error [Press release]. PwC

• PwC Mauritius. (n.d.). Audit and General Ledger Analysis using Halo. PwC

• Reuters. (2024, July 10). China dials up scrutiny of Big Four audit firms after Evergrande probe — sources. Reuters

• Reuters. (2024, December 19). China introduces new rule to tighten scrutiny on foreign accounting firms’ domestic operations. Reuters

• Scottish Financial News. (2025, June 23). Big Four slash graduate jobs as AI takes over entry-level tasks. Scottish Financial News

• The Finance Story. (2024, June 4). Big 4 firms scramble to win the consulting race. Investing… The Finance Story

• The Times. (2025, January). Robots could speed up audits — but they’re unlikely to cut fees. The Times

• The Wall Street Journal. (2025, August). As AI comes for consulting, McKinsey faces an “existential” shift. The Wall Street Journal