In mid-July, Satya Nadella published an essay on his personal blog with an unassuming title: “The Reverse Information Paradox.” It deserves more attention than it’s getting, because it names something many leaders have felt but few have articulated — a structural problem sitting underneath every AI subscription an organization signs.

The original information paradox comes from the economist Kenneth Arrow. To sell information, you have to reveal it — and once revealed, the buyer no longer needs to pay for it. Sellers of knowledge have always lived with that risk. Nadella’s observation is that the AI era flips the paradox onto the buyer. To get real value out of a frontier model, you have to feed it your context: your data, your judgment calls, your corrections, your edge cases, the accumulated know-how that makes your organization different from the next one. As he puts it, “You essentially pay for intelligence twice, once with money, and again with something even more valuable.”

Every prompt refined, every output corrected, every workflow encoded (Nadella calls this exhaust), and it doesn’t evaporate. It flows somewhere. The question is whether it compounds inside your walls or someone else’s.

Zoom out

Zooming out, the deeper question is not which model your organization uses. It’s where your organization’s learning accrues. When your best people spend a year teaching a rented model how your business works — what good output looks like, which exceptions matter, how your customers actually talk — that year of teaching is an asset. The uncomfortable follow-up question is who holds it.

Here is the thesis: access to intelligence is becoming abundant. Ownership of learning is becoming scarce. The models themselves are converging faster than anyone predicted. What doesn’t converge — what can’t be downloaded — is the accumulated, corrected, context-rich learning loop an organization builds around its own work. That loop is either an asset you own or rent you pay. And a growing number of serious players are restructuring around the difference.

Paying twice

Nadella’s essay is worth taking seriously not just for the diagnosis but for the asymmetry it describes. AI providers learn continuously from how customers use their systems — usage patterns, prompts, corrections — while customers gain almost no insight into what the provider extracts. Value capture flows toward the infrastructure owner, not the knowledge creator.

If that sounds abstract, walk it through your own organization. Your service team has spent the year correcting the AI's tone — thousands of small judgments about how your customers actually want to be spoken to. Your analysts have refined the prompts that finally produce usable output from your data. Your best salesperson has quietly taught a copilot which objections matter and which are noise. Now ask the uncomfortable question: if you switched vendors tomorrow, how much of that teaching would come with you? For most organizations, the honest answer is none of it. That is what paying twice looks like on the ground.

His prescription is a set of four protections organizations should insist on: control (own your private evals and memory), capability (build proprietary training environments), choice (keep orchestration independence so you can switch models without rebuilding everything), and cost (efficiency as a discipline, not an afterthought). Strip away the framing and it amounts to a simple principle: use AI without surrendering the unique knowledge that distinguishes your firm.

It’s tempting to note that Microsoft has its own commercial interests in this argument. It does. But the framing is spreading well beyond Redmond, and the market behavior backing it up is harder to dismiss.

The rental market is cracking

Consider what Hugging Face CEO Clem Delangue told TechCrunch’s Equity podcast recently. Hugging Face has become something like the GitHub of AI — the central platform where open models and datasets are shared — and roughly half of the Fortune 500 now uses it. From that vantage point, Delangue describes a recurring migration pattern: “companies start out on frontier APIs, but as they scale, the costs push them towards open source models.”

This is not an ideological conversion. It’s a lifecycle. Renting frontier intelligence is the rational way to start — fast, zero infrastructure, state of the art on day one. But as usage scales, two curves cross: the cost curve of paying per token forever, and the capability curve of open models that are now close enough for most real workloads. When they cross, ownership starts winning on economics alone. The knowledge-leakage problem Nadella describes is the second, quieter push in the same direction.

Delangue’s larger worry is concentration — “that a handful of big companies could end up controlling everything.” Whatever you make of that concern, the adoption pattern he describes is observable in the data, and we’ll get to it.

Betting against one-size-fits-all

Then there's Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, which released its first model Inkling, a large open-weight, natively multimodal system on July 15. What makes Inkling interesting is what the company didn’t claim. Thinking Machines states plainly that it is “not the strongest overall model available today, open or closed." That’s not modesty — it’s the entire thesis. As described in The Economics of Intelligence post, frontier models are not needed for most work, lesser models do just fine at much of what enterprises need. The company’s bet is that organizations customizing their own models will outperform organizations renting generic frontier ones, and its revenue comes not from the model but from Tinker, its customization platform. The model is open; the business is helping you make it yours.

One more detail rewards attention. Inkling’s post-training drew partly on outputs from other open models — including Moonshot AI’s Kimi K2.5. An American frontier lab, staffed by OpenAI alumni, building openly on a Chinese open model. That’s not a footnote. That’s what a compounding open ecosystem looks like in practice: improvements stack across organizations and borders, in public, at a pace no single closed lab can match internally.

The evidence keeps arriving

If owned intelligence were just a compelling essay and two well-positioned companies, it would be a trend piece. The adoption data says it’s more than that. Start with Moonshot AI itself. When the Beijing-based startup released Kimi K2 in July 2025, it startled the industry: an open-weight trillion-parameter model with elite coding performance at a fraction of frontier pricing. A year later, the Kimi line runs several times cheaper than comparable closed models on real workloads. And the users aren’t hobbyists. Chamath Palihapitiya moved his companies’ workflows from Amazon Bedrock onto Kimi K2, calling it “way more performant.” Airbnb CEO Brian Chesky has said the company leans on Alibaba’s open Qwen models over ChatGPT for customer service because Qwen is “fast and cheap.” South Korea’s Univa reported Qwen cut its document-processing costs by 30 percent.

The aggregate numbers tell the same story:

• Roughly a third of all tokens processed on OpenRouter went to open models in late 2025; today open models carry well over half, and eight of the ten most-used models on the platform are Chinese.

• The measured performance gap between the best open and closed models sits at about 3.3 percent (it nearly closed entirely in 2024 before the frontier labs pulled back ahead last year), while inference costs at fixed capability have been falling roughly tenfold per year.

• Meta’s Llama family passed one billion cumulative downloads. The Qwen family alone has spawned over 170,000 derivative models on Hugging Face — more monthly derivatives than any Western lab, a lead Alibaba has held since early 2025.

• Chinese open models accounted for 41 percent of Hugging Face downloads in 2025, overtaking the United States — which is why even OpenAI, the company whose name became synonymous with closed frontier AI, released its own open-weight models and watched them reach the monthly download volume of all of DeepSeek’s models combined within months.

None of this means closed frontier models are finished — they still hold the capability edge at the very top, and for some workloads that edge is worth every penny. What the data shows is narrower and more important: the default is shifting. Open, customizable, owned infrastructure has crossed from ideological preference to defensible business decision.

The geopolitical asterisk

None of this adoption is happening in a vacuum, and the numbers should not be read as an all-clear. Washington is looking at the same charts. The pending No Adversarial AI Act would bar Chinese models from federal agencies outright, this year’s defense authorization already prohibits DoD contractors from using DeepSeek in contract work, and in July, House committees began sending letters to Cursor, Airbnb, and DoorDash asking pointed questions about the Chinese models in their stacks. Nobody serious is proposing to ban private companies from running open weights — most experts concede that once weights are freely downloadable, prohibition is close to impossible — but procurement rules, entity listings, and export-control gray zones can make a Chinese model a genuine liability for any organization that sells to the government or ships data into a China-hosted API.

Notice, though, what the risk actually attaches to. Nearly every concern regulators have raised — data flowing to foreign servers, censorship baked into hosted services, security behavior of the hosted agents — is a risk of renting Chinese models as services, not of running open weights on infrastructure you control. Self-hosted weights phone home to no one. And the hedge against the ban scenario is precisely the asset this essay argues for: if a model you depend on becomes untouchable next year — by statute, by procurement rule, by your own board’s risk appetite — the organization that owns its evals and kept its orchestration portable can actually execute the switch. The organization welded to a single vendor relives its worst migration. Geopolitical risk doesn’t weaken the case for owned intelligence. It is one more argument for it.

What ownership actually means

Here is where the conversation usually goes wrong. “Owned intelligence” gets heard as “host your own trillion-parameter model,” and most organizations correctly conclude they have no business doing that. But that was never the claim — and this is where Nadella’s framing is more useful than the open-source purist’s.

Ownership is a spectrum, and the model weights are the least important thing on it. What an organization should actually own:

• Your evals. An eval is less exotic than it sounds: a set of examples pulled from your own real work — fifty customer inquiries, contract clauses, claim summaries — each paired with what a good answer looks like. Run any model against it and you know how it performs on your work, not on a vendor's benchmark. Without that, every vendor claim is unfalsifiable and every migration is a leap of faith.

• Your learning loop. The corrections, preferences, and edge cases your people generate daily are training data — yours or someone else’s. Captured deliberately, they become fine-tuning material, retrieval corpora, and institutional memory. Left as exhaust, they become a competitor’s moat.

• Your orchestration layer. The routing, prompting, and workflow logic that connects models to your business should not be welded to any single provider. Choice is only real if switching is cheap.

• Your adapted models — eventually. For organizations with the scale to justify it, fine-tuned open models running on infrastructure you control. This is the last step, not the first.

The honest objection

A fair-minded CTO would push back: this is easy to say and expensive to do. Running open models well takes talent most mid-size companies don’t have. The frontier labs still ship the best models, and the gap, however narrow, reopens with every major release. And the “pay twice” framing overstates the leakage — enterprise API agreements typically prohibit training on customer data, and most organizations’ prompts are less proprietary than their leaders like to believe.

All partly true, and worth conceding without retreating. Most organizations should keep renting frontier intelligence for most workloads today — the case for wholesale self-hosting is weak outside genuine scale. But notice that the objection only defeats the maximalist version of the argument. It says nothing against owning your evals, which costs discipline rather than GPUs. It says nothing against capturing your learning loop, which is an operating-model choice, not an infrastructure one. And contractual promises about your data, whatever their legal force, do nothing to change the structural asymmetry Nadella identifies: the provider’s system still compounds from the aggregate patterns of everyone’s usage, while your organization’s learning — unless you deliberately capture it — compounds nowhere at all.

The choice was never rent-everything versus own-everything. It’s whether the things that make your organization distinct are accumulating in a place you control.

What this means

For workers: the skills migrating into the open ecosystem — model evaluation, fine-tuning, orchestration — are becoming the trade skills of the next decade, and they are learnable now, on free tools, without anyone’s permission. More immediately: the judgment you exercise when you correct an AI’s output is valuable. Organizations that recognize and capture that judgment will treat it as the contribution it is. Seek them out, and push yours to become one.

For organizations: start with an audit, not a migration. Where does your AI learning currently accrue? If the honest answer is “our vendor’s servers and our employees’ heads,” you have an ownership problem regardless of which logo is on the invoice. Build the private eval suite first. Instrument the learning loop second. Keep orchestration portable third. The model question — open, closed, or the hybrid most will land on — becomes far easier once those three assets exist, because you’ll finally be able to measure the answer instead of guessing it.

For society: the geography of this shift matters. The center of gravity in open models has moved decisively toward China, and the Inkling example shows how quickly open ecosystems compound across borders. Whether Western economies participate in that compounding or merely consume closed alternatives is being decided now, in thousands of unglamorous procurement decisions. Concentration of intelligence infrastructure in a handful of firms — Delangue’s worry — is not inevitable. It’s the default we get if organizations keep choosing convenience over ownership one renewal at a time.

The asset you’re already creating

The reverse information paradox has one merciful property: unlike Arrow’s original, it has an answer. The seller of information could never un-reveal what they sold. But the buyer of intelligence can choose, starting today, where their learning accrues.

Your organization is generating the asset either way. Every day your people work with AI, they produce corrections, preferences, and judgment — a stream of exactly the knowledge that makes you different. The only open question is whether it compounds on your balance sheet or evaporates into someone else’s.

Owned intelligence isn’t a product you can buy, which is precisely the point. It’s a discipline you build. And the organizations that build it will spend the next decade teaching their own systems — while everyone else pays, twice, to teach someone else’s.

For most of modern business history, we have priced work around people.

A line worker is paid by the hour. A software engineer draws a salary. A consultant bills by the hour or the project. A CEO is compensated for a broad mix of judgment, leadership, accountability, and performance. The person may do a dozen different things in a day, but the economic unit stays the same: the individual. We price the role, not every cognitive move inside it.

That convention has shaped how organizations think about labor. More experience, more education, more specialized expertise, more responsibility — each one usually costs more. Compensation rises with capability, scarcity, risk, and expected value. This is the basic logic of human labor economics, and it has held for a long time.

AI breaks the unit.

With AI, we are not hiring another worker. We are renting slices of intelligence. We pay for input, output, context, reasoning, tool use, retrieval, memory, and increasingly, autonomous action. The unit of cost is no longer the person. It is the token, the model call, the agent loop, and the level of certainty the task demands. That sounds like a technical detail. It is closer to a strategic one, because as AI embeds into work, intelligence becomes something you can route, meter, and price.

The Free Trial Phase Is Ending

For the last few years, the AI market has run on a strange kind of subsidy. Twenty dollars a month bought access to models that would have seemed impossible only a few years earlier. Enterprises experimented broadly. Developers ran increasingly capable coding assistants. Teams treated AI as a cheap layer of productivity sitting on top of the organization. None of this reflected the real cost of building or serving frontier intelligence. It reflected a land grab — labs racing for users, mindshare, data, and enterprise distribution, with pricing set to win adoption rather than to make money.

Now the market is repricing. Alvarez & Marsal calls it the end of the AI flat-rate era: between late 2025 and mid-2026, the major vendors moved enterprise contracts toward consumption-based pricing. Anthropic put Claude agents on a meter, and its newer models consume more tokens per task than their predecessors, which raises cost even at the same rates. OpenAI rolled out enterprise usage analytics and spending controls. GitHub Copilot moved to usage-based billing.

The costs are real enough to change behavior at the largest companies. Microsoft pulled back internal Claude Code licenses after token billing reached roughly $2,000 per engineer per month — a case Fortune describes as exposing AI’s real cost problem. Investors eventually want the economics to make sense, and that pressure is now showing up in product decisions. The direction is consistent: providers want price to track actual compute.

This does not mean AI is becoming unaffordable. It means AI is becoming accountable. The early question was “what can this model do?” The next one is “what does this work actually cost?”

Agentic AI Burns More Than You See

The cost question gets sharper as companies move from chatbots to agents. A chatbot answers a question. An agent pursues an outcome.

To reach that outcome, an agent may break the task into steps, inspect files, search systems, call tools, revise its approach, weigh options, draft intermediate plans, and critique its own work before it returns anything. What looks like a single clean answer can sit on top of a long chain of hidden activity. The user sees the result; the system sees the bill.

This is why token economics matters. A token is the metered unit of machine cognition, and almost everything can become part of the meter — every prompt, document, search result, reasoning step, tool call, policy file, and final response. As models grow more capable, they often grow more computationally intensive. Reasoning models spend longer thinking. Agentic workflows loop through more steps. Coding agents inspect more files, run more tests, generate more alternatives. The output improves, and so does the cost. Anthropic’s Opus 4.8 burned more tokens than 4.7 and Fable 5 is burning even more.

That creates a management problem leaders haven’t had to face before: the smarter you ask the system to be, the more you tend to pay. Not always — some models will get cheaper, smaller and open-source models keep improving, and efficiency gains are real. But at the frontier, the pattern is clear enough. Deeper reasoning, longer context, more tool use, and greater autonomy consume more resources. The cost of intelligence used to be hidden inside a salary. Now it shows up in usage.

The Cost of Certainty

Not all AI work needs the same level of certainty. A brainstorm is not a legal analysis. A marketing headline is not a financial risk assessment. A first draft is not a board memo. An internal summary is not clinical decision support. Yet many organizations send every task to the most capable model they have access to.

That is expensive, and it is usually unnecessary.

We already understand this in human work. We don’t ask a senior partner to format a spreadsheet, or the CEO to book meetings, or a specialist physician to handle intake questions. We route work by expertise, risk, urgency, and value. AI deserves the same discipline. Some tasks need speed, some need traceability, some need domain expertise or regulatory compliance or a human signature — and some just need a good-enough answer at a low price.

The mistake is treating AI capability as a single ladder where the best model should always handle the most work. The better mental model is a portfolio: cheap models and fast models, specialized and open-source models, frontier models, internal and external agents, retrieval systems, evaluation layers, human review points, governance controls. The advantage comes from matching the right form of intelligence to the right work. The useful question isn’t “which model is best?” It’s “best for what, at what level of risk, and at what cost?”

The Cost of Expertise

The same unbundling is happening to expertise. In human markets, expertise is expensive because it is scarce. A senior engineer, a regulatory specialist, a physician, an attorney — each commands a premium for specialized judgment, for seeing patterns others miss and avoiding mistakes others make.

AI begins to take that apart. A general model handles many tasks at a decent level. A specialized model handles a narrow one with more precision. A workflow-specific system can encode domain rules, retrieve institutional knowledge, follow policy, and complete a business process end to end.

That capability introduces a design choice most organizations haven’t named yet. Do you pay a frontier model to reason through everything from scratch, or invest in a structured workflow that hands a smaller model the right context, rules, examples, and tools? Do you rent general intelligence at a premium, or build repeatable organizational intelligence that lowers the cost of every future run?

A high-end model is often the right call for ambiguous, high-stakes, novel work. But for repeatable workflows, the more durable advantage usually comes from codifying context, standards, decision rules, and evaluation criteria — so a cheaper model can do the job well. The company shouldn’t only rent intelligence. It should build systems that make intelligence cheaper to apply. This is where cost management stops being a procurement question and becomes an organizational design question.

From Labor Allocation to Intelligence Allocation

This is a real shift in how companies manage work. In the human economy, leaders allocate labor — deciding which people, teams, and vendors do which work, then building org charts, job descriptions, budgets, and approval flows around human capability. In the AI economy, they will also allocate intelligence: which work goes to humans, which to AI, which to human-AI teams, which earns expensive frontier reasoning, and which runs on a cheaper model or a piece of structured automation.

That is a business architecture decision as much as a technical one, and every workflow will carry a set of choices behind it. What is the outcome, and what quality does it require? What is the cost of being wrong? How much context does it need, and how much autonomy should the system have? Where does a human belong in the loop? Which model capability is sufficient, what gets logged and reviewed, and what is the acceptable cost per completed unit of work?

When AI was mostly experimentation, leaders could focus on access — give people tools, let teams explore, run pilots, find use cases. That phase was necessary. But an enterprise with thousands of employees using AI across writing, coding, analysis, support, operations, and finance can’t manage it as a novelty. It has to manage it as an operating cost, a productivity lever, a risk surface, and a source of advantage at the same time. AI adoption becomes a cost-allocation problem, not only a capability one.

The New AI Cost Curve

It’s tempting to assume AI simply drives marginal costs toward zero. Sometimes it does. A process that once took hours of human labor can compress into minutes of machine work — a service workflow, a compliance review, a software migration, a research task — all far cheaper once redesigned around agents.

But near-zero marginal cost is not zero cost. Compute, frontier models, long context windows, reasoning, tool calls, human oversight, governance, evaluation, integration, vendor lock-in, and mistakes all carry a price. And the real curve is more interesting than a slide to zero, because AI creates appetite as fast as it cuts cost — deeper analysis, more personalization, more simulation, more always-on intelligence.

Economists have a name for this dynamic. In 1865, William Stanley Jevons observed that as steam engines became more efficient, Britain burned more coal, not less — efficiency made the resource cheap enough to justify entirely new uses. The Jevons paradox is now playing out in machine cognition: as intelligence gets cheaper per unit, organizations consume far more of it.

So the result isn’t automatic savings. It’s a new management discipline — understanding where AI compresses cost, where it expands usage, where it lifts quality, where it removes labor, where it generates fresh demand, and where it hides expense. This form of intelligence accounting will increasingly be an important discipline as models increase in both usage and cost per token.

Open Source Moves to the Center

As costs rise, open-source and lower-cost models matter more. If every task leans on the most expensive proprietary model, the economics get hard to sustain. Route routine, low-risk, well-structured work to cheaper models and reserve frontier models for the highest-value tasks, and the curve bends back.

DeepSeek made this argument impossible to ignore when its R1 model delivered near-frontier reasoning at a fraction of frontier cost in January 2025, and the trend has only accelerated. OpenRouter’s analysis of more than 100 trillion tokens shows open-weight models now handling roughly a third of all token volume on the platform, and open models from DeepSeek and Tencent currently lead its usage rankings outright.

Moonshot AI’s Kimi K3 launch further shows that cheaper Chinese frontier models are competitive at lower costs proving that not all useful intelligence has to come from the most expensive source and that the market is already pricing that reality in. Thinking Machines released their open source model Inkling, further representing the need in the market for better control over models and costs.

Open source isn’t sitting on the periphery of the AI economy anymore. It’s becoming part of the control layer. That doesn’t mean every organization should rush to self-host or abandon proprietary platforms — open models bring their own trade-offs in infrastructure, security, support, compliance, and operational complexity (the Kimi K3 release is met with a potential ban from the White House). But strategically, they create leverage. They reduce single-vendor dependency, they make model routing possible, and they force a sharper question into every workflow: what level of intelligence does this actually require? The space is moving fast. In the past week The Kimi K3 launch and

The Leadership Imperative: Cost-Aware Decision-Making

The next phase of AI leadership runs on cost-aware decision-making. Not cheap. Precise. The shift is to stop treating AI as a free-floating layer of productivity and start treating it as an intelligence supply chain, where different forms of intelligence carry different costs, capabilities, risks, and constraints — and leadership’s job is to align them.

The discipline is simple to state and hard to practice: match the cost of intelligence to the value and risk of the work, keep humans where judgment, accountability, empathy, or final authority decide the outcome, wrap governance around anything where autonomy creates risk, and measure everywhere. The companies that get this right will scale AI without letting usage sprawl into uncontrolled expense. The ones that get it wrong will learn that AI pilots can be cheap while AI operations are not.

The Bottom Line

AI is not only changing how work gets done. It is changing how work gets priced.

For decades, organizations managed the economics of human labor through roles, salaries, hourly rates, utilization, and headcount. In the agentic era, they will also manage the economics of machine intelligence — tokens, model tiers, agent loops, autonomy levels, context windows, evaluation systems, and cost per outcome. Intelligence is becoming divisible. It can be deployed in smaller units, routed across systems, measured with growing precision, and priced in ways that expose the real cost of thinking, reasoning, generating, deciding, and acting.

The winners won’t be the organizations that use the most AI. They’ll be the ones that know when to spend on expensive intelligence, when cheap intelligence is enough, when only human judgment will do, and when the smarter move is to redesign the work entirely.

The future isn’t just AI adoption. It’s intelligence alignment — matching the right form of intelligence to the right work, at the right level of risk, at the right cost.

Today, most companies are still approaching AI like a software rollout. Pick a tool, run a few workshops, encourage experimentation, publish a policy, and wait for transformation to arrive. It does not arrive. AI is not another layer of software sitting on top of the organization. It changes how people think, how work gets done, how decisions are made, how knowledge moves, and how value is created.

That is why so many AI efforts feel active but not productive. Employees are experimenting, vendors are pitching, leaders are asking for updates. There is movement everywhere. But movement is not progress. Without an operating model, the organization accumulates activity instead of building capability.

Zooming out, the question leaders should be asking is not “which AI tools should we buy?” It is “what kind of organization do we need to become to absorb, direct, govern, and adapt to AI as it changes the nature of work?”

The companies that win with AI will not be the ones with the best tools. They will be the ones with the best operating model.

The Market Is Converging on the Same Conclusion

This is no longer a contrarian position. BCG surveyed 625 CEOs and board members this year and found them aligned on AI in theory but divided in practice — divided specifically on strategy, speed, and ROI. MIT Sloan and BCG’s research on the emerging agentic enterprise carries a subtitle every executive should sit with: “A Tidal Wave of Adoption, a Trickle of Strategy.” Their data across more than 2,000 organizations: agentic AI reached 35% adoption in roughly two years, another 44% plan to deploy it soon, and most of it is arriving ahead of any strategic management framework. Microsoft’s 2026 Work Trend Index names the same gap the Transformation Paradox: workers are ready, their organizations are not. IBM’s CEO put it plainly at Think 2026: “Running AI in the enterprise requires a new operating model.” And OpenAI — a model company — launched a deployment company with $4 billion behind it, a tacit admission that the models are not the hard part.

The consultancies have gone further and put structure on the claim. BCG’s operating guidance is a 10-20-70 rule: 10% of the effort is algorithms, 20% is technology and data, 70% is people and processes. McKinsey’s research on the agentic organization argues the next paradigm will be built on five enterprise pillars: business model, operating model, governance, workforce and culture, and technology and data. When the firms selling billable hours tell you the technology is the smallest part of the work, believe them.

The signals agree on two things: the impact will be large, and the work is far more than buying tools. The promised gains — growth, speed, cost structure — rest on an assumption most of these reports leave unexamined: that the organization is ready to metabolize AI into its operations. Most are not. The benefits of AI are real, but real transformation requires structure, alignment, and discipline. None of that ships with a license.

The Adaptive Intelligence Operating System

Notice what the frameworks above have in common: they describe what the agentic enterprise looks like. They are written for organizations with transformation offices and enterprise budgets. What they do not describe is how a mid-size company — where the CEO is the transformation office — actually gets there.

That is the gap this framework fills.

I think about this work through six elements. Together they form what I call the Adaptive Intelligence Operating System (AIOS): Intent, People, Work, Data & Systems, Governance, and Adaptation.

These are not abstract categories. They are the practical building blocks of AI adoption, and when one is missing, the whole effort wobbles in a predictable way. Great tools with no purpose. Executive enthusiasm with no employee trust. Automated workflows that were broken to begin with. Use cases built on messy data. Speed with no controls. Pilots that never produce learning.

That is how companies end up in AI theater — the appearance of transformation without the substance of it.

1. Intent — Why are we doing this?

Intent is where adoption should begin, and it is where most companies are least clear. Leaders say they need an AI strategy, but underneath that statement there is often no shared answer to a basic question: what is AI supposed to help this organization accomplish? Improve margins, create capacity, sharpen decisions, reduce low-value work, protect against disruption — these are different goals, and they lead to different priorities.

The failure mode is AI without strategic intent. It looks productive because there is a lot happening, but it does not compound. The organization collects pilots and tools; it does not build momentum toward a defined outcome.

The fix is a usable north star: a clear AI position statement, the strategic outcomes AI should support, the boundaries the organization will not cross, and the criteria used to evaluate opportunities. Without intent, AI adoption becomes a thousand disconnected experiments. With it, the organization can decide where to focus.

2. People — Who is being asked to change?

AI adoption is not only technical change. It is human change — and it is easier to buy tools than to help people rethink their relationship to work. The common mistake is assuming access equals adoption. Giving everyone Copilot or Claude does not mean they know how to use it well, what is expected of them, what is safe, or how their role will evolve.

The failure mode is treating AI literacy as a generic training problem. A two-hour “Intro to AI” session will not change the organization. A finance manager, an operations director, and a customer service representative do not need the same AI education — they need AI in the context of their actual work, risks, and opportunities.

The fix is role-based enablement, built on psychological safety. If people believe AI is a headcount reduction strategy dressed up as innovation, they will resist it, hide from it, or use it defensively. The people element is not soft. It is central.

3. Work — What actually changes?

Eventually AI has to touch the work itself: the workflows, handoffs, decisions, and repetitive cognitive labor that define how the organization operates. The opportunity is not doing the same work slightly faster. It is rethinking how work should be structured when intelligence is abundant — where handoffs can be removed, where humans move from production to review, where AI takes the first draft so people can focus on judgment.

The failure mode is automating fragments instead of redesigning workflows. Faster emails and meeting summaries on top of an unchanged process yield pockets of productivity, not transformation. Worse, AI applied to a broken process is not transformation at all. It is accelerated dysfunction.

The fix is to classify the work: human-led where judgment and accountability matter, AI-assisted where a person stays responsible, AI-automated where risk is low and rules are clear. That classification is what makes value measurable — “this workflow takes 12 hours a week; we believe AI can take it to five” is operational progress, not experimentation.

4. Data & Systems — What is AI building on?

AI is only as useful as the information environment and infrastructure around it. Where does knowledge live? Is it accurate, current, accessible, permissioned correctly? Or is it scattered across systems, spreadsheets, inboxes, and people’s heads? The same questions apply to the systems layer: which tools are sanctioned, how they connect, and what AI — increasingly, AI agents — can reach and act on.

The failure mode is expecting AI to create intelligence from chaos. AI does not clean up a messy information environment. It exposes it, and sometimes amplifies it — confident answers from incomplete information. A second failure mode is ignoring access: as AI moves from generating text to taking action, the question shifts from what the model can do to what it should be allowed to touch.

The fix is a focused readiness review around priority use cases — not a boil-the-ocean data transformation. What information does this workflow need? Where does it live? Is it good enough? Who owns it? What must never be exposed? Companies do not need perfect data to begin. They need to know which data problems matter.

5. Governance — How do we adopt AI responsibly?

Governance has a branding problem. People hear the word and think committees, forms, and slowdowns. But good governance is not what stops AI adoption. It is what makes adoption possible at scale — and it matters more as organizations move from tools that generate text to agents that retrieve data, trigger workflows, and take action.

The failure mode comes in two extremes. Too little governance: shadow AI, sensitive data in public tools, outputs trusted without review, leaders with no idea what is running where. Too much: every idea stuck in review while innovation goes underground. Neither works. The goal is not maximum control. It is appropriate control, tiered by risk — fast where it is safe, careful where it matters.

The fix is a simple governance model: approved tools, prohibited uses, data handling rules, review requirements, risk tiers, and accountable owners — plus a lightweight intake process for new use cases. The companies that scale AI responsibly will not be the ones that say yes to everything. They will be the ones that know how to say yes, no, not yet, and only with controls. That is governance doing its job.

6. Adaptation — How do we keep learning?

AI adoption is not a one-time transformation. The models, the risks, the economics, and the competitive landscape will all keep moving. Workflows that look advanced today will feel dated in six months. This is why companies need more than an AI strategy — they need an adaptation loop: sense what is changing, interpret what it means, decide, act, measure, adjust.

The failure mode is pilot purgatory: experiments that never convert into operational change, use cases with no measured outcomes, policies that never get updated. A close cousin is treating AI strategy as an annual planning exercise. That cadence is too slow for this environment.

The fix is a regular review rhythm — monthly or quarterly — where leaders examine which use cases are producing value, where adoption is strong or stalled, what risks have surfaced, and what to invest in next. This is where observability becomes critical: not surveillance, but a management capability. Without visibility into what is actually happening with AI across the organization, leaders are flying blind. With it, AI becomes a managed system of organizational learning.

The Real Work Is Integration

The six elements are not separate lanes. Intent gives AI direction. People determine whether it is understood and adopted. Work is where value gets created. Data & Systems determine what AI can reach and rely on. Governance creates the trust required to scale. Adaptation keeps the organization learning as the environment changes.

Focus on tools alone and you miss the system. Governance alone creates friction. People alone builds enthusiasm without measurable value. Workflows alone optimizes activity that may not matter. Strategy alone changes nothing on Monday morning.

AI adoption requires all six working together — not perfectly, but intentionally. That is the difference between companies that use AI occasionally and companies that become more adaptive, more intelligent, and more capable over time.

Leaders should stop asking only what AI tools to buy. Better questions: What should AI help us accomplish? Which roles will be most affected? Which workflows should change? What data and systems are required? What governance is necessary? How will we measure progress and adapt?

Those are not technology questions. They are operating model questions — and they are leadership’s to answer. The organizations that answer them deliberately will not just adopt AI. They will evolve with it.

If you want an honest read on where your organization stands across these six elements, that is the work I do with executive teams — reach out to learn more or start your free AI readiness assessment here . And if this framing is useful, subscribe.

In the last post I laid out four themes shaping this moment — Abundance, Asymmetry, Adaptation, and Applied Intelligence. This piece will connect those themes into the organization because the same forces that are reshaping markets are about to reshape how companies are built.

Let’s start with a pattern that has held for two centuries. In the industrial era, advantage came from organizing labor and machines. In the digital era, it came from organizing software, data, and networks. In the agentic era, it will come from organizing intelligence.

Many conversations about AI never reach that level. They stay fixed on the tools — which model is better, which copilots to deploy, which workflows can be automated. Those questions are real, and leaders have to answer them. But they describe the surface. Underneath, something more structural is happening. AI is not just adding new tools to the organization. It is changing where intelligence lives and how it moves. The companies that adapt fastest will be the ones that start treating that as a design problem — and start thinking seriously about something both old and new: organizational intelligence.

Organizations Are Intelligence Systems

Strip an organization down to its function and it does a handful of things:

• It gathers information.

• It interprets signals.

• It makes decisions.

• It coordinates action.

• It learns from outcomes.

That is a description of an intelligence system. For most of business history, nearly all of that intelligence lived in people. Knowledge lived in experience. Decisions lived in meetings. Process lived in documents and in the tribal memory of whoever had been there longest. Software changed the storage — data moved into databases, and interfaces governed how it went in and came back out. Automation handled the rule-bound edges, where machines could operate inside fixed boundaries. But the core of the work still ran on human cognition.

AI moves that line. For the first time, meaningful intelligence can sit in more than one place at once:

• In human judgment

• In AI systems

• In workflows and automation

• In codified institutional knowledge

Distribution is now the story. The leadership question shifts from which AI tools should we adopt to how should intelligence operate across this organization — across people, systems, and the seams between them. The first is a procurement decision. The second is an architecture decision.

The Missing Layer in AI Adoption

Most companies experimenting with AI today eventually hit the same wall. The models are capable. The access is there. And yet basic questions have no clear answer:

• What is AI actually allowed to do here?

• Where must human judgment stay in control?

• How should decisions be structured when a machine is part of them?

• How should work flow between people and systems?

Without answers, AI gets used in scattered ways. One person uses it for research. Another for writing. Someone automates a spreadsheet. Each use is genuinely helpful, but not scalable. The activity is real; the structure is missing.

The shift happens when an organization stops leaving this to chance and begins to define how intelligence operates. That is the point where AI stops being a tool sitting on top of the company and starts becoming part of how the company actually runs.

The Architecture of Organizational Intelligence

One useful way to see this is as a layered architecture. Different levels of the organization define different kinds of intelligence — and different constraints on it. Each layer inherits from the one above and translates it into something more specific.

General intelligence principles. The broad guidance for how AI operates anywhere inside the organization: ethical guardrails, decision philosophy, and the baseline requirements for human oversight.

Industry context. The regulatory and environmental constraints the business lives inside — healthcare compliance, financial regulation, security and privacy obligations. These set hard edges before any workflow is designed.

Company operating model. Strategy and governance: how intelligence serves the actual business objectives, where automation is encouraged, where human judgment is required, and how decisions get documented.

Departmental translation. How individual teams turn that strategy into working practice — marketing workflows, sales processes, operations systems — each with its own shape.

Role-level intelligence. How a specific role interacts with information, makes decisions, and executes: which tasks AI assists with, and which decisions stay human.

Individual adaptation. How each person actually works with AI tools and assistants in the texture of a normal day.

The point of the architecture isn’t bureaucracy. It’s coherence. Right now, in most companies, these layers don’t talk to each other. A principle set at the top has no path down to the person doing the work, and the person doing the work has no way to see the constraints they’re operating inside. Organizational intelligence is what connects them.

From Organizational Intelligence to Spec-Driven Organizations

In an earlier piece I wrote about the engineering mindset shift — how modern software increasingly begins with a clear specification rather than jumping straight to implementation. You define what you want and the constraints it has to satisfy, then let execution follow.

That same logic is starting to apply to organizations. When intelligence is distributed across humans and AI systems, ambiguity gets expensive fast. The organization needs sharper definitions of what work exists, how decisions are made, what constraints apply, and where AI is allowed to participate.

In engineering, this is spec-driven development. Applied to a company, it points toward something parallel: the spec-driven organization. Not rigid, not bureaucratic — the spec is not a binder nobody reads. It is a clear, living definition of how intelligence should operate, inside which both humans and AI execute. The structure is explicit so the work doesn’t have to be improvised every time.

As with much of AI, this realization begins at the individual level and then scales to the enterprise. A growing trend in localized (personal) AI is the concept of a second brain. Written about by Tiago Forte in Building a Second Brain and further expanded into AI by Andrej Karpathy’s LLM Wiki, Nate B Jones and others, the second brain is a digital representation of knowledge comprised of the information, rules, decision logic, consequences, and more necessary to function at the individual level. The concept is now scaling into enterprise knowledge with the release of Googles Open Knowledge Format.

Why Organizational Intelligence Matters

Organizational Intelligence is the shift that every business leader should be paying attention to in 2026. Why? Because it is the new bottleneck. As AI continues to surpass benchmarks, the next phase of AI adoption is less about experimentation and more about organizational design. The companies that get real value from AI will tend to do three things well.

They make intelligence visible. They understand where decisions actually happen and how work actually flows — not the org chart version, the real one.

They define clear constraints. They are explicit about where AI can assist, where humans decide, and how accountability is assigned when the two are combined.

They adapt continuously. They treat workflows, roles, and decision structures as things that change as the technology changes, rather than fixed once and forgotten.

When those conditions hold, AI does more than make individuals faster. It raises the intelligence of the organization itself.

The Next Competitive Advantage

For most of modern business history, advantage came from a familiar set of sources: capital, scale, technology, distribution. In the age of AI, one more factor joins that list, and it may matter as much as any of them — how intelligently the organization itself operates.

Not how smart its people are. That has always mattered and always will. The new variable is how well the organization as a whole gathers information, makes decisions, and adapts — how good it is at putting the right intelligence, human or machine, against the right problem. Companies that treat that as something to be designed will move faster than companies that leave it to chance. Which means the question worth sitting with is no longer “how do we use AI?” It is something more fundamental: how does intelligence operate inside our organization — and who is responsible for shaping it?

This is where strategy meets philosophy. Epistemology is the branch of philosophy concerned with the nature, origin, and limits of human knowledge. It the study of how knowledge is acquired, stored, evaluated, and deployed. This is the exact challenge organizations face as AI-powered agents bring great power, but require the knowledge, rules, evals, and structure to be understood, trusted, and scaled. A companies ability to leverage the power of AI to expand it’s knowledge ingestion, processing, and deployment to create value will be a differentiator as the world speeds up int he AI era.

Where to Start

Organizational intelligence sounds like a destination. In practice it’s a path, and the first steps are not exotic.

Build awareness. Get honest about what AI can and can’t do today, and where it actually fits the work you already have. Hype and fear are both distractions here; the goal is a clear-eyed map.

Document your workflows. You can’t allocate intelligence across processes you haven’t made visible. Writing down how work really flows is unglamorous and almost always revealing.

Build a culture of adaptation. The structures will keep changing. The organizations that thrive are the ones that treat that as normal rather than as disruption.

Close the feedback loops. Capture what’s working, route it back into how the system operates, and improve from there. Intelligence that doesn’t learn from its own outcomes isn’t intelligence for long.

None of this requires a finished theory of the agentic enterprise. It requires starting to treat intelligence as something you organize — deliberately, visibly, and on purpose — rather than something you hope emerges from a pile of tools. That choice is available now. The companies that make it early will spend the next few years compounding an advantage the others won’t see until it’s structural.

The Case for an Organizational Operating System

As this shift plays out, companies will need infrastructure to support it — something that helps them map how intelligence operates, document workflows and decision structures, define where AI participates and under what governance, coordinate people and systems across teams, and keep adapting as capabilities move. A recent post by Aaron Levie, CEO of Box reinforces the importance of this type of knowledge management in agentic enterprises.

This is the same realization showing up at the largest firms studying enterprise AI. At Think 2026, IBM laid out a blueprint for an “AI operating model” — its framing for how companies move from fragmented experimentation to AI that actually runs the business, connecting intelligence, action, operations, and trust into one system. McKinsey recently announced Wave, an AI Operating model as a new operating model for the AI era, built around the business model, the operating model, governance, workforce and culture, and technology and data.

The labels differ, but the conclusion converges: scattered AI experiments are not a strategy. What organizations need is a structure for how intelligence — human and artificial — works together on purpose.

The next post will explore how AI operating systems will provide the structure organizations need to scale AI and organizational intelligence.

Recently, Demis Hassabis, CEO of Google Deepmind referenced AI as “10 industrial revolutions at 10 times the speed”. The industrial revolution lasted about 100 years. What he is saying is AI could feel like 10 of those in 10 years. McKinsey echoed a similar sentiment in its recent paper on Agentic Organizations by stating ““AI is leading the largest organizational paradigm shift since the Industrial and Digital Revolutions.”

Those are some big claims. But there is some merit to big statements about the gravitational pull of AI. While AI, depending on how you define it, still has many limitations (real-time learning, inefficiencies, rising costs, etc…), it is important to remember one thing. This is the dumbest AI will ever be. Ever. This acts as a reminder to us all:

Stop looking at where AI is today, look at the trajectory.

It has been 42 months since the release of ChatGPT. 42 months. Think of what has changed. The 10 Industrial Revolutions in 10 years statement hits deeper in that context. By that logic 1 month = 10 months (0.83 years) of change pressure. Whether or not that math holds precisely, the point is directionally useful: leaders should stop evaluating AI as a static tool and start evaluating it as a compounding force. AI is increasingly compounding output, compressing time, and changing how we think about work, value, and the dynamics of intelligence.

With that framing, let’s look at 4 ways AI is reshaping our world.

Abundant Intelligence

Every major technology era changes what becomes abundant.

The industrial era changed the economics of physical labor.

The software era changed the economics of information.

The internet changed the economics of access and distribution.

AI changes the economics of intelligence itself.

This raises some questions

• How are we defining intelligence?

• What changes when we have abundant intelligence?

• As intelligence becomes abundant, what new capability must organizations develop?

As we look into the AI mirror and reflect on ourselves it is important to metabolize a simple fact - AI is not one thing, it is many things (learning, creativity, analysis, communication, discovery, planning, etc…). Likewise, our own human intelligence is not one thing. Intelligence, artificial or organic, is and has never been one single thing. In his book Frames of Mind, Howard Gardner lays out his theory of multiple intelligences. He distinguishes between different capabilities such as linguistic, logical-math, bodily-kinesthetic, interpersonal and more. As we increasingly will be asking the question - “is this a task for a human or for AI?” - it is important to stop treating intelligence as a single general human trait and understand that intelligence is neither singular nor equal. We each have different skills and different levels of those skills. So too, each of the models all have different skills at varying levels. AI is indeed making intelligence abundant, but not universally distributed and an important skill of this collaborative future will be in understanding intelligence at a deeper level. This means a shift from leaders asking “Can AI do this?” to asking “What kind of intelligence does this work require?”

AI is Asymmetric

Symmetry represents a degree of equality. I put in x and get out a relative y. Asymmetry represents the opposite; I put in x and get back y*100. This is the idea that a modest amount of AI capability that creates disproportionate leverage. With abundant intelligence and enough GPUs AI can scale many many things. This has been the idea behind manufacturing production and software for generations. It is how the Allies won WW2 (scalable manufacturing) it is how Toyota expanded globally, and how Microsoft and Salesforce built billion dollar companies. AI creates a new kind of asymmetry - one scaled on intelligence. Once one has one agent, they can have multiple agents, those agents can scale and produce more code, more feedback, more designs, more, more, and more still.

As it scales, it also it compresses time. Time is the most expensive resource on the planet We pay ourselves in time, we bid projects on time. Time to market, time to information, time for us as human beings to live our lives. The value function of time is relative, but considerable. How much would you pay to have 8 more hours in a day? This is why AI investment is a frenzy, because of this asymmetry and the impact it will have for organizations and individuals.

More Questions:

• What happens when AI can scale asymmetrically and compress time proportionally?

• What impact does that have on the billable hour?

• What downstream impacts does this have on work?

Asymmetry is neither good nor bad. AI can scale poor decisions just as fast as it can scale good ones. What is important is to understand the concept of asymmetry in our interactions with machines of scale and the proportionate opportunities and risks it brings.

Adaptability Wins

“The measure of intelligence is the ability to change.” — Albert Einstein

A truism of human history and just as critical in the age of AI is our ability to adapt to a changing world. As intelligence becomes abundant and AI scales asymmetrically, change is the only constant. That change is now happening 10x faster.

Because AI is asymmetric and can scale, it inherently shifts where value lives. The more capable the machines become, the more the human value shifts upward to orchestration, leadership, and other inherently human skills. Pressure from AI is also changing the value function away from hours worked towards outcomes. The current window between the explosion of Artificial Intelligence and our ability to metabolize it into organizations, workflows, and the value function will eventually close. There will be a market repricing and equilibrium as AI takes on more of the load. The concept of a 10x developer brings more value to a business than a 1x developer. A company producing outputs at 10x will generally beat a company producing 1x, and an economy producing 10x will draw more investment than one at 1x. Innovation inevitably shifts where value lives and markets, organizations, and people alike will have to adapt with it.

Applied Intelligence

With all of these in mind, abundant intelligence, asymmetrical AI, and the adaptability imperative is that the future will reward those that can properly apply intelligence (human and/or machine) to the world around them.

This has always been true:

Technology creates possibility, application creates value.

It was true in earlier waves of computing. And it is even more true now.

As intelligence becomes increasingly available through machines, the strategic question changes. The advantage will not come from having access to intelligence. Most organizations will have access to powerful AI. Fewer will know how to organize around it. The advantage will come from knowing what kind of intelligence is needed, where it should be applied, how much it should cost, and how much human judgment should remain in the loop.

The value function here is a critical component that needs more attention. The underrepresented part of that equation is cost. As we watch token costs accelerate (and they will continue to rise), cost must be a key consideration in applied intelligence. The argument that AI will replace humans across the board is discounting the reality that the investors in OpenAI expect material returns that will drive that AI engineering cost north to the point where human and machine intelligence are not only evaluated by output, but by cost. This reality is already coming to light in headlines as organizations shift spending between people and AI.

After all, we must remember, the value function is an optimization problem. Value = utility/cost. Beyond the hype, businesses still have to balance the books.

Ultimately, the future of AI adoption is not simply more intelligence. It is better-aligned intelligence: the right intelligence, applied to the right task, at the right level of cost, risk, and control. While some functions will increasingly be machine only, many will remain in a collaborative intelligence space in which humans apply the proper intelligence (human and/or machine) to real problems.

Application

So, with these themes established, how do we apply them?

For organizations, the work is not simply to “adopt AI.” It is to build the organizational capacity to apply intelligence deliberately. Start this by identifying where intelligence creates value in the organization: where decisions are made, where judgment is required, where analysis happens, where creativity matters, where knowledge moves, and where work slows down. Then ask what kind of intelligence is actually needed. Does the work require speed, pattern recognition, empathy, compliance, domain expertise, creativity, trust, or human judgment? Some work should be automated. Some should be augmented. Some should remain human. Much of it will live in the middle, where humans and machines collaborate. Next, figure out how to manage this intelligence across humans and machines. How do you extract knowledge, store it, and monitor the deployment of it. Finally, do the math. Make cost visible in the context of the processing and output of intelligence. Understanding not only costs today, but costs in 2 years will be an important factor in applied intelligence.

For individuals, the same logic applies at a personal level. The question is no longer simply “How do I use AI?” The better question is, “Where does my intelligence create the most value, and where can machine intelligence extend it?” Each of us will need to become more aware of the shape of our own intelligence: what we are good at, where we struggle, what drains us, what requires our judgment, and what can be accelerated or improved with AI. The people who adapt well will not be those who blindly outsource their thinking, but those who learn to work in partnership with machines while protecting the parts of themselves that matter most: taste, judgment, empathy, creativity, discernment, and purpose. In a world of abundant intelligence, the individual advantage will come from knowing yourself clearly enough to know what to amplify, what to delegate, and what to keep human.

These primary themes are central to our thesis at Adaptive Intelligence and shape how we work with helping our clients build the future and will provide context for future articles on the evolution of AI and its impact on our organizations and individuals.

Anthropic just released Agents for Financial Services.

This is not just a product announcement. It is a signal for where enterprise work is headed:

Agents + vertical knowledge + enterprise data + approval workflows = agent-driven execution of real business deliverables.

In this case, Anthropic is packaging agents for work like pitchbooks, KYC screening, earnings reviews, model building, valuation review, month-end close, and statement audits.

The question is no longer just: “Can this tool answer a question?” It is becoming: “Can this agent complete a meaningful piece of work inside the systems, standards, data, and controls of the business?”

This creates a strong potential upside:

✅ Faster turnaround time
✅ Reusable best-practice workflows
✅ Better handoff between tools like Excel, PowerPoint, Word, and Outlook
✅ More repeatable execution of complex knowledge work

It also increases the need for discipline:

🟠 Governance
🟠 Customization
🟠 Human review
🟠 Observability and auditability
🟠 Explainability around how work was completed
🟠 Changing cost structures

This is changing the narrative of the future of work from: employees using AI tools to organizations redesigning work around agent-enabled workflows, with humans still responsible for judgment, review, escalation, and accountability.

Anthropic is moving beyond generic LLMs and agents into packaging tool + digital labor to meet some objective (market research, pitch builder, valuation, audit, etc…). It does this through:

➡️ Skills
➡️ Connectors
➡️ Subagents

What this means is that:

➡️ Work is becoming modular
➡️ Expertise is becoming packaged
➡️ Execution is becoming more agentic
➡️ And governance needs to evolve just as quickly

This reinforces my Substack article on The Engineering Mindset Shift.

The engineering of work is expanding beyond software.

Finance is one example. But the same pattern is coming for healthcare, manufacturing, insurance, agriculture, legal, consulting, and every other domain where knowledge work depends on repeatable deliverables, specialized judgment, and trusted data.

This is the new paradigm.

It means leaders will need to understand their organizations in a more explicit way:

• What work is repeatable?

• What work requires judgment?

• What data can be trusted?

• What decisions need review?

• What risks need controls?

• What outcomes need to be measured?

The companies that win will not be the ones that simply “use AI.” They will be the ones that learn how to redesign work for an agentic operating environment. The people that will win will be the ones that are able to integrate these agents into their evolving roles, are able to manage work and agents effectively, and retain the human element in a rapidly changing world.

The key is to adapt and to distinguish between signal and story.

Finance is the signal. The operating model is the story.

AI transformation is now a priority for most business leaders. But moving from experimentation to real implementation has proven more difficult than many expected. Stalled pilots, growing complexity, and unresolved questions about ROI are creating frustration, wearing down patience, and forcing harder conversations in the C-suite and boardroom. So what is actually going wrong? Is it a failure of understanding, strategy, execution, or is something more fundamental missing?

Many transformation efforts do not fail because leaders lack vision. They fail because change spreads faster than leadership’s ability to see it clearly. Strategy gets approved, pilots get launched, tools get purchased, and teams begin experimenting. From the outside, it looks like progress. But inside the business, the real picture is often fragmented. Use cases multiply without a common framework. New tools appear before standards exist. Workflows begin to shift before roles are redefined. Risks build in pockets. Infrastructure strains quietly. What leadership has in many cases is not a transformation management problem. It is an observability problem.

AI makes this more urgent because it does not enter the organization as a contained change. It affects multiple layers at once. It changes how work is done, how decisions are made, what technical foundations are required, what kinds of risks emerge, and what capabilities people need to build. Recent research from MIT Sloan and BCG argues that AI adoption is moving faster than organizations are redesigning processes, decision rights, workforce structures, and governance. McKinsey makes a parallel case, describing AI as a broader organizational paradigm shift that touches business model, operating model, governance, workforce, and technology all at once.

That is why observability matters. Not observability in the narrow engineering sense of logs and technical telemetry, but executive observability into the transformation itself. Can leadership see where value is emerging, where risk is building, what tools are spreading, how work is changing, whether leaders are aligned, and whether the workforce is actually becoming more capable? If not, then the organization is not really steering AI transformation. It is reacting to it.

What follows are six areas where leadership teams need a much clearer line of sight.

1. Opportunity and ROI

Every executive team wants to know where the value is. But in many organizations, AI activity and AI value are being confused. A growing number of experiments, tools, and enthusiastic users can create the impression that progress is happening, even when the enterprise still lacks clarity on where meaningful business impact is being created. That is a dangerous gap. If leaders cannot see which use cases are tied to cost reduction, revenue growth, speed, quality, or strategic differentiation, they cannot allocate capital or attention well.

This is one of the reasons so many organizations end up with excitement but not enterprise value. McKinsey warns about the risk of having AI everywhere except in the P&L, while MIT Sloan and BCG note that many firms are adopting AI before they have the strategic frameworks needed to direct it toward real returns. Observability here means being able to see the full use case portfolio, the status of each initiative, the business problem it is meant to solve, the expected value, the realized value, and which efforts deserve scale versus which are simply generating noise.

Without visibility, leadership cannot distinguish transformation from experimentation.

2. Governance and Risk

The second area is governance and risk, and this is where many leadership teams feel the most exposed. AI introduces new questions around privacy, compliance, quality, explainability, bias, decision authority, and reputational risk. But the bigger issue is that these risks rarely emerge in one obvious place. They develop across many local experiments, tool choices, workflows, and behaviors. If leadership cannot see those patterns, governance becomes reactive.

MIT Sloan and BCG describe one of the core tensions of AI adoption as the challenge of supervising systems that operate with varying levels of autonomy. McKinsey similarly argues that governance in the AI era must become real-time, embedded, and data-driven rather than periodic and paper-based. That has a direct implication for leaders. Governance is no longer just a policy function. It is an operational visibility function.

Executives need to see what systems are in use, what data they touch, what levels of autonomy they have, what controls are attached, what exceptions are surfacing, and where human oversight is still required. Governance without observability is largely ceremonial. Governance with observability becomes control.

3. Infrastructure and Tools

Many organizations start their AI journey at the surface layer. A tool is introduced. A team experiments. A vendor adds a feature. A few workflows improve. But underneath that visible activity are deeper structural questions about architecture, integration, data quality, vendor concentration, security, and technical debt. This is where leadership can easily underestimate what it will take to scale.

McKinsey points toward an AI environment defined by more distributed technology ownership, modular systems, evolving protocols, and increasingly important proprietary data assets. MIT Sloan and BCG point to the corresponding leadership challenge: AI investments are not static. They require ongoing retraining, adaptation, monitoring, and adjustments as the technology evolves.

That means leadership observability must extend below the use case layer. Leaders need to know what tools are actually in use, how fragmented the environment is becoming, what systems are critical dependencies, where data readiness is insufficient, and whether the enterprise is building a coherent foundation or simply accumulating disconnected solutions. Tool adoption may look like progress. But without visibility into the infrastructure beneath it, scale can become fragile very quickly.

4. Operating Model

This is the category many organizations underappreciate. AI transformation is not mainly about tool adoption. It is about work redesign. It changes workflows, handoffs, decision rights, managerial spans, and the balance between execution, supervision, and judgment. If leadership cannot see that layer clearly, it will miss the most important change underway.

MIT Sloan and BCG argue that leaders now face the need to rethink workflows, governance, roles, and investment models together rather than simply automate isolated tasks. McKinsey similarly describes a move toward AI-first workflows, flatter networks of outcome-aligned teams, and humans increasingly positioned above the loop rather than inside every step of the process.

This is where observability becomes a strategic asset. Which workflows are merely being augmented, and which are being fundamentally redesigned? Where are humans still in the loop, and where are they moving above the loop? Which tasks are disappearing, which are being compressed, and which are becoming more exception-based? Those are operating model questions, not just tool questions. And they determine whether the organization is truly transforming or merely layering AI on top of old assumptions.

5. Leadership Alignment

Perhaps one of the more challenging aspects of AI transformation is around alignment. Even when good work is happening, AI transformation can still stall if the leadership team does not share a common view of what is changing and what matters most. One executive may see AI as a productivity lever. Another may see it as a governance issue. Another may focus on infrastructure. Another may worry mainly about workforce disruption. All of those concerns are legitimate, but without an integrated picture, the organization fragments at the top.

MIT Sloan and BCG explicitly argue that the old separation between technology decisions and strategy decisions no longer holds because AI simultaneously affects processes, roles, decision rights, and accountability. McKinsey makes a similar point by framing the shift as one that spans business model, operating model, governance, workforce, and technology at once.

This is a critical to adoption:

AI transformation is not a technology problem governed solely by a CIO/CTO, it is a business problem that requires alignment and coordination across the leadership team.

That means observability is not just about generating data. It is about creating shared visibility for the leadership team. Executives need to be able to look at the same transformation picture and discuss tradeoffs from a common frame. Where is value emerging? Where is risk rising? What parts of the business are ahead? Where is intervention needed? When leadership lacks that shared line of sight, AI becomes a set of disconnected conversations instead of a managed enterprise shift.

6. Change and Upskilling

No transformation becomes real until people change with it. That is why the final area of observability is change and upskilling. Organizations often talk about AI literacy in broad terms, but leadership needs a much clearer picture than that. It needs to know who is actually building useful capability, where resistance is forming, what roles are being affected first, and whether managers are equipped to lead through the transition. This form of change management is not a traditional process optimization or scaling effort, it is how abundant intelligence changes the nature of how the system operates and the balance of human+AI effort in the organization.

Both reports point to the fact that AI is reshaping roles, expectations, and workforce design. MIT Sloan and BCG note growing expectations around changes in operating models, role definitions, and organizational layers. McKinsey describes the emergence of new human profiles: AI-fluent supervisors, deeper exception-oriented specialists, and frontline employees whose human skills become more valuable as systems take on more execution.

Observability here means more than tracking training completions. It means seeing where adoption is real, where confusion is high, which teams are building capability fastest, what skills are becoming newly critical, and where support is needed before the organization falls behind its own ambitions. Upskilling without observability tends to become generic. Upskilling with observability becomes strategic.

The Real Leadership Imperative

Taken together, these six areas create something far more important than a dashboard. They create the conditions for AI leadership. Opportunity and ROI reveal whether AI is creating real value. Governance and risk show whether the organization is staying in control. Infrastructure and tools reveal whether the foundation can support scale. Operating model visibility shows whether work itself is changing. Leadership alignment determines whether the enterprise is moving coherently. Change and upskilling show whether people are keeping pace.

This is the deeper issue many organizations are now facing. They do not lack AI activity. They lack integrated visibility into what that activity means. They can see pilots, but not patterns. They can see tools, but not structural change. They can see energy, but not capability gaps. They can see policy, but not whether governance is actually operating in practice.

The organizations that will handle AI best will not simply be the ones that adopt the most tools or move the fastest. They will be the ones that can see clearly enough to make better decisions as they move. Because in a transformation this consequential, what leadership cannot see, it cannot govern. And what it cannot govern, it cannot turn into advantage.

If your organization is adopting AI faster than leadership can clearly see and manage it, that gap needs to close. I help executive teams build the visibility, structure, and operating discipline needed to guide AI transformation with intention. Reach out to explore and subscribe for more.

The image of a match strike illuminating a dark room is powerful. It represents enlightenment, information, and possibility. The light provided gives us context and safety and it starts with taking a simple action: striking a match.

Two other images circling the internet recently provide powerful context as well.

The Anthropic capability vs. observed usage graph from the Economic Impact of AI report.

And this one.

These images speak to the gap between the few AI power users and the masses. While AI gets further and further ahead the truth is that the number of people actually using it regularly relative to the total population is still small.

There are several reasons for this

• Companies roll out tools without training

• Fear and ego are powerful blockers

• Blank prompts are not great UX

A big one though really is that most people just don’t know where to start.

They approach getting started with AI as a knowledge problem. They think they need the right course, the right tool, the right prompt framework, the right use case, or the right moment. So they wait. They read. They watch. They compare. They ask smart questions about which model is best and where the market is heading.

And still, they do not begin.

That is the more important story.

Because for most individuals, and for many organizations, the biggest barrier to AI adoption is not lack of access. It is not even lack of understanding. It is behavioral resistance. It is hesitation in the face of uncertainty. It is the very human instinct to avoid what feels unfamiliar, ambiguous, or potentially exposing.

That is why one of the most important truths about AI adoption is also one of the simplest:

The fastest way to get more comfortable with AI is to use it.

Not perfectly. Not strategically. Not after a six-week plan.

Just use it. Then use it again.

That may sound almost too basic to matter. It is not. It is consistent with decades of behavioral science and with what we are now seeing in the workplace: experience changes perception faster than theory does.

Action metabolizes fear.

This means that taking action, any action, counteracts a behavioral pattern of inaction caused by fear (conscious or unconscious) by establishing a new mental model of growth.

“A journey of a thousand miles begins with a single step.”

— Tao Te Ching / Laozi

The Real Barrier Is Not Intelligence. It Is Activation.

To better understand this let’s put it in the context of physics. The topic of inertia is an underappreciated one in the context of human experience.

Newton’s first law states:

“a body at rest stays at rest and a body in motion stays in motion…”

Doing the same thing over and over creates safety and it creates one of those two conditions - we stay at rest or we stay in motion (the motion we understand). For many, anything that challenges that understanding is seen as a potential threat that triggers a fear response in the limbic system. This is an automatic response (Kahneman System 1) and directly impacts action.

While cognitively we understand this, that fear can trigger paralysis, the reality is that inaction is all around us. A lot of smart people are still on the sidelines with AI. Not because they are incapable. Not because they do not care. Not because they do not see the trend. In many cases, they are precisely the people who understand that something important is happening.

But understanding that something matters is not the same as crossing the threshold into action.

That is where adoption stalls.

In psychology, avoidance is not passive. It is self-reinforcing. The longer we postpone engagement with something uncertain, the more cognitive weight it accumulates. The task starts to feel bigger. The stakes start to feel higher. The ambiguity grows. The mind interprets delay as evidence that the thing must be difficult, risky, or unpleasant.

That is how small hesitation becomes durable resistance.

This matters for AI because many people are not resisting AI on ideological grounds. They are resisting the discomfort of beginning. They are resisting the moment of contact. The first prompt. The first awkward output. The first reminder that they may not know what they are doing yet.

So the challenge is not merely educational. It is behavioral. And that means the first move is not deeper analysis. It is action.

Confidence Does Not Usually Come First

Albert Bandura’s work on self-efficacy is useful here. His framework identifies several sources of confidence, but the most powerful is mastery experience: actually doing something and discovering, through direct experience, that you can handle it. The basic lesson is simple. People build belief in their capability less through explanation than through successful contact with the task itself.

That insight maps cleanly onto AI adoption.

You can watch demos of AI all day. You can read think pieces. You can hear peers describe how they use it. All of that can help. But none of it is as powerful as opening the tool and trying it on real work.

A workplace study in Behaviour & Information Technology found exactly this dynamic in another technology context. Workers who had direct hands-on engagement with new robotic technology developed stronger enthusiasm than workers who only observed demonstrations. In other words: participation outperformed observation.

That is not a minor point. It gets at something deeper about human adaptation. We do not become ready and then act. More often, we act, and readiness follows. This behavior applies to many aspects of life. I have heard this about startups for a decade.

If you are waiting to feel ready, you may be waiting for a condition that never comes.

AI Adoption Is an Exposure Problem

There is another lens that helps explain why this matters.

A lot of fear fades not because someone was persuaded, but because reality turns out to be more manageable than anticipated. In exposure-based psychology, newer models emphasize expectancy violation and inhibitory learning: people build new learning when feared outcomes fail to materialize in the way they expected. The old fear is not magically erased; it is weakened by lived evidence.

That is highly relevant to AI.

Many people have not actually had a bad experience with AI. They have an anticipated one.

They expect it to be confusing. Overhyped. Hard to use. Unreliable. Embarrassing. A waste of time. They expect to feel behind, or exposed, or dependent on a machine they do not fully trust.

Then they try it.

And often the result is not magical, but it is useful. A summary that saves ten minutes. A better draft. A stronger list of ideas. A clearer outline. A decent starting point.

That moment matters.

Not because it proves AI can do everything. But because it breaks the expectation loop. It creates new evidence. It changes the user’s relationship to the technology from abstract speculation to grounded judgment.

This is one of the core themes of Adaptive Intelligence: people and organizations do not adapt through theory alone. They adapt through repeated contact with new realities.

Experience Changes Belief

We can see this in the workplace data as well.

Gallup found a striking perception gap between people who had used AI in customer interactions and people who had not. Among employees with firsthand experience, 68% said AI had a positive effect. Among employees without that experience, only 13% believed it would. That is not a small difference in opinion. It is a major gap in lived reality.

BCG’s 2025 global AI-at-work research showed something similar at scale. Employees with stronger leadership support for AI experimentation were far more positive about generative AI, with positive sentiment rising from 15% to 55% when leaders actively supported use. The same report also found that regular AI users were more likely than non-users to worry that jobs could disappear over the next decade.

That second finding is especially important. Experience does not always make people less concerned. Sometimes it makes them more clear-eyed. But that is still progress. Because there is a meaningful difference between vague fear and informed concern. One is paralyzing. The other can be acted on. That is where real adaptation begins.

The Wrong Question Is “What Is the Perfect Use Case?”

This is where many people, and many companies, get stuck.

They assume the first step is to identify the highest-value use case. The perfect workflow. The most strategic implementation. The big ROI win.

Eventually, yes. That matters.

But behaviorally, that is often the wrong first move.

When people are already hesitant, adding a demand for perfect selection increases friction. It raises the cognitive cost of starting. It tells people that before they are allowed to experiment, they must first be strategically certain.

That is backwards.

The better opening move is lower stakes and higher repetition.

Use AI for things that are already in front of you:

• Draft an email.

• Summarize a long document.

• Pressure-test an idea.

• Rewrite a paragraph.

• Generate meeting questions.

• Turn rough notes into structure.

Get in the reps.

The point is not that every one of these is transformational. The point is that each interaction builds familiarity. Familiarity lowers resistance. Repetition builds calibration. Calibration builds judgment.

And judgment is the scarce resource.

As AI makes execution more abundant, the premium shifts to knowing when to use it, where to trust it, where to verify it, and where human discernment still matters most. Take note, this will be a distinguishing characteristic between AI and human value in an AI-Powered world.

Adaptive Intelligence Starts Smaller Than People Think

People often imagine adaptation as a large strategic move.

• A reorg.

• A transformation initiative.

• A task force.

• A platform rollout.

• A six-month roadmap.

But real adaptation often begins at a much smaller scale.

It begins when an individual changes behavior.

One person starts using AI to think better, write faster, or reduce friction in low-stakes work. That changes their comfort level. That changes their expectations. That changes the examples they can share with others. Over time, those experiences become local norms (social norms significantly matter in behavior change). Then team habits. Then process redesign. Then management attention. Then strategy.

This is how many changes actually spread. Not only top-down. Not only bottom-up. But through repeated behavioral contact that gradually changes what people believe is normal, useful, and possible.

From the CEO to the frontline worker:

Action metabolizes fear.

Not because action eliminates uncertainty. It does not. It metabolizes fear because it converts uncertainty into information. Before action, the mind has only anticipation. After action, it has evidence. That evidence might be positive, mixed, or disappointing. But even disappointing experience is useful, because it is now real. It is no longer vague. It can be evaluated. Improved. Scoped. Directed. That is far more powerful than indefinite hesitation.

What This Means for Individuals

Readiness is often not what appears before action. It is what emerges after a few repetitions.

So the practical advice is simple:

Use AI today on something small and real. Not because that one use will change your life. But because beginning changes your posture. You move from observer to participant. From abstract concern to direct experience. From generalized fear to specific judgment. That shift matters more than most people realize.

What This Means for Leaders

If you lead a team, the implication is bigger than “encourage innovation.” Your job is not only to pick tools. It is to reduce the behavioral friction of adoption. That means creating permission. It means lowering the stakes of early experimentation. It means making it normal to try, compare, reflect, and refine. It means understanding that hesitation is not always opposition. Sometimes it is just uncertainty without a safe entry point.

The Gallup and BCG findings both point in the same direction: people who use AI tend to see the world differently than people who do not, and leadership support materially affects whether that experimentation happens. So if you want adoption, do not lead only with policy, architecture, and ROI models. Lead with permission to begin.

The Deeper Point

The future of AI adoption will not be determined only by which model is best or which vendor wins. It will also be determined by something much more human:

Who is willing to cross the starting line.

Because this transition is not just technical. It is human. It is behavioral. It is cognitive. It is organizational. It is a test of whether individuals and institutions can update themselves through contact with a new reality.

That is the work of adaptive intelligence.

Not passive awareness. Not abstract belief. Not waiting for certainty.

Action.

Because in the age of AI, the people who learn fastest are not always the ones who studied the most first.

They are often the ones who started.

The Engineering Mindset Shift

When Anthropic announces new capabilities and SaaS stocks drop, it tells you something important.

The headlines read like this: AI is coming for software. Entire SaaS categories are “in the crosshairs.” The fear narrative is loud. But the death of software is almost certainly the wrong story. The more interesting signal isn’t destruction.

It’s a mindset shift. And it’s much bigger than code.

Claude Code and the Abstraction of Engineering

The meteoric rise of Claude Code is having an asymmetric impact on software engineering. Engineers can now describe software in natural language. Claude enters “plan mode,” generates design documents, business requirements, and architectural decisions — then spins up agents to write and test the code.

Engineers and designers still matter deeply. But the center of gravity has moved. Instead of starting with implementation, work begins with specification. Instead of manually executing every layer, humans define intent and constraints — and agents handle execution. On the surface, this looks like a productivity boost. Underneath, it is a structural reorientation of how work gets done.

The Architecture Behind It

Claude Code is not magic. It is engineering discipline encoded into an agentic system.

Software has always been rules layered on rules:

• You describe what you want built (requirements).

• You define constraints (architecture, standards).

• You write code in a structured language.

• You validate against tests and acceptance criteria.

That’s the SDLC.

For decades, we’ve abstracted lower levels of complexity. We stopped writing machine code when compilers emerged. Then frameworks abstracted even more. Then cloud platforms abstracted infrastructure.

Claude Code abstracts again — but this time at the specification layer.

The human increasingly operates at the level of:

• Intent

• Architecture

• Judgment

The AI operates at the level of:

• Implementation

• Consistency

• Speed

The critical separation isn’t human vs. machine.

It’s specification vs. execution.

And that separation is portable.

From Code to Cowork

Anthropic quickly realized something profound: the architecture behind Claude Code applies beyond software engineering. Enter Claude Cowork. It brings agentic interfaces and structured plugins to office workflows — orchestrating work across your desktop, Google Workspace, Excel, PowerPoint, DocuSign, and beyond.

What changed?

Nothing fundamental.

Work is still:

• Inputs

• Dependencies

• Rules

• Outputs

Engineering thinking simply formalizes it. Work isn’t just a series of tasks. It is a system of dependencies, handoffs, and constraints that can be designed, monitored, and refined.

That’s engineering.

Engineering Principles, Everywhere

Look at nearly any business domain.

Marketing runs on frameworks:

• The 4 Ps (Product, Price, Place, Promotion)

• Brand standards

• Channel strategy rules

• Platform algorithms

Finance runs on frameworks:

• GAAP

• EBITDA

• CAC/LTV

• Cap rate methods

Operations runs on:

• OODA loops

• Just-in-time inventory

• Demand forecasting models

Law runs on:

• Tort law

• Contract structures

• IP doctrine

• Jurisdictional constraints

These are rule systems.

They are domain-rich, layered, and adaptable.

Think of them like a recipe. There are standard ingredients for chili — but every company adds its own flavor. Global rules blend with company-specific constraints.

AI makes that blend scalable and increasingly controllable.

You can now define:

• A comprehensive brand profile

• Ethical and legal guardrails

• Pricing logic

• Financial assumptions

• Operating constraints

And from those specs, agents can generate campaigns, financial analyses, contracts, landing pages, inventory models — in minutes.

Spec-Driven Development: The Real Shift

Claude Code is built on this concept of spec-driven development. Before execution begins, you collaborate with AI to develop a validated specification.

You describe intent.

The AI asks clarifying questions. It surfaces edge cases. It forces architectural decisions early. You iterate on the plan — not the output.

Only once the spec is coherent do agents execute.

This is what “plan mode” in Claude Code actually represents: A collaborative design loop where human intent becomes machine-operable structure. The breakthrough is not code generation. It’s structured co-reasoning before execution.

Skills, Layers, and Constraints

Anthropic’s system architecture reflects this layered discipline.

At the outer layer: Constitutional AI — principles governing what the system can and cannot do.

Inside that: domain rule systems (TypeScript conventions, accounting logic, legal doctrine).

Inside that: Skills — reusable, structured capabilities like “marketing strategist” or “website designer.”

And now, companies can build their own skills.

This is composable engineering applied to knowledge work.

Enterprise workflows are becoming:

• Versionable

• Testable

• Observable

• Specifiable

That is a radical shift from tribal knowledge and undocumented process.

Why Trust Has Been the Bottleneck

Institutions like Gartner, McKinsey & Company, Massachusetts Institute of Technology, Stanford University, and University of California, Berkeley have all published extensively on AI adoption challenges.

The recurring issue is trust. Black-box systems create hesitation. AI capability have evolved dramatically for several years, but felt uncontrollable. Engineering discipline reduces it.

When workflows are:

• Specified

• Logged

• Monitored

• Tested

They become auditable. Accountability becomes clearer, not murkier.

Accountability in an Engineered World

Engineering brings observability:

• Logs

• Tests

• Monitoring

• Error boundaries

As agentic systems spread, this discipline extends to business workflows.

Accountability doesn’t disappear. It becomes structured.

The organizations that thrive will not be those who automate blindly — but those who engineer responsibly.

Agents, Parallelism, and Scaling

Once you have agent, you can have many agents and multi-agent systems are quite powerful. Claude Code can deploy multiple agents simultaneously to execute downstream work based on upstream specs. Discrete tasks become parallel threads.

Speed is not just acceleration — it’s decomposition plus orchestration (human or agentic). This isn’t incremental automation. It’s an operating layer for business logic.

Two years ago, production-level code generation wasn’t viable.

One year ago, it was emerging.

Today, AI-built systems are building production systems .(Claude Code and Cowork were built using agents).

Important to consider even when faced with imperfections and failures: Trajectory matters more than the current state. This is the worst AI will ever be.

The Engineering Mindset Shift

So what is the shift and what do we do?

It’s the recognition that:

• Work can be represented as specifications and protocols

• Processes can be dynamically orchestrated

• Business functions can be designed as composable systems

• Agents can become operational primitives

This is not the productivity feature that many think AI represents.

It’s a new operating model.

The defining skill is no longer execution. It is system design.

The Hidden Organizational Problem

Here’s the uncomfortable truth:

Most businesses don’t have their operating procedures written down clearly enough to power this shift. Processes evolved. Standards drifted. Tribal knowledge filled the gaps.

Spec-driven organizations require explicit articulation.

The real work isn’t “using AI.” It’s defining the rules your business already runs on — but never formalized.

What Comes Next

The engineering mindset spreads outward. From IT to finance, to HR, to compliance, to marketing, to operations, and more.

As agents gain context and cross-application orchestration becomes normal, the engineering mindset diffuses across the enterprise.

The future of work isn’t “work harder.” It’s design smarter.

And the companies that embrace engineering thinking — not just in software, but in how they model and orchestrate work — will define what productivity means in the AI era.

The World Economic Forum recently took this question head on and released it’s annual future of jobs report detailing four potential futures and the impacts of AI on jobs.

Zooming Out

Most conversations about AI and jobs still orbit the same question: Will AI take my job?

That question is important to the individual, yes, but it misses the broader impact.

Zooming out, we should be asking: What kind of economy are we building through the choices we make today—about skills, systems, and how humans work with machines?

The WEF paper lays this out with unusual clarity by describing four plausible futures for jobs by 2030—each shaped by two forces:

1. How fast AI capabilities advance

2. How ready the workforce is to work alongside that intelligence

The paper also identifies the real challenge/opportunity.

The technology curve matters. But human readiness matters more.

We must assume that the tech, this AI wave or the next, will continue to advance. The more important and complex question is how we adapt to it.

The Four Futures (and What They Really Mean)

The Forum doesn’t predict one future. It maps four. Think of them less as forecasts and more as guideposts and warning signs.

1. Supercharged Progress

Exponential AI + High Workforce Readiness

In this world, AI capabilities accelerate rapidly—but people keep up.

Work doesn’t disappear; it recomposes. Humans stop executing tasks and start orchestrating systems. People manage fleets of AI agents. Productivity explodes. New roles emerge faster than old ones vanish.

But there’s a catch: governance, ethics, and safety nets struggle to keep pace. This is a world of abundance and instability.

The key insight: AI literacy and systems thinking become baseline skills—not specialties.

2. The Age of Displacement

Exponential AI + Low Workforce Readiness

Same technology curve. Very different outcome.

Here, AI advances faster than education, reskilling, or organizational redesign. Automation becomes the default response to talent shortages. Jobs vanish faster than people can adapt.

Productivity rises. Profit margins expand. Social cohesion fractures.

This is what happens when organizations treat AI as a cost-cutting tool instead of a human-machine operating model.

3. The Co-Pilot Economy

Incremental AI + High Workforce Readiness

This is the least flashy—and arguably most sustainable—future.

AI progress is real but pragmatic. The hype cycle cools. Instead of chasing autonomy, organizations invest in augmentation. Humans stay in the loop. Decision-making remains human-led, AI-supported.

Work changes gradually. Productivity gains accumulate. New hybrid roles emerge—part technologist, part domain expert.

This is the future where co-intelligence is designed, not assumed.

4. Stalled Progress

Incremental AI + Low Workforce Readiness

Here, neither the technology nor the workforce fully delivers.

AI improves, but adoption is shallow. Skills gaps persist. Productivity gains are patchy and uneven. Inequality widens—not because AI is too powerful, but because it’s poorly integrated.

This is what happens when organizations dabble instead of commit.

The Pattern Hiding in Plain Sight

Across all four futures, one pattern is consistent:

AI outcomes are not determined by models. They’re determined by readiness.

The Forum’s data reinforces this:

• Over half of executives expect AI to displace jobs

• Fewer than one in eight expect it to raise wages

• Demand for AI literacy skills is already surging at an exponential rate

That gap—between expectation and preparedness—is where risk lives.

Co-Intelligence Is the Competitive Advantage

What the report never explicitly says—but clearly implies—is this:

The winners in every future are the individuals and organizations that learn how to think, decide, and operate with AI, not just use it.

Co-intelligence isn’t about prompt tricks or tool fluency. It’s about:

• Knowing what decisions should remain human

• Designing workflows where AI amplifies judgment instead of replacing it

• Developing the ability to orchestrate agents, systems, and information flows

• Embedding learning into the flow of work, not as a one-off training

In other words: AI literacy is table stakes. Decision literacy is the moat.

The “No-Regret” Moves For Both Humans and Businesses

The Forum outlines “no-regret” strategies for organizations. They translate surprisingly well to individuals navigating the future of work:

• Start small, build fast, scale what works

  Experiment with AI in real workflows, not hypothetical ones.

• Align technology with talent

  Don’t learn AI in isolation. Learn it in context—your domain, your decisions, your craft.

• Invest in human-AI collaboration

  Treat AI as a co-pilot, not an autopilot.

• Build judgment, not just efficiency

  As AI absorbs routine work, human value shifts to sense-making, ethics, creativity, and leadership.

• Stay adaptable

  The half-life of skills is shrinking. Learning how to learn with AI is the meta-skill.

The Future of Work Is Not Prewritten

The most dangerous assumption we can make is that the future of work will “just happen to us.”

It won’t.

It’s being shaped—right now—by how seriously we take AI literacy, how intentionally we design organizational intelligence, and who makes the decisions and takes on the responsibility. We are active participants in the process.

The four futures aren’t predictions. They are mirrors of our decisions, the ones we are already choosing today. Individuals must be curious, organizations must develop intentional strategy for transformation, and society must remain vigilant as the nature of work changes how economies are structured.

AI is no longer just accelerating software development — it is changing the nature of software engineering itself.

Inside leading AI organizations like Anthropic, the act of writing code line by line is giving way to something fundamentally different. Engineers increasingly operate as supervisors, reviewers, and system designers — guiding AI-generated outputs rather than producing every artifact by hand. Execution is becoming abundant. Judgment is becoming scarce.

This shift is often discussed as a productivity story, but that framing is incomplete.

What’s really happening is an operating model inversion. When AI can generate code, tests, documentation, and even architectural scaffolding at machine speed, the constraint moves upstream — away from implementation and toward decision-making, intent-setting, governance, and design quality.

The tactical execution downstream is becoming more accessible and the upstream strategic more important.

In other words, the bottleneck is no longer how fast software can be built. It’s whether organizations know what to build, why to build it, and how to integrate it responsibly into real-world systems.

This is the power of AI Augmented Software Engineering (AI ASE) and it has profound implications for executives and technical leaders alike. Team size becomes less predictive of output. Traditional measures of engineering velocity start to break down. Competitive advantage shifts from execution capacity to clarity of intent, quality of decisions, and the ability to orchestrate human–AI collaboration at scale.

AI-Augmented Software Engineering is not about replacing developers. It is about redefining where human value sits in a world where execution is cheap, fast, and increasingly automated.

In this piece, I’ll explore what AI ASE actually looks like in practice — how roles are changing, where new risks emerge, and why the organizations that succeed will be the ones that redesign their workflows, decision structures, and leadership models before AI forces them to.

Because the real question is no longer whether AI can write software.

It’s whether we are prepared to lead in a world where it already can.

The Dawn of Spec-Driven Engineering: Navigating AI-Augmented Software Development

For decades, we’ve treated software engineering as a craft of syntax—of writing the precise lines of code that tell a machine what to do. But as we enter 2026, the bedrock is shifting. We are moving from an era of manual construction to one that is AI assisted with a potential future in which it is AI Native.

For the executive, this is a massive shift in unit economics and speed-to-market. For the technologist, it’s a fundamental redefinition of what it means to “build.” The definition of “lead developer” could increasingly shift to an orchestrator of autonomous agents.

What It Is & How It Works

AI-Augmented Software Engineering is the integration of Agentic AI into the entire Software Development Lifecycle (SDLC). It is not just a smarter “autofill” for code; it is a collaborative partnership where AI handles the low-level execution while humans manage high-level intent.

At its core, AI ASE works through Contextual Synthesis. These tools ingest your codebase, documentation, design specs, and even your Slack history to understand not just what you are building, but why. When a developer gives a prompt, the system doesn’t just suggest a snippet; it plans a multi-file refactor, writes the tests to verify it, and spins up a temporary environment to prove it works.

The Power Players: From Assistants to Agents

The market is rapidly maturing beyond simple chat interfaces. We are now seeing “agent-first” environments that operate with a high degree of autonomy.

• Claude Code (Anthropic): A CLI-first tool that brings the reasoning power of Claude directly into the terminal, capable of executing complex research, planning, and debugging tasks at increasing levels of sophistication.

• Cursor: A fork of VS Code that treats AI as a first-class citizen, allowing for “Composer” modes that can write entire features across multiple files simultaneously.

• GitHub Copilot (with Workspace): The incumbent giant. Moving from inline suggestions to “Copilot Workspace,” where you can go from a GitHub Issue to a proposed plan to a Pull Request without writing a single line of manual code.

• Google Antigravity: Recently launched, it allows leaders to spawn and monitor multiple autonomous agents working in parallel—effectively turning a single developer into a manager of a small “bot squad.”

• OpenAI Codex: OpenAI’s developer-focused agent for code generation and transformation—able to take a natural-language task (or an issue-level spec), reason about an approach, and then produce concrete code changes across a project (including refactors, tests, and documentation), typically with an emphasis on correctness and iterative fix/review loops.

To bridge the gap between these powerful agents and our siloed data, the industry is coalescing around the Model Context Protocol (MCP). MCP can be thought of as a universal connector for AI—a standard that allows models to securely and instantly “plug into” your specific databases, local files, and enterprise tools without requiring custom, one-off integrations for every new agent. By decoupling the AI’s reasoning from the data sources it needs to access, MCP is rapidly becoming the foundational plumbing that allows tools like Cursor or Claude Code to act as true experts on your unique codebase.

Vibe Coding

The term “Vibe Coding” was created in 2025 as a term that represents the shift from writing code line-by-line to directing software into existence. Instead of obsessing over syntax, developers operate at the level of intent, constraints, and outcomes—describing what they want built and letting increasingly capable AI agents handle planning, implementation, and iteration. The craft moves from keystrokes to judgment: framing the right problem, setting boundaries, reviewing outputs, and steering the system as it converges. In this mode, coding starts to feel less like construction and more like composition—humans providing taste, context, and accountability, while machines supply speed, scale, and tireless execution. Vibe coding makes engineering more accessible to product owners, designers, and enthusiasts, but also comes with the dangers of creating another term from 2025: “AI Slop” - more on this later.

The Rise of Spec-Driven Development

AI ASE tools are giving rise to a new model of engineering known as Spec-Driven Development (SDD). In this model, the “source of truth” is no longer the code itself, but the Specification. Engineers describe the software more than write it from scratch, monitoring what the coding agents produce and revise their guidance and add as they go. This form engineering is human and AI Agents working together to create Co-Intelligence in the codebase. In this case, the engineer’s job shifts from coding syntax to ensuring the spec is bulletproof and that the resulting code is acceptable. Spec-driven development can be seen by using Claude Code “plan mode” to work with the AI to develop the plan before downstream coding agents write code.

The New Software Engineer: Designer of Systems

As AI changes the nature of software engineering, the value function changes for current and future engineers. While traditional syntax coding becomes automated with agents it seems it will be increasingly important for traditional engineers to embrace more of an understanding of Design. This means adopting Design Skills: understanding user experience, systems thinking, and product strategy. Engineers will no longer be limited to coding, they will be Architecting systems and products with a fleet of AI workers. The opposite is true for product designers teams. With technical tools increasingly accessible to them they would be well to learn more of the shallow technical aspects to more quickly build and iterate through design (while leaving the deep technical production buildout to engineers.

A Tale of Two Reports: The Reality of AI in 2025

Debate about AI ASE is intense and continuous. To truly understand where we are in this journey, we must look at the data. Two pivotal reports from 2025 offer contrasting yet equally crucial insights into the real-world impact of AI-Augmented Software Engineering.

The METR study, “Early 2025 AI Experienced OS Dev Study,” presents a rigorous, small-scale Randomized Controlled Trial (RCT). It served as a sobering “reality check” for individual productivity, revealing that experienced open-source developers actually took 19% longer to complete tasks when using AI, despite their own perception of increased speed. This report illuminates the “micro” friction experienced by elite individual contributors, highlighting the overhead of the learning curve, the need for careful prompt engineering, and the subtle burden of managing “AI slop”—code that requires more refactoring and validation.

In stark contrast, the Google/DORA report, “2025 State of AI-Assisted Software Development,” offers a massive global survey of nearly 5,000 professionals, viewing AI through a broader, organizational lens. This report frames AI as both a “mirror and an amplifier” for existing development practices. While METR details the individual’s struggle with immediate productivity, Google’s “macro” findings suggest that successful AI adoption is indeed driving overall software delivery throughput, but with a critical caveat: this acceleration is most pronounced when the organization possesses robust cultural and technical “safety nets.” These include clear AI policies, strong testing practices, and the foundational elements outlined in the DORA AI Capabilities Model, designed to mitigate the inherent instability and challenges that increased speed can introduce.

A key element here is to understand when navigating conflicting reports is the importance of sample size. Generally speaking lower sample size = higher variability, meaning that the results can skew more readily than in bigger sample sizes. The METR study is a small sample size of 16 developers is open to scrutiny moreso than the 5,000 respondent DORA Report. Taking sample size into consideration, the DORA report likely speaks more truth about the real impacts of AI ASE has on the future of software engineering.

Software Development Economics

If this does in fact turn out to be the case, the economics of Software Development change. In the Short-term software engineering becomes more accessible as costs drop with fewer expensive human engineers required to build software. Teams can ship products 2-5x faster, and the marginal cost of a new features as well as modernizing old systems plummets. This is the honeymoon phase of AI ASE and may last 3-5 or so years.

In the Long-term, the companies building agents have to pay their investors back and the price of AI ASE tools and Agents drives upward - perhaps even to a consumption based hourly agent cost as the concept of “digital labor” becomes more widely accepted. These costs will likely be cheaper than human labor, but AI risks must be accounted for in the cost calculations.

Challenges

AI ASE tools and agent-based coding are by no means a certainty, but they are fairly likely to continue in their adoption. Potential rewards are too high to pass up the opportunity for business leaders and engineers will gladly hand off the more mundane parts of their jobs. There are however a number of challenges remaining and questions to be answered.

Reliability

As previously mentioned the term “AI Slop” exists for a reason. This is code that works but is unmaintainable, redundant, or subtly flawed. Engineers with less experience may watch the code be written, give it a quick look, fall prey to heuristic processing and call it good. A few days or weeks later it collapses, is found to have massive security holes, or can’t be touched for fear of it breaking. This reliability issue will be a challenge for leaders and dev teams alike to approach with intention as these risks can be mitigated with proper governance, oversights, design principles and strategic imperatives.

Architectural Coherence

Because of the nature of using agents to write code and even architecture, architectural coherence could become harder—not easier—in an AI-augmented environment. When multiple agents or tools generate code in parallel, systems can quickly drift into a patchwork of patterns, abstractions, and assumptions that technically work but don’t hang together. Without explicit architectural guardrails—clear conventions, shared mental models, and enforced constraints—AI accelerates local optimization at the expense of global design integrity. The result isn’t broken software, but brittle systems that are harder to reason about, extend, and govern over time.

Human Factors

With AI writing code and architecture, there can be some unintended consequences on the human side. These challenges are less about tools and more about cognition, culture, and accountability. As AI takes on more of the implementation work, developers must shift from writing code to exercising judgment—framing problems clearly, setting constraints, and evaluating outputs they may not have personally authored. This creates an illusion of understanding, where code appears correct but its deeper logic or tradeoffs are opaque, turning debugging into an exercise in reconstructing machine intent rather than following human reasoning. Skill development also becomes uneven: junior engineers risk bypassing foundational learning, while senior engineers are forced to evolve into system stewards, reviewers, and directors of automated labor. At the organizational level, questions of ownership and responsibility become blurred—when an agent proposes an architecture or implements a critical change, accountability still rests with humans, but the decision trail is harder to trace. Layered on top of this are cultural frictions: fear of role erosion, distrust of machine-generated work, and resistance to rapidly changing workflows. Together, these human factors form the real bottleneck of AI-ASE—not the capability of the models, but our ability to adapt our roles, incentives, and norms to remain deliberate, accountable, and aligned as software creation becomes increasingly machine-mediated.

Tool Sprawl

The multitude of tools popping up is getting difficult to keep track of. New tools and frameworks continuously are released and much like traditional software engineering preferences, each developer has their own their stack. With every team member using different agents (Cursor, Copilot, etc.), organizations are losing a “unified source of truth” and could struggle with fragmented security credentials and inconsistent code quality.

Co-Intelligence In the Organization

When looking into how to leverage Co-Intelligence in the codebase, leaders should be thinking of the broader implications of what this shift towards upstream intent driving downstream execution in an AI powered world means for the rest of the enterprise. If detailed “specs” can be not only interpreted by execution agents, but also co-written by spec agents, how could this concept of Co-Intelligence be applied to other functions of the business such as Marketing, Finance, HR, Operations, and more? This question is being pondered by all organizations, including competitors. Much like computers and the internet, AI is inevitably disrupting old models. Today’s competitive advantage is no longer the ability to execute (which is being commoditized), but the ability to properly identify problems and define solutions. Leaders who can move their teams toward a “Spec-Driven” culture—where clarity of thought and precision of requirement are the primary KPIs—will win. This is where adaptive intelligence is not an option but a requirement of remaining relevant in tomorrows world.

A New World

While the future of software engineering is highly debated, one thing is evident: AI is pushing engineering, product development, and organizational strategy into uncharted territory. The next few years will likely consist of two steps forward and one-step back, but pandoras box has been opened and leaders will need to rapidly adjust to new business threats, opportunities and models of producing value for clients and customers. This will require important strategic conversations about AI adoption, retaining core business identity, and leveraging the latest innovation to stay competitive and succeed in the new AI-powered world.

Adoption of anything is complicated.

But the answer is more simple than we often want it to be.

People need a good reason.

One of the recurring challenges I’ve seen with emerging technologies is that they follow a familiar pattern. Excitement spikes. Expectations inflate. Capital flows. Thought leadership proliferates. Then reality shows up and asks a very basic question: So what?

This pattern is well illustrated by the Gartner Hype Cycle. A rapid ascent driven by promise, followed by a sharp drop when that promise fails to translate into lived value. Eventually, some technologies recover and find their footing. Others quietly fade away.

I’ve watched this cycle up close for years in XR—Virtual Reality (VR) and Augmented Reality (AR). Every new headset, platform, or breakthrough momentarily re-energizes the market, only for momentum to stall again when the novelty wears off and the friction remains. The Apple Vision Pro is an amazing platform, but even the consumer electronics giant has struggled to find a foothold in XR.

At some point along the way, it hit me:

People need a good reason to put something on their face.

Read that again, and then replace the ending.

People need a good reason to __________

– get into a self-driving car

– talk to Alexa

– use AI regularly

– trust an automated decision

– change how they work

This is not a technology problem.

This is a human behavior problem.

Technology adoption does not fail because the tools are insufficiently powerful. It fails because the reason to change is unclear, misaligned, or emotionally unconvincing.

From an individual perspective, this is why new tech is fun for 10 minutes and then fades away. It’s why VR headsets sell at Christmas and then collect dust on shelves, it’s why ChatGPT use skyrocketed, but OpenAI can’t cashflow the business. It’s why most people still don’t understand the superpower that AI gives us.

From an executive perspective, this is where many AI initiatives quietly break down.

Organizations talk about efficiency, automation, and transformation. But individuals experience disruption, risk, loss of control, and uncertainty about their role. When those two narratives collide, behavior does not change—no matter how capable the technology is.

Changing human behavior requires a reason that is:

• Personally meaningful, not abstract

• Safe enough to experiment with

• Aligned with incentives and identity

• Reinforced through daily workflows, not one-time mandates

This is why AI “pilots” often stall. People comply just enough to appear supportive, but not enough to fundamentally change how decisions are made or work gets done.

Executives often ask, “How do we get people to use AI?”

The better question is, “What reason have we given them to change?”

Until AI is framed not as a tool to be adopted, but as a capability that meaningfully improves judgment, reduces cognitive load, or protects decision quality, behavior will remain unchanged.

This is also why adoption cannot be delegated solely to IT, innovation teams, or data science groups. Behavioral change lives at the intersection of leadership, incentives, trust, and operating models. It is a systems problem.

If you are leading an organization right now, the most important work is not selecting the right AI tools.

It is answering, clearly and credibly:

• What decisions improve if AI is used well?

• What risks are reduced, not increased?

• What becomes easier for people tomorrow than it is today?

• What doesn’t change, so people feel grounded?

Reason Drives Adoption

The ones who found a good reason to put VR or AR devices on their face will do so happily and regularly. The hardware fades into the background because the value on the other side is clear.

The same is true with AI and other emerging technologies. Every day, early adopters find ways to weave these tools into their lives—not because they are more technical, but because the tools meaningfully help them think, decide, or work better.

At that point, it stops feeling like adoption. It becomes habit.

People don’t adopt technology in the abstract. They repeat behaviors that reliably produce value. When the reason is clear, personal, and trustworthy, behavior changes on its own.

Give People a Reason

If your organization is experimenting with AI but struggling to move from curiosity to consistent behavior change, pause before adding more tools.

Start by articulating the reason.

Make it human. Make it specific. Make it safe.

That clarity—not the technology itself—is what determines whether adoption actually happens.

Invent A Reason

If you as an individual are not finding ways to incorporate AI into your life (business or personal), create your own reason (ask ChatGPT or Gemini) and discover the power of AI to learn, grow, and adapt to a changing world.

Example Prompt:

I am interested in ______ and want to see how AI can help me optimize learning, produce higher quality content, and grow in some way. I need a good reason to use the tools available, explore and push through some of the behavioral psychology holding me back. Give me a convincing reason and create a plan for me to engage regularly to engage with AI tools.

Give it a go and set an intention to use the reason for 30 days and measure the impact. Remember discipline > motivation.

Connect and follow to learn more and engage with AI.

If your head is spinning about AI, you are not alone. Opposing opinions on the impact of AI on our lives and work are published every day. Confusion persists around how to approach, implement, and extract value from AI as its pace of advancement outstrips our ability to integrate it into work and daily life.

There is a simple reason for this cognitive dissonance.

AI is challenging the world that we understand.

Mental Models

Focus on the “the world that we understand” part for a moment. We understand the world around us through mental models. Much like the algorithms that human behavior inspired, we run on our own algorithms (data, patterns, understanding, behavior) called mental models. We make sense of the world by building mental models—simplified frameworks that help us interpret complexity, make decisions, and act with confidence. These mental models create order and are the basis for how we interpret and engage with the world around us.

AI is challenging those mental models at a cognitive, organizational, and philosophical level.

Cognitive: The belief that what we call thinking, reasoning, and synthesis are uniquely human domains.

Organizational: How value is created, the structure of work, and control.

Philosophical: Traditional understanding of agency, meaning, value, and society.

Destruction and Creation

These challenges to our mental model require new ways of thinking.

John Boyd, a military strategist and deep thinker, wrote a short and deceptively dense paper in 1976 titled Destruction and Creation. His aim was simple:

To show how we break apart old ideas and assemble new ones so we can act effectively in a changing world.

Boyd discusses how mental models shape our reality and how a changing world requires a cycle of destruction and creation of new models.

This article is about that process:

• Why our current concepts become rigid

• Why they break

• And why breaking them—intentionally—is the precursor to clarity

In the age of AI, it’s the operating system for AI-ready leadership.

Our Models Inevitably Stop Matching Reality

As AI rapidly evolves and markets shift, the world no longer fits the picture we built of it. Boyd points out that this mismatch as inevitable and it requires destruction of old models.

Not once.

Not occasionally.

But over and over.

This process is constant and humans have always destroyed old models in order to create new ones. We break our existing mental models (Destruction), and we rebuild new ones (Creation).

Destruction is Necessary

We love the “creation” part: innovation labs, new tech, pilots, proofs of concept. But Boyd reminds us:

You cannot create a new model until you break apart the old one.

Why?

Because if you don’t consciously unstructure old assumptions, you will unknowingly reuse the same building blocks in the same way—and produce the same results.

Said differently:

You cannot build a future using the logic of the past.

For leaders, this means some uncomfortable but necessary questions:

• Which assumptions about my business are outdated?

• What do I believe about AI that might be wrong?

• What mental models have hardened without me realizing it?

• Where am I trying to improve instead of rethink?

• Where am I protecting stability at the cost of adaptability?

This internal “destruction” of models isn’t chaos, it’s the necessary space for clarity to emerge. It is a precondition for true transformation.

In our mental model there are nodes and edges.

Nodes = Concepts

Edges = Relationships between concepts

Boyd explicitly argued that destruction happens by breaking concepts apart so they can be recombined into better ones.

Framing systems this way gives leaders a practical mechanism for change—by intentionally breaking inherited relationships between concepts rather than trying to optimize the concepts themselves.

Creation

Creation, in Boyd’s framing, is not invention from nothing, but the deliberate recombination of existing concepts into new, more coherent wholes.

“creativity is related to induction, synthesis, and integration since we proceeded from unstructured bits and pieces to a new general pattern or concept. We call such action a creative or constructive induction. It is important to note that the crucial or key step that permits this creative induction is the separation of the particulars from their previous domains by the destructive deduction.”

Once rigid structures are broken apart, individual elements—ideas, assumptions, experiences, signals—can be reassembled in ways that better reflect a new reality. This process is inherently creative, not because it produces novelty for its own sake, but because it generates new meaning through synthesis. Creation requires openness, experimentation, and tolerance for ambiguity, as not every recombination will hold. Over time, however, successful recompositions form broader, more adaptable mental models—ones capable of absorbing change, integrating new information, and guiding effective action in an evolving world.

For leaders, this means treating creation not as a burst of innovation, but as a disciplined act of synthesis grounded in reality, feedback, and experimentation.

The Cycle: Expanding From Narrow to Wide

Boyd described this process as an ever-widening spiral of:

Structure → Unstructure → Restructure → Unstructure → Restructure → …

Each cycle produces a broader, more flexible, more general mental model—one better suited to the complexity of the real world.

Over time:

• Rigid models give way to flexible ones

• Narrow thinking expands

• Local optimization becomes global strategy

• Leaders stop looking for certainty and start managing adaptation

“Over and over again, this cycle of Destruction and Creation is repeated until we demonstrate internal consistency and match-up with reality.”

This is the path from rigid organizations to adaptive ones.

Applying Boyd to AI Transformation

Here’s where this becomes essential for modern leaders.

AI isn’t a tool.

AI is a new cognitive partner.

You’re not just adding automation—you’re reshaping how decisions are made, how teams operate, how information flows, how customers interact, and how value is created.

Using Boyd’s lens, most organizations today are trying to implement AI inside a existing mental model—and this is why progress feels hard, constrained, or chaotic.

AI transformation requires the destruction of outdated assumptions.

Some examples:

❌ Old Mental Model

“We need humans to make every important decision.”

✔️ New Mental Model

“Humans design decision frameworks; machines handle scale, speed, and analysis.”

❌ Old Mental Model

“AI augments productivity.”

✔️ New Mental Model

“AI reshapes the flow of work, the structure of teams, and the nature of roles.”

❌ Old Mental Model

“Let’s look for AI use cases.”

✔️ New Mental Model

“Let’s redesign how value is created in an AI-native environment.”

❌ Old Mental Model

“We should automate tasks.”

✔️ New Mental Model

“We should reimagine business processes entirely.”

❌ Old Mental Model

“We need a strategy for AI”

✔️ New Mental Model

“We need a strategy for how humans and AI learn, adapt, and make decisions together.”

This shift is not incremental.

It is structural, cognitive, and behavioral.

It requires leaders who are comfortable breaking and rebuilding their own thinking.

Dismantling Traditional Work

An example of this in the context of AI adoption is how to fit AI into existing organizational workflows. An enterprise attempting to introduce AI into operations finds that progress stalled because work was conceptually tied to fixed roles and job descriptions. Tasks, expertise, and accountability were bundled together into indivisible units tied to roles in the organization. In order to address this, it must intentionally dismantle this traditional model by breaking roles into underlying concepts—capabilities, judgment, automation potential, and learning requirements. identifying and retaining the concepts while breaking down the edges in order to create anew.

Once unstructured, these elements can be recombined into AI-native workflows where machines perform continuous analysis, humans provide contextual judgment, and roles became fluid, adaptive constructs rather than static titles. Only once the current model has been destroyed can we effectively create.

The Leadership Mindset for the AI Era

If Boyd were advising CEOs today, he’d tell them:

Your competitive advantage is not your data, your models, or your technology.

It’s your willingness to continually destroy and recreate your mental models.

Leaders who thrive in this environment share several traits:

1. They embrace conceptual unstructuring.

They question assumptions, break patterns, and invite new interpretations.

2. They welcome ambiguity as a signal.

Uncertainty isn’t a threat—it’s evidence the old model is reaching its limits.

3. They build systems that evolve.

Processes, structures, and incentives that adapt faster than competitors.

4. They cultivate generative thinking.

They pair humans (creativity, judgment, context) with machines (speed, scale, pattern recognition).

5. They don’t cling to “what worked.”

They build what will work in a world reshaped by AI.

What This Means for Your Organization

Almost every organization today is operating with mental models formed before AI existed. They are adding AI to old decision frameworks, old organizational assumptions, old workflows, and old beliefs about what humans and machines should do. But transformational technologies don’t reward incremental updates. They reward leaders who can shatter and rebuild their ways of thinking—the way John Boyd argued we must if we want to survive and thrive.

The next 12–24 months will separate those trying to “bolt on” AI from those who are willing to rethink the fundamentals:

• How decisions are made

• How work flows

• How teams collaborate with machines

• How strategies are formed

• How value is defined

• How risk is managed

• How knowledge is created

This is not a gentle evolution.

It is a re-architecting.

And the leaders who will win are the ones who embrace Boyd’s cycle:

Destroy what no longer matches reality.

Create the models that will.

Repeat, endlessly

This cycle of destruction and creation is the path to truly adaptive intelligence.

The Cost of Clinging to Old Models

Organizations that resist this cycle don’t fail dramatically—they fade gradually.

They spend 18 months on AI pilots that produce marginal gains. Meanwhile, competitors who destroyed and rebuilt their mental models are operating at different speeds, different cost structures, different value propositions. The gap compounds monthly.

Their best talent leaves—not for better pay, but for environments where they can rethink, not just optimize within constraints they can see but can’t name.

They layer new AI capabilities onto old organizational structures, creating coordination nightmares that slow them down further. More meetings to “align” incompatible mental models. More frameworks to justify what isn’t working.

Leadership knows something is wrong but can’t articulate it. So they add more initiatives, more tools, more consultants—all while the core mental models remain intact.

Boyd was clear about this: when your model stops matching reality but you keep using it anyway, you don’t lack information—you lack the ability to process new information through an outdated framework.

The most dangerous position isn’t uncertainty. It’s false clarity—believing you understand AI because you’ve deployed tools, while your fundamental assumptions about work, decisions, and value creation remain unchanged.

This is how market leaders become cautionary tales.

Final Thoughts: AI Requires a New Way of Seeing

Boyd wrote that survival depends on our ability to:

“Generate a changing and expanding universe of mental concepts that match a changing and expanding universe of reality.”

That sentence could have been written for this moment in history.

AI is not asking us to simply upgrade our systems.

It’s asking us to upgrade our thinking.

To let go.

To rethink.

To restructure.

To expand our conceptual universe.

Leaders who embrace this cycle will not only navigate AI transformation—they will define it.

But here’s the challenge: destruction and creation are easier to understand than to execute. Most leadership teams are so embedded in their current mental models that they can’t see which assumptions need breaking. The system resists examining itself. This is where external perspective becomes essential—not to provide answers, but to facilitate the unstructuring process itself. My work focuses on partnering with leadership teams to identify which mental models are constraining progress, deliberately dismantle them, and rebuild adaptive decision-making frameworks. The goal isn’t just AI adoption—it’s building organizations that continuously evolve their thinking as fast as the technology does.

If your team is ready to move beyond incremental AI pilots to fundamental transformation, let’s connect.

The biggest barrier to AI adoption isn’t technical — it’s human.

AI is evolving fast — but humans are still far more complex, unpredictable, emotional, and ultimately still in control (for now). And that’s exactly why AI adoption is not a technical problem — it’s a human one.

Most organizations already own the tools. What they don’t have is the psychology, culture, and leadership to use them.

This article will explore The Human Side of Ai Adoption based on a recent talk presented at the Applied Ai conference at the University of St. Thomas. The talk focused on human psychology over technology underpinning the belief that the tech is rapidly advancing while human evolution and adoption are much more complex.

Human vs. Machine

While Ai rides the headlines, humans have actually been quite effective for a long time. The reality is that both humans and machines bring considerable benefits (and complexities) to the table. While the rush is to automate as much as possible, the narrative is continuously shifting between how effective Ai is at producing work vs. consistently producing quality work. A recent study from Stanford and Carnegie Mellon compares human and Ai workflows and finds that humans ensure work quality while machines optimize work efficiency.

This shows both the potential and limitations of current Ai systems relative to human output. Ai is great at producing, but it is limited in its reasoning capabilities to make consistent quality outcomes. As to how “intelligent” Artificial Intelligence is today is a hot topic of debate to cover in a future post. A brief comparison of human and machine strengths could be viewed as follows.

In other words: humans are still the quality layer — and will be for the foreseeable future.

Adoption

That said, Ai is clearly a human species level disruption and will evolve, perhaps faster than we will be able to fully understand. It will be adopted along darwinian terms (adapt or exit the pool) as with most innovations across human evolution. This adoption is however unevenly distributed and is under intense debate. Conflicting views on adoption and value are catching headlines as investment in Ai grows exponentially. A recently published Wharton study - a now 3 year study on adoption of Generative Ai in enterprises shows 82% of business leaders using Gen Ai weekly. The conflicting MIT NANDA report cites 95% of business are not seeing the expected ROI from Ai investment.

Gartner, McKinsey and the other big consulting firms that predicted Ai adoption are also reporting that 70% of AI initiatives fail.

Here’s the thing

Ai adoption isn’t a technical challenge — it’s a human one.

Not because of data or models — but because people, culture, and systems aren’t ready for what the technology demands.

We’ve prioritized technology over humanity.

And in doing so, we’ve created an imbalance that no amount of code can fix.

The key is to leverage the best of both worlds, what Ethan Mollick refers to as Co-Intelligence by blending the strengths of humans and machines. To properly do this, we must first understand ourselves as much as we understand Ai.

Human Psychology

Despite the impressive capabilities of Ai, humans are vastly more complex than machines. We have entire libraries and schools filled with exploration and understanding of human psychology, social mechanics, and behavioral dynamics. We have a subconscious and cognitive biases that we are not even aware of. We make snap judgements, have an intuition, and we dream.

Daniel Kahnemans book Thinking Fast and Slow details system 1 and system 2 thinking. System 1 thinking is fast, automatic, and intuitive, operating without conscious effort, while System 2 thinking is slow, deliberate, and analytical, requiring conscious effort and reasoning for complex tasks. The interplay of these systems is constant, unique to individuals, and a basis for behavioral dynamics that occur within us as individuals, as parts of an organization, and society at large.

Behavioral Dynamics

Behavioral dynamics gives us a way to understand how people actually change—how their habits shift, how teams respond to new tools, and why organizations adopt (or resist) new technologies. It looks at the patterns beneath our actions and the forces that influence them, from personal motivations to social pressure to the design of the environment around us. What’s useful here is that behavioral dynamics blends insights from psychology, complex systems, physics, and statistics to explain not just what people do, but how those behaviors evolve over time.

Every organization sits atop four interdependent layers of behavior:

1. Emotional – Fear, trust, and psychological safety

2. Cognitive – Biases, heuristics, and meaning-making

3. Cultural – Group norms and collective emotion

4. Structural – Systems, incentives, and governance

Anyone looking at adoption of systems (printing press, steam engine, computers, internet, Ai, etc…) would be wise to explore how these behavioral layers work to optimize for success.

The following sections will look at each of these along 3 key aspects.

1. Brief context (definition)

2. How it manifests with Ai

3. What human-centered Ai looks like

1. Emotional Dynamics: Safety Before Strategy

Humans are wired for survival, not change. When uncertainty spikes, the amygdala sounds the alarm long before the prefrontal cortex can reason. Fear, anxiety, and loss aversion take the wheel. Signals may not even make it to the pre-frontal cortex for logic-based reasoning.

How it Manifests with Ai:

“Will this replace me?”

“I don’t understand it.”

“What happens to my role?”

For many, Ai is scary, complicated and confusing. People can’t listen when they are afraid. CEO’s shouting to use Ai or leave is not particularly helpful in creating an environment of safety.

Human Centered Ai

Create Safety

Psychological safety is the foundation of AI fluency. Until people feel safe, they can’t learn, collaborate, or innovate. Leaders who begin with empathy, transparency, and reassurance calm the nervous system so curiosity can return. Only then does experimentation — the lifeblood of AI adoption — become possible. Start by creating a constructive environment of safety and support to reach them.

• Normalize experimentation and small wins

• Reward learning over perfection

• Empower “AI champions” across departments

• Create safe spaces for trial, error, and dialogue

• Properly educate on how to use Ai

Build Trust

Building trust with machines is full of challenges as detailed in a previous post.

• Transparency → Explain decisions, show limits

• Empathy → Understand fears and show respect for human expertise

• Model Behavior

• Competence → Deliver reliable results

• Replace “AI will replace jobs” with “AI will reshape work”

• Communicate purpose first, technology second

• Be specific about what will change and what won’t

• Share success stories early and often

2. Cognitive Dynamics: Bias, Ambiguity, and Oversimplification

Our brains crave certainty. They rely on shortcuts — heuristics — to reduce cognitive load. Under stress, Kahneman’s System 1 dominates.

That’s why even brilliant teams fall into:

• Ambiguity aversion: “Let’s wait until it’s proven.”

• Oversimplification: “AI = automation = job loss.”

How it Manifests with Ai:

Heuristics

One way this manifests is through what are known as “heuristics”. Heuristics are mental shortcuts our brains use to make quick decisions with limited information, trading accuracy for speed and efficiency. The book “Emotions and Affect in Human Factors and Human-Computer Interaction“ includes the excerpt “Heuristic processing refers to constructive but truncated, low-effort processing, which is likely to be adopted when time and personal resources such as motivation, interest, attention, and working-memory capacity are scarce”. Ai introduces circumstances in which heuristics can flourish. One need only look at Vibe Coding to see this at play. A junior developer asks Ai to build a platform and says “Looks good to me - ship it!” then quickly realizes they cannot understand what it’s doing” (as outlined in this article).

Dunning-Kruger Effect

This is a microcosm of what is known as the Dunning-Kruger Effect. The Dunning–Kruger effect shows up when people with limited knowledge think they know far more than they do, while those with deeper expertise often question themselves. In Ai adoption, this occurs all the time: leaders with a surface-level understanding assume Ai is plug-and-play, set unrealistic expectations, and underestimate the complexity of real integration. Meanwhile, the people who actually understand the nuances—data, workflows, governance, behavior change—tend to be more cautious and sometimes even doubt their own expertise. That gap between perceived ability and actual capability creates misalignment, slows progress, and ultimately limits the real value AI could deliver. The impact can be substantial as poorly informed decisions scale poorly.

Perfectionism: The Polite Face of Fear

Another cognitive ‘feature’ is that of perfectionism. Perfectionism feels like professionalism — but it’s often fear in disguise. It can be paralyzing and hold teams back.

In Ai adoption, it sounds like:

“We’ll launch once the model’s perfect.”

“We can’t experiment until the policy’s final.”

Cognitive Load

Cognitive load also plays a critical role in adopting new technologies. The challenge is that our brain has some limitations in knowledge gathering. Cognitive Load Theory suggest that we have a limited amount of working memory. As we learn new and challenging things we can hit a wall. A previous post goes into detail on how overloading the brain shuts us down.

Human Centered Ai

Awareness

Countering the effects of cognitive dynamics starts with awareness. Awareness of ourselves when exploring and making decisions about technologies we don’t fully understand. Awareness of what others are aware of when doing the same and asking the appropriate questions necessary to navigate heuristics, cognitive biases, and gaps in understanding. Designing appropriate frameworks for decision making and creating feedback loops can help improve awareness of biases. For example, heuristics are shortcuts that can make change feel safer by closing mental loops prematurely. But learning requires keeping those loops open and a sense of curiosity. That curiosity isn’t a personality trait; it’s a practiced cognitive discipline. Reading this post and books like Thinking Fast and Slow are ways in which to build awareness of our behavior (often misunderstood) so that we can do something about it.

Environment

Once an awareness is built, leaders can then design environments that slow thinking down — where reflection and inquiry are rewarded as much as execution. This creates an environment that fosters open conversations and iterative learning. Perfectionism and cognitive load can be overcome through openness and an iterative mindset. Throughout, it is imperative that leaders listen to experts and workers alike in the adoption process, create measurements that matter, and don’t succumb to their own cognitive biases.

3. Cultural Dynamics: The Illusion of Alignment

Culture is collective emotion. It defines what feels safe to say — and what doesn’t. People can form culture extremely fast, and from many variables they may or may not even be aware of. Culture is complex, personal, and based on experience. Individually it can be based on pride, fear, and other system 1 “feelings”.

How it Manifests with Ai:

Groupthink

We are all subject to groupthink. A confident CEO says “we are going to adopt Ai and gain the upper hand in market adoption”. The leadership team understands the mandate and acceptance masquerades as consensus. As the pressure of using Ai mounts, teams nod along, executives proclaim momentum, and dissent quietly disappears. Critical thinking goes out the window as people defer to “experts” and move things along. The result is ethical blind spots, fragile adoption, and missed innovation. Groupthink isn’t just happening at the organizational level, it is happening at the societal level as well. There is a lot of pressure to figure out how to use Ai from peer groups (developers, marketers, etc…), the providers (OpenAi, NVIDEA, Microsoft) and in the media.

Cultural Blind Spot

“Culture eats strategy for breakfast”

Misalignment in organizations happens often. Not just misalignment, but an environment in which leadership isn’t even aware that there is misalignment. This can happen particularly in technology where well intentioned and intelligent technologists elevate to decision making positions while their incentives are based in technology without a fundamental understanding of human group dynamics. This overemphasizing the tech while ignoring the human element and causes a lot of pain and wastes a lot of time and money.

Human Centered Ai

Cultural Awareness

Here again, it is is imperative to understand some basic human psychology and behavioral dynamics. Astute leaders can “feel” when something is off and are able to manage through challenges in teams, personalities, skills, and implementation. Understanding why groupthink happens, how to address it, and the dynamics of groups is an important step towards being able to do something about it. Also helpful is to understand how the innovation pendulum swings and how to balance an innovative mindset with a bit of healthy skepticism.

Normalize Dissent

This healthy skepticism should also be fostered across teams. Healthy Ai cultures normalize dissent. They assign devil’s advocates, invite diverse voices, and treat disagreement as due diligence. To combat groupthink it is imperative to retain these critical thinking abilities, especially so with hyped technologies. This means asking questions of peers, leaders, and others. It includes voicing an opinion that may be contrary to the Ai hype that is tall on dreams and low on implementation details.

If your Ai committee never argues, your algorithm is probably biased.

4. Structural Dynamics: Systems Shape Behavior

Even when people want to change, structures often won’t let them. Most organizations reward efficiency, not learning. KPIs favor output over insight. Budgets measure activity, not adaptation.

How it Manifests with Ai:

Broken Reward Systems

The simple breakdown of reward systems are to use the carrot or the stick. When forced (stick), people will do the minimum and unconsciously resist. ROI is poor. When inspired (carrot), people proactively learn, seek growth, and produce outsized returns.

An example of this can be seen in a number of recent all company emails sent from CEO’s mandating a “use or AI or get fired” policy like this one from the freshly hired CEO of Opendoor.

Human-Technology Imbalance

Broken reward systems can be a reflection of prioritizing technology over humans. Ai systems optimize for speed, efficiency, and automation — but while we invest heavily in the tech, we under-invest in trust, learning, and meaning. The result is that Technical readiness outpaces human readiness.

Human Centered Ai

Rewards Systems

Ask yourself “What are the incentives for people in your organization to adopt Ai?" If you expect that software engineer to do the work of 5 people, and they come somewhere close to that, do you expect to keep them around if they don’t have meaningful incentives for them? Competing companies will be offering those incentives. Make sure incentives align with expectations as those expectations change with new systems.

Rebalance and Rethink

Evaluating and rebalancing the human-technology dynamic means designing structure and systems for people — their psychology, cognition, and culture — not just their output. Failing to do so will create the unintended consequence of Ai uncovering how broken current systems and efforts really are. This also includes rethinking structure and systems altogether. In a world in which Ai can take on more of the load, it is forcing us to rethink how the system is architected in the first place. Job responsibilities will change, corporate structures must evolve. Many organizations are creating centers of excellence that crosscut departments (data + ops + HR + design) into strategic teams to map out how best to incorporate Ai into the enterprise.

Measure what Matters

We measure what’s easy — not what matters.

Measuring progress is not always easy or straightforward. Many times this manifests as vanity metrics. Impact comes from integrating technology into human systems, not counting how many models built and deployed. Leaders must redesign systems around learning velocity — how quickly the organization turns information into improved behavior.

Business Risks of Ignoring Human Factors in AI Adoption

The risk/reward calculus is not just in what Ai can help improve, it also carries risk. Getting the human side wrong can create problems that are equally magnified by Ai.

• Misallocated investment

• Talent retention issues

• Shadow workflows

• Compliance blind spots

• Model misuse

• Organizational cynicism

• Competitive disadvantage

Leadership for a Human-Centered Future

As with most things in organizations, all of this comes down to strong, informed, and balanced leadership. As Ai evolves, leaders are not just process owners, but are increasingly behavioral architects. They understand that while technology moves fast, humans move with meaning. This starts with appreciating the human side of Ai adoption, the behavioral dynamics at play and that Ai is an amplifier. It can amplify what makes us human — creativity, empathy, and connection, or it can amplify noise, power dynamics, and expose failures.

AI doesn’t make organizations more intelligent.

It reveals whether the organization was intelligent to begin with.

The future of AI isn’t artificial — it’s profoundly human.

While technical aspects like data maturity are critical to incorporating Ai, psychological maturity is equally important. Ai success will belong to those who understand both the machine’s logic and the human mind behind it.

Ready to learn more? Connect and see if your team is ready for Ai adoption.

Last week I was fortunate to attend The AI Conference 2025, a gathering of researchers, entrepreneurs, and executives exploring the state of artificial intelligence in San Fransisco. If last year was about experimentation, this year was about collision—between hype and hard reality, innovation and infrastructure, ambition and ethics.

The conversations revealed not only where AI is advancing, but also the friction points shaping adoption. Much like the Agentic AI debate, the themes of this conference weren’t confined to technical breakthroughs. They cut across human, organizational, and societal dimensions that will determine AI’s long-term impact.

Beyond Models: The Expanding Frontier

Large language models still capture attention, but the real momentum is moving toward what comes next: agentic intelligence, compact architectures, domain-specific foundation models, and multi-modal systems capable of text, image, video, and audio generation in unified workflows.

The challenge is clear: these systems remain costly, power-hungry, and prone to brittleness. Efficiency and alignment are no longer side projects—they’re prerequisites for scale.

Constructive Action: Begin experimenting with smaller, domain-specific models. For teams or organizations, focus on matching model choice to real-world use cases rather than chasing the biggest architecture. Think long-term architecture and how to productionalize Ai systems.

Infrastructure and the Unseen Battle

Conference sessions repeatedly underscored that the flashiest demos mean little without robust scaffolding. Data is great, but of limited value if Ai can’t use it. MLOps pipelines, data governance, deployment frameworks, and regulatory readiness are now central to the AI conversation. This includes data quality, visibility, and observability.

Just as cities need roads and utilities, AI demands the unseen infrastructure that makes systems usable and trustworthy.

Constructive Action: Invest in data pipelines, monitoring, and governance early. Building infrastructure alongside innovation prevents costly rebuilds later.

Ethics and the Trust Imperative

Bias, privacy, alignment, and dual-use risk were no longer sidebar debates—they were front and center. If AI is to support healthcare decisions, financial trades, or autonomous logistics, the question is no longer “Can it work?” but “Do we trust it to act safely?”

Trust is not earned by capability alone. It emerges from transparency, accountability, and consistent track record. Agent observability was front and center at the conference, speaking to the importance of knowing what the Ai is doing.

Positive Action: Define clear trust thresholds. Decide in advance which tasks can be automated and where human review must remain in the loop. Document and communicate these guardrails to build confidence.

Applied AI: From Hype to Proof

Panels on healthcare, biotech, and logistics showed a pivot from speculation to evidence. Case studies demonstrated early wins in clinical imaging, fraud detection, and supply-chain optimization. Yet, references to the MIT NANDA report on the State of Ai in Business 2025 were made, citing “High curiosity, low conversion”—proof that adoption is still uneven.

The signal: progress is real, but disciplined integration matters more than hype.

Constructive Action: Start small. Pilot AI where outcomes are measurable, then expand based on evidence. Build proof points before scaling enterprise-wide. Consider all aspects of getting to a production implementation.

The Startup Signal

The floor was filled with startups. The question for investors and founders alike: Where does durable value emerge? The consensus was pragmatic. The winners will be those who marry technical capability with workflow integration, cost discipline, and domain expertise—not just those who chase headlines. Ai is an equalizer and moats are increasingly challenging. This suggests a shift from “AI-first” pitches to “problem-first” execution.

Constructive Action: If you’re building, anchor your value proposition in solving a concrete problem. If you’re investing, prioritize startups that integrate seamlessly into existing workflows.

Impact and Action

For workers:

AI is shifting from novelty to tool. The question will soon be: how are you using AI to extend your skills? Those who treat AI as collaborator, not competitor, will thrive.

• Action: Learn how to apply AI directly in your role—whether for research, drafting, or analysis. Small gains now compound into career resilience.

For organizations:

The winners will be those who adopt responsibly: creating trust guardrails, preserving human strengths, and balancing efficiency with engagement.

• Action: Establish an AI adoption framework that pairs productivity goals with cultural and ethical safeguards.

For society:

As AI moves into sensitive domains, we must set thresholds for autonomy and ensure empathy and creativity remain central to human identity.

• Action: Advocate for regulation that balances innovation with safety, and stay active in shaping the narrative around how AI should serve humanity.

Closing Thoughts

The Ai Conference 2025 wasn’t about a single headline announcement—it was about signals: where AI is accelerating, where it is stalling, and where human judgment must catch up. The need to transition from experimentation to deployment is clear but getting there is more complex than many anticipate - a perennial challenge with technology as seen in the ever recurring hype cycle.

Ai will be transformational. But the path forward will be shaped as much by infrastructure, trust, and human creativity as by algorithms. The opportunity before us is clear: adopt with care, build with purpose, and ensure AI serves as a collaborator in shaping a more resilient, human-centered future.

Follow for more on Ai, Agents, and Adoption.

Agents are everywhere. Salesforce, AWS, OpenAI, Gartner, Accenture—everyone has a solution, a strategy, or a use case for AI agents. The hype cycle is peaking, with promises of production-ready pipelines, turnkey automation, and agents that can replace human labor. Gartner’s 2025 Hype Cycle for Generative AI places Agentic AI squarely at the Peak of Inflated Expectations, and yet research from Carnegie Mellon and Duke shows agents fail at 70% of assigned tasks. Gartner further predicts that by 2027, 40% of Agentic AI projects will be canceled due to escalating costs, unclear business value, and inadequate risk controls. MIT’s State of AI in Business 2025 reports that 95% of businesses surveyed said that Ai is failing to deliver expected value.

While challenges loom, even a 30% success rate in Ai tasks today points to massive potential in the long run. Despite the hype cycle doom and gloom about practical implementation, the nature of work is changing across industries as Ai rapidly destroys benchmarks across the board. Zooming out, Agentic Ai could be a transformation on the scale of the industrial revolution, but realistically it will have to go through substantial challenges along the way.

This article will examine the first of four on Agentic Ai considerations and challenges across four dimensions:

• Human: Trust, reasoning, empathy, and the user experience gap.

• Technical: Model selection, security, data quality, and evaluation.

• Organizational: Workflow readiness, change management, and costs.

• Societal: Ethics, job displacement, unintended consequences, and regulation.

Understanding these barriers is critical—not to slow adoption, but to guide it toward sustainable, safe, and valuable outcomes. Let’s begin by considering the differences, similarities, and challenges when comparing Ai and humans, and the mechanics of Agentic Ai adoption.

Human

Reasoning: The Perception vs. Reality Gap

A central question in the agent debate is whether large language models (LLMs), large reasoning models (LRMs), or any other architecture can truly “reason”. The recent back-and-forth between Apple’s The Illusion of Reasoning paper and Anthropic’s The Illusion of the Illusion of Thinking highlights the nuance.

The debate exposes a critical distinction between actual reasoning (systematic, logic-based thought) and perceived reasoning (outputs that appear logical to humans). Agents can often produce convincing results, but those results may not come from a reasoning process comparable to human cognition. This distinction between perception and reality matters for trust, adoption, and the types of decisions we allow agents to make. In workflows or high-stakes decision-making, the difference between genuine reasoning and a persuasive illusion can be the difference between a breakthrough and a costly mistake.

Trust: Deciding When to Delegate

As discussed in a previous post on Ai's Trust Problem, trust is one of the foundational principles of human society. While trust is critical to human interaction, how do we trust the decisions an Ai Agent makes? As human-to-human trust is a complex mix of factors, the answer will similarly depend on the complexity of the decision, acceptable risk levels, the agent’s track record, and influence.

Complexity: While low complexity means fewer decisions and things that could go wrong, higher complexity requires deeper levels of trust. Consider high-stakes examples: medical diagnoses, self-driving vehicles, financial trades, manufacturing controls. Trust in each of these areas can be extremely complex.

Acceptable Risk Levels: While we have high benchmarks for Ai, Humans don’t have a perfect track record in many areas of work or life. We accept certain levels of risk every day. Even if agents don’t achieve perfect accuracy, they are already outperforming humans at certain tasks. With human in the loop for review and guidance, we will increasingly grow more comfortable with accepting certain risk levels the same way we do with human work. More complex, mission critical or safety related work may be more challenging to adopt and could hinge on economics and regulation. For example, once self-driving cars are statistically safer than human drivers, insurance incentives may push consumers toward them. Price will also come into play. Even if the Ai doesn’t get something perfect, can it do it better than a human at the price I am willing to pay?

Track Record: A strong track record serves as crucial evidence of an Ai system's reliability and safety when organizations consider deploying Agentic Ai solutions. Since Agentic systems operate with greater autonomy and decision-making authority than traditional systems, potential adopters need concrete historical evidence that the agent can consistently perform as intended, handle edge cases appropriately, and maintain alignment with human values across diverse scenarios. Without demonstrated performance over time and across varied real-world applications (difficult to find at this early stage of AAi), organizations face significant uncertainty about whether an agentic system will act predictably and beneficially when given increased operational independence.

Influence: Trust is also not created in a silo. It is often built on a number of data points from those around us and their experience. A very human characteristic is to extend trust through the experience of others. For example, Many may not step into a Waymo self-driving car until they hear a trusted resource speak to the safety features.

Just as trust is shaped by shared experiences and social proof, our willingness to work alongside AI agents will also depend on how convincingly they can mirror the human qualities—like personality and empathy—that make collaboration feel natural.

Consciousness, Personality, Empathy, and Relationships

Humans are accustomed to collaborating with other humans—a dynamic that shapes our expectations for Ai interaction. Our subconscious communication and personalities are part of what makes the world go around. While Ai is indeed improving its personality with conversational aspects like “great question!”, it is still far from being a coworker with a personality that connects with the human condition. This is because it can’t truly empathize - at the end of the day it’s a machine running statistics.

Ai (LLMs, agents, or otherwise) cannot feel. Ai agents don't have a family. It can't talk about how the kids are doing in baseball or hockey, the latest series on Netflix they saw or how difficult it is to build a startup. It can't relate to past experiences, physical pain, or love. We don't have the same interest in impressing or pleasing Ai the way we do with other humans. It's not part of a societal class or group, it can't help us network, and it’s not impressed at that high school football trophy.

It is however being programmed to appear to have empathy and intuition and the illusion of awareness, and in many cases, it can do a decent job. Mustafa Suleyman, CEO of Microsoft Ai, just published a piece on “Seemingly Conscious Ai (SCAI)” to further detail the distinction between perception and reality. “Simply put, my central worry is that many people will start to believe in the illusion of AIs as conscious entities so strongly that they’ll soon advocate for AI rights, model welfare and even AI citizenship.” The article lays some important foundations around what Ai can appear to be capable of, the practical limitations, and downside risk of believing it to be conscious or aware.

This becomes even more of an important consideration when seeing the growing number of healthcare applications being developed where empathy is a feature. Companies like Hippocratic Ai and Slingshot Ai are building empathy into their patient facing solutions to improve healthcare outcomes in spaces where there are not enough clinicians or other barriers to care. Questions arise such as how Ai is “making decisions” about its responses. Is the machine basing a response on true empathy and intuition like us humans do, or is it giving its best statistical guess, based on how it was programmed, with limited information (no body language for example).

All of this to say that Ai is distinctly not human. While it can appear to have human characteristics, it’s not actually a co-worker, or a therapist, or a boss, and will thus have challenges in relatability and as Suleyman discusses, some potentially substantial risks to people and society.

Intuition and Instinctual Drive

Beyond reasoning and consciousness, Language Models and Agents still don't fully "get it". A recent paper on the Advances and Challenges of Foundation Agents calls this out with “Agents lack emotional intuition and instinctual drives; their learning depends entirely on the form and fidelity of the reward signal.” The article categorizes the various functions of the brain into levels of development with many of the frontal lobe executive control and cognition functions that manifest intuition being in the “significant room for improvement” category.

Without question Ai has gotten quite good at recognizing emotions, sentiment, and more. While Ai agents can use emotion detection models to recognize patterns like valance vibrating to detect emotion in audio, it still has a number of limitations. Variation in emotional response across cultures and demographics, limited inputs, and limited training data will keep Ai at a disadvantage in its “gut feel” for some time to come. This is because ultimately, Ai is not conscious, as such, it does not have a subconscious, with underlying 'feelings' or goal-oriented instincts...a sixth sense. It lacks emotional instincts and approaches the world entirely through logic, when in some cases there is simply no logic, but pure emotions that drives human behavior.

Here again, researchers are continuing to develop models and entirely new architectures that better mimic these human characteristics. The overlap of neuroscience and Ai is a fascinating space to consider as we try to understand how the different parts of the brain contribute to true consciousness. Important to consider on this journey is that much of conscious awareness has nothing to do with the “language” that our “language models” are built on. It is a much more complex network of different brain functions, inputs, and processing. Turns out humans are quite good at some areas of performance while Ai dominates others.

Human vs. AI Performance

To explore this comparison further, let’s consider data and processing. Humans can process environmental data through multiple senses (vision, hearing, touch, etc.) in parallel and at extremely high rates. Research shows that the human nervous system gathers on the order of 1 billion bits per second from sensory inputs – a bandwidth comparable to a high-speed internet connection. This massive sensory throughput gives people an instantaneous, richly detailed picture of their surroundings that current machines cannot directly replicate. No existing AI has a truly comparable multi-sensory processing ability; our biological senses evolved to work in unison, whereas in Ai, different sensors (e.g. cameras or microphones) operate separately. In other words, humans integrate sight, sound, touch, smell, and taste effortlessly in real time, a feat beyond the scope of today’s AI agents which lack those modalities or the comprehension to fuse them like we do. Even though the human brain intakes an enormous stream of data, it consciously processes only a tiny fraction – roughly 10 bits/s – for decision-making. This “bottleneck” means our conscious thought is comparatively slow, but the raw sensory input bandwidth remains astounding.

On the other hand, when it comes to sifting through huge datasets or text corpora at blistering speed, AI far outpaces any human. Modern AI agents and large language models (LLMs) are trained on hundreds of billions to trillions of words, leveraging vast information sources that no person could absorb in a lifetime. For example, the latest generation of AI systems can consume over eight trillion words in a single month of training – roughly 1,000 times more text than a human would read in their entire life. Once trained on this colossal knowledge base, an Ai can retrieve facts or analyze patterns within milliseconds, scanning through millions of data points almost instantaneously. In practical terms, an Ai can “read” and remember entire libraries or databases without breaking a sweat, whereas a human would require many years (if not centuries) to even come close. Studies note that GPT-3, for instance, was pre-trained by processing about 3 \times 10^11 tokens of text (hundreds of billions of words), vastly more than a person could ever learn from. No human researcher can match the combination of speed and scale at which algorithms today digest information – whether it’s scanning thousands of documents for a pattern or crunching a gigabyte of data in seconds. The machine’s advantage is particularly evident in tasks like big-data analysis, web search, and high-speed computation, where AI systems operate on a level of throughput that leaves human capabilities in the dust.

It’s clear that humans and Ai excel at different kinds of information processing, making it “not much of a competition on either side.” In terms of sensory input and real-world perception, humans maintain a dramatic lead – our ability to assimilate rich, multi-modal information from the environment in real time is still unrivaled. A person can effortlessly recognize faces in a crowd, catch a whiff of smoke and locate its source, or feel the slightest change in terrain underfoot – all within moments. Ai agents, by contrast, do not yet truly experience the world through such integrated sensory streams, limiting their throughput in this domain. But when it comes to speed of processing stored information and scale of knowledge, Ai is the clear winner. LLMs and similar Ai systems recall and manipulate data at orders-of-magnitude faster speeds than human thought, drawing on far more information than any one person will ever know. A well-trained Ai can answer a factual question in seconds by effectively “consulting” billions of documents it has seen, whereas a human might spend days researching the same question.

Ultimately, this comparison highlights how fundamentally different human and Ai strengths are – and why direct comparisons must consider the context of throughput (speed and modality) to be meaningful. The amount of data that humans have available at run time is staggering, while our ability to access knowledge and process is limited while Ai lacks multi-modal real-time processing and an architecture for intuition and instincts. New Ai designs are likely needed to better mimic those human characteristics, leaving us humans the operators of Ai, at least for a bit.

User Experience

With humans as the operators, let’s talk adoption.

As LLMs have made what we call Ai today more accessible, Ai still has a UX problem. Many if not most, still struggle with the “blank page syndrome” and ponder “What do I ask this thing to do?” While the LLMs continue to get better with suggestions and conversational attributes, the rules-driven, dropdown software space has beaten a lot of creativity out of the workforce. Salesforce, ecommerce sites, and business platforms have removed much of the options and narrowed decisions into a funnel filled with checkboxes, dropdowns and next buttons. So, while GenAi has made Ai much more accessible, it still needs to adapt how we work today (or we have to adapt to how Ai works - likely both). It will be a messy back and forth on how it will change the nature of work. I would not be surprised to see dropdowns and hierarchical decision trees to fill in the logical and creativity gaps - that much of today’s software is built on - being built into LLMs, agents, and more. Even as the UX evolves, it could evolve into something that humans just don’t want to use because productivity does not match the passion behind craft.

Craft vs. Productivity

The value proposition is clear. Ai is actively helping people do more with less and adding economic value to companies. There is however a cautionary tale and unintended consequences that could emerge. What if we automate the parts of people’s jobs they enjoy and they grow increasingly dissatisfied with their job? Recent Stanford research studied and found exactly that. This post exposes the challenges in designing user friendly systems that people will use and not revolt against. Is it more enjoyable to do the work of learning to paint and being proud of something unique and the results of much time and effort or is generating a hundred images on midjourney a rewarding craft. As often happens, we have needs in conflict. We need to ship something (productivity) while the craft (actual skill) suffers in face of tools that increasingly produce better results than the majority of humans. This gap will likely cause growing pains with Ai adoption and is important for both workers and companies to be mindful of.

Synthetic vs. Human

Then of course there is the debate of whether us humans will ‘accept’ synthetic products. The 80’s proved that we will largely accept synthesizers but for a few hardcore purists that want live instruments. We haven’t burned down any companies that have automated chatbots the last 20 years. But as always, it depends. Getting critical medical instruction from an Ai agent is very different than one asking if you took your medication. Watching a clumsy robot perform Swan Lake is not remotely the same as the grace of a human studying for 20 years. While society will flex with many things and accept some degree of agents and synthetic products, it also has the potential shut down companies who don’t appreciate human nature.

Expanding Consciousness

As it turns out, the advent of generative Ai is driving us to further explore ourselves through this comparison to Artificial Intelligence, helping us better understand ourselves (as individuals, organizations, and broader society) in the process. Ai and agents are challenging our systems, job descriptions, identity, and much more. It has many concerned about what their children will do to earn a living. It has sparked an Ai arms race that could have big unintended consequences if not properly managed. While it is making major changes and setting off concerns it is also reflecting the importance of the human condition, relationships, and our value in the universe. As stated earlier, Ai will be massively transformational on humanity, but what that means is still unclear and ours to interpret and create a plan of action as we move towards Ai powered co-intelligence. The important thing will be to expand our conscious understanding of the impacts of Agentic Ai and take action to adapt to a rapidly changing world.

Impact/Action

So how does all of this impact you, your job, and broader humanity?

What this means for workers:

Workers will increasingly be asked how they are using Ai to improve their skills and their output. They will also find themselves balancing efficiency with meaning. Ai can automate routine tasks, but if those tasks are also the most enjoyable or identity-forming, dissatisfaction may rise. The challenge for workers will be to cultivate skills that complement—rather than compete with—Ai, such as creativity, intuition, and emotional intelligence. Workers who see Ai as a collaborator rather than a competitor will be better positioned to thrive in hybrid human–machine environments. Ultimately, much like the calculator, the computer, and the internet, investing in learning how to use Ai and adopting it into practice is a wise plan.

What this means for organizations:

For organizations, the decision isn’t whether to use Ai agents but how to integrate them responsibly. The winners will be organizations that create guardrails for trust, preserve elements of human craft, and design systems that support—not replace—human strengths. Companies must weigh productivity gains against the hidden costs of worker disengagement or customer pushback to “synthetic” interactions. Strategic adoption means using Ai for scale while still embedding the human touch where it matters most—especially in brand, customer relationships, and safety-critical operations. Beware the hype, overapplication of technology, the innovation pendulum, and remember that humans (talent, relationships, instincts) are critical for the long-term.

What this means for society:

At the societal level, the perception vs. reality gap in Ai reasoning, empathy, and intuition raises profound questions about trust, regulation, and values. Just as we set standards for safety in cars or medicine, society will need frameworks for when Ai can make decisions independently, and when humans must remain in the loop. The broader cultural challenge will be deciding how much “synthetic humanity” we are willing to accept—and ensuring we don’t lose sight of the very human qualities of empathy, creativity, and intuition that define us. Paramount is the need to build a future where Ai augments humanity rather than eroding it.

Up Next…

The next post will explore some of the technical considerations when approaching Agentic Ai including infrastructure, data, and evaluation of agentic platforms.

Follow for more.

As AI continues its shift from static models to dynamic agents, organizations face a growing ecosystem of platforms promising to replicate human behavior, reason through complexity, and act autonomously across digital environments. In Part 3 of our series on Agentic AI (see Pt1 and Pt2), we explore the platforms and protocols enabling this transformation—from cloud giants like Azure and AWS to experimental systems like LOKA and CrewAI. This post breaks down the technical foundations, platform-specific implementations, and evolving standards shaping the future of intelligent agents.

Background - Models, Apps, and Infrastructure

Let's start out with some background of what agents are made of. As the previously cited Hitachi Ventures article illustrates, agents are comprised of apps, models, and infrastructure. Similarly, Sequoia Capital's narrative on the evolution towards reasoning models discusses the evolution of SaaS models to incorporate digital labor directly into platforms.

In part 2, we also discussed the 3 core functional areas of agents - perception, cognition, and action. With some liberty, we can line these areas up accordingly.

• Perception - infrastructure

• Cognition - Models

• Action - Apps (tools)

With that background, let Let's dig into the technical composition of AI Agents.

Technical Composition

While the lines can blur, the above list gives us a sense of the relationship between the inputs and outputs of Agents. Breaking this down a little further we can separate out infrastructure into hardware and data, models into their respective types, and apps into tools, communication, and so on. Let's start with the brains behind the operation: models.

Models: The Brains Behind Agents

While much of the infrastructure and apps has been around for years, the missing piece to the Agentic puzzle was cognition. Increasingly, Large Language Models are evolving towards reasoning models. Reasoning models are different from LLMs in that they "slow down and think". They integrate the capabilities of LLMs with chain-of-thought reasoning, multi-step planning, memory, and more elements we will go through below. This comes with a shift in resources from training to inference or test time compute. Reasoning models generate more tokens as discussed in this article about how benchmarking is getting more expensive.

The release of DeepSeek R1 can be characterized as part of this broader shift in emphasis—from massive pretraining compute toward leveraging more test-time compute and retrieval-augmented mechanisms to improve reasoning and performance. This and other architectural breakthroughs are helping evolve agents by demonstrating an enhanced ability to maintain coherence across multi-step tasks and execute complex reasoning chains without losing context. Benchmarks are continuously being set higher and passed by the evolving reasoning models.

Deepseek and other modes are getting better at decomposing problems and generating intermediate steps before producing solutions, effectively showing emergent planning capabilities not explicitly built into its architecture. An example of this is OpenAi research asking questions to collect more context before executing it's work. This missing link acknowledges that we don't always provide all of the context necessary in a request - something that the reasoning model can interpret much better than a pre-trained LLM. This architectural efficiency and emergent reasoning challenged the industry's scaling-focused paradigm and inspired a new wave of models prioritizing architectural refinement over raw parameter count.

Data: The Fuel

As a foundational element of models, data plays a key role in this evolution and how data is used is evolving rapidly. How, when, and where data is used fundamentally shapes agent capabilities and limitations. Agents are increasingly able to incorporate various forms of runtime data. This increasingly includes Multimodal - text, images, raw data, etc... - such as environmental data from sensors or system monitors and external information retrieved from APIs or databases. This expansion in access to data is critical in expanding the capabilities of agents as much of our world today is data-driven and the basis for decision making and execution. This expanding data and reasoning capabilities is driving the need for increasingly more sophisticated means of memory.

Memory: Holding Context

Memory is a critical component of how us humans work. Our ability to leverage long-term memory of workflows, templates, client knowledge is indeed important, but equally important is the ability to use short-term contextual memory to complete a task. For example, booking a trip to Thailand requires not just a general understanding of how to book a trip, but all of the parameters important to doing so correctly for the situation at hand. These factors include dates of travel, number of passengers, ages, preferences, budget, flexibility, desired and undesired airlines, etc... While a small set of these can go into a prompt it is much more desirable for the system to remember these as it tries various attempts to book travel and moreso for increasingly complex tasks.

Likewise, memory systems enable AI agents to maintain context and coherence across interactions, significantly enhancing their ability to complete complex tasks effectively. AI agents are continuously being upgraded with sophisticated memory systems to deliver consistent, contextual responses. These memory architectures can be divided into several key components that work together to create more human-like interactions.

Short-term memory represents a recent development that has significantly enhanced how Large Language Models and agents operate. This capability includes conversation history that allows agents to maintain context throughout an interaction. Without this functionality, every response would be disconnected from previous exchanges, creating fragmented and frustrating user experiences. The implementation of short-term memory enables agents to reference earlier parts of a conversation and maintain coherence, making interactions feel more natural and continuous. Major cloud providers have recognized the importance of this feature, with both Amazon and Azure offering robust caching mechanisms to support conversational context maintenance.

Long-term memory systems store persistent information for future reference, allowing agents to build upon past interactions across extended periods. Vector databases store embeddings of previous interactions, enabling semantic searching across historical data when needed. Knowledge graphs represent relationships between entities, creating structured maps of information that agents can navigate to understand connections between concepts. Episodic memory systems organize experiences chronologically, creating a timeline of interactions that provides temporal context. Semantic memory systems store conceptual knowledge in accessible formats, giving agents a foundation of understanding to draw upon.

The deeper technical aspects of memory come down to sophisticated memory management - determining what information gets recalled and what doesn't. This process involves complex algorithms that prioritize relevant information based on context, ensuring that responses incorporate the most pertinent details. Memory integration represents another critical challenge when dealing with multiple potential sources of information. Modern approaches include retrieval-augmented generation workflows that enhance responses with relevant retrieved information from various sources. Memory-based reasoning systems combine multiple information sources to produce more informed and contextual responses by synthesizing diverse knowledge. This memory management is highly complex and is both impressive and disappointing in how it prioritizes information. A rapidly expanding work-in-process, this is possibly one of the more complex aspects of how agents work as it significantly impacts decision making.

Tools: Taking Action

In order to take action, agents need tools. These tools (or Apps) extend an agent's capabilities beyond its core model, enabling interaction with external systems and specialized functionality. Tools include information retrieval systems like search engines and knowledge bases, code execution environments for computational tasks, APIs for accessing external services, and specialized processors for domain-specific tasks such as mathematical problem-solving.

Much like a human having access to many tools, agents are being supplied with more and more of these same tools. Tools and apps vary by platform. In OpenAi, one can call a “GPT”. In Azure an agent can call an azure function or app service. In the AWS ecosystem agents can access Lambda functions that can take action or connect to tools to perform actions. Many of these tools are existing platforms that can be connected to perform bigger multi-step, multi-platform actions.

As more and more complex tasks can be accomplished, this requires coordinating multiple tools in a process known as orchestration. This might involve using tools in sequence for multi-step processes, executing them in parallel for efficiency, implementing conditional branching based on intermediate results, or creating feedback loops to refine tool usage based on outcomes. This orchestration requires communication.

Communication

Today's AI agents don't just answer questions—they manage workflows, coordinate with other systems, and maintain persistent understanding across numerous interactions. This shift requires robust protocols for handling the nuanced exchange of information between models, applications, and users. Think traditional TCP/IP or SMTP but for agents.

Model Context Protocol

Anthropic's Model Context Protocol (MCP) represents a significant advancement in addressing these challenges. At its core, MCP serves as an interface between models and applications, maintaining crucial contextual information throughout interactions.

In simple terms, MCP functions like a smart notebook that an AI carries to track important elements in conversations and tasks. Rather than starting from scratch with each interaction, the AI can reference this "notebook" to understand what's been discussed, what actions have been taken, and what still needs attention.

From a technical perspective, MCP implements a structured approach to managing contextual awareness in AI systems. It organizes information into retrievable memory units, employing techniques such as:

1. Long-short term memory management to balance immediate needs with historical context

2. Embeddings for semantic understanding of content

3. Hierarchical context retention that prioritizes the most relevant information

This architecture allows AI agents to maintain continuity, adapt to evolving user needs, and improve decision-making over time. Important to note is OpenAi following suit in adopting MCP.

Some examples include using Claude with

• Blender to create 3D content as in this post.

• Unity as in this post

• Figma workflows

MCP servers can be found online in a growing number

• https://mcp-get.com/

• https://smithery.ai/

• https://github.com/modelcontextprotocol/servers

• https://cursor.directory/mcp

• https://opentools.com/

Agent 2 Agent Protocol

Google also just launched a communication protocol called Agent to Agent, or A2A.

This article does a good job of comparing A2A and MCP by stating "A2A facilitates agent-to-agent communication, while MCP enables agent-to-tool interactions." Both of these protocols will be needed as agents need access to tools and each other (much like us humans do).

LOKA

The LOKA Protocol (Layered Orchestration for Knowledgeful Agents) is a decentralized framework designed to foster trustworthy and ethical AI agent ecosystems. LOKA is an Open Source layered framework that introduces a Universal Agent Identity Layer (UAIL) for decentralized, verifiable identity; intent-centric communication protocols for semantic coordination across diverse agents; and a Decentralized Ethical Consensus Protocol (DECP) that could enable agents to make context-aware decisions grounded in shared ethical baselines. Anchored in emerging standards such as Decentralized Identifiers (DIDs), Verifiable Credentials (VCs), and post-quantum cryptography, LOKA proposes a scalable, future-resilient blueprint for multi-agent AI governance.

APIs

While MCP represents one approach to Ai communication, the broader ecosystem includes various API-based solutions that enable different forms of interaction between LLMs and applications. These APIs typically provide:

• Structured input/output formats (often JSON-based) for predictable data exchange

• Authentication and rate-limiting mechanisms to manage access

• Endpoints for specific functions like text generation, classification, or tool usage

• Webhooks for asynchronous communication patterns

Together, these communication protocols and APIs form the nervous system of modern AI agents, allowing them to interact with users, access external tools, and coordinate with other systems in increasingly sophisticated ways.

Enterprise Platforms

Now, let’s take a look at some platforms starting with the hyperscalers (Google, Amazon, Microsoft). These ecosystems are well positioned for Agentic Ai as they already have access to much of the worlds infrastructure and apps.

Azure

Microsoft, making a $10b investment into Open Ai early on has worked diligently to incorporate ChatGPT models into the Azure ecosystem. With access to permissions to data and apps similar to a human user, the model can interact with data where many already have it in - in sharepoint, SQL databases, Powerpoint decks, teams, custom applications, etc...

Azure Agentic Ai Architecture show how agents fit into the overall azure infrastructure including security and permissions. In Azure agents are given permissions and access to resources similar to how they are given to humans to maintain security and leverage their existing infrastructure. Actions can be taken throughout the system using communications protocols outlined above.

This ecosystem can be developed in Azure through Azure Ai Foundry, a more developer-centric environment with full control, or in Copilot Studio, a low/no code environment to connect models to data and apps.

AWS

Similarly, AWS has now invested $8b into Anthropic and has incorporated a many of Anthropics Claude models into the AWS Bedrock platform. Here, one can build standalone Agents as well as multi-agent systems in a low-code/no-code interface. Here one can select the models, add instructions, connect RAG data sources and customize an agent.

One can quickly set up an agent with some base instructions and settings in a no-code environment and be off and running with a proof of concept in minutes. Similar to the above Azure diagram, AWS has published articles that lay out the agent architecture and how agents integrate into the broader AWS ecosystem. This again represents the treatment of agents in a similar regard to humans by providing access to tools and data.

Google

Google recently launched Agentspace, an enterprise Ai platform that integrates AI agents, Gemini, and enterprise search. Similar to Microsoft and AWS, Agentspace powers agents within the Google ecosystem to accomplish complex tasks such as planning, research, content generation, and automation with a single prompt. The platform also connects to third party apps including Jira, Microsoft SharePoint, and ServiceNow, that provide access to both structured and unstructured data. Features like NotebookLM Enterprise for data synthesis and audio summaries, and Agentspace Enterprise Plus for creating custom AI agents allow for low/no-code solutions as well as highly custom solutions. As mentioned above, Google has also launched Agent2Agent (A2A) protocol to manage multi-agent systems.

Ai Labs

The Ai labs themselves are also working on expanding their Agentic capabilities.

OpenAi

What began as a conversational interface has rapidly evolved into a dynamic platform where autonomous, goal-directed actions could be executed through user-defined “GPTs”—custom agents capable of integrating with external tools, APIs, and data sources. OpenAI’s emerging Deep Research platform exemplifies this trajectory: it enables users to generate comprehensive research plans, perform live internet searches, synthesize structured outputs, and manage complex writing workflows—all within a single agentic thread. With the launch of GPT-4o, OpenAI has further pushed boundaries by fusing real-time image generation, voice interaction, and multimodal reasoning into a single, unified system. Meanwhile, the release of o3 and o4-mini models reflects OpenAI’s strategic layering of lighter-weight agents that can operate at different performance and latency thresholds—signaling a future in which swarms of specialized, cooperative agents can handle diverse tasks across domains.

Anthropic

Anthropic is also making strides of its own in Agentic capabilities. Their Claude models now incorporate key agentic methods including Computer Use, which allows the AI to execute code, browse websites, and interact with computational environments. The popular model company is leaning hard into agents as part of a broader strategy to power the future of work. As mentioned, the recently released Model Context Protocol (MCP) enables Claude to connect with existing platforms while maintaining contextual awareness throughout multi-step processes.

Agent Platforms

Beyond the hyperscalers and Ai labs a number of companies offer platforms to build and deploy agents.

Crew.ai is an advanced framework designed to orchestrate multiple AI agents working together toward a shared objective. Unlike single-agent models that perform isolated tasks, CrewAI enables the creation of intelligent agent “teams” where each agent has a specialized role—such as a researcher, writer, or planner—working in tandem to accomplish complex workflows. These agents can communicate, delegate tasks, and refine outputs across different LLMs and tools.

Langchain is an open-source framework that empowers developers to create AI-powered applications that dynamically chain together multiple components, such as large language models, memory, APIs, and databases. It provides a structured way for AI to reason and make decisions based on external data rather than just static prompt-based interactions.

Fetch.AI focuses on creating autonomous economic agents (AEAs) that operate in decentralized environments. These agents are designed to interact with digital and real-world markets, autonomously making decisions and negotiating transactions.

Emergence is an AI framework designed to develop self-evolving agents that can adapt to changing environments and optimize their strategies over time. Emergence allows agents to iteratively improve through reinforcement learning, simulation, and evolutionary algorithms. This makes it particularly powerful in fields requiring continuous adaptation, such as robotics, financial modeling, and AI research itself.

Ema is an emerging agentic AI platform designed to create autonomous, adaptive, and highly interactive AI agents capable of executing complex tasks with minimal human oversight. Ema connects different models and apps and sets up workflows in a low/no-code interface.

Genspark uses a multi-agent framework where different specialized AI agents handle different parts of a search query — such as summarizing, verifying facts, checking biases, and sourcing multimedia.

Agent Index

The number of agentic platforms continues to expand with the above representing only a few. As the promise of agents and complexity grows, better means of tracking platforms and capabilities is needed. The AI Agent Index, developed by MIT-affiliated researchers, is the first public database cataloging deployed agentic AI systems—those capable of planning and executing complex tasks with minimal human oversight. It documents each system’s technical components, intended applications, and safety measures, providing a structured framework to enhance transparency and inform stakeholders about the capabilities and risks associated with these autonomous agents.

Vertical Platforms

A topic for a full separate post is a list of the vertical platforms in specific industries. These companies represent a broader movement across verticals—from law and customer service to compliance and healthcare—toward AI systems that behave more like collaborators than tools. Each embodies unique agentic capabilities, but they all point to a common future: one where AI doesn’t just assist, but acts with purpose. A small sample of these companies includes the likes of Harvey, Sierra, Norm.ai, Hippocratic Ai.

Harvey an agentic AI platform for legal professionals, augments the work of lawyers by performing complex tasks like drafting legal documents, summarizing case law, and conducting legal research. Sierra is a customer experience AI platform designed to serve as an autonomous customer service agent, capable of managing entire customer workflows such as troubleshooting, returns, and onboarding. Norm.ai is focused on compliance and regulatory workflows. It functions as an AI agent that interprets legal and compliance documents, maps them against company policies, and flags areas of risk. Hippocratic AI operates in the healthcare space with a mission to provide safe, empathetic AI agents for non-diagnostic clinical tasks such as patient follow-up, medication reminders, or pre-op preparation.

Shifting Tides

“You cannot step twice into the same river, for other waters are continually flowing in.”

— Heraclitus

With the rate of change in Ai constantly increasing, it is difficult to keep one's bearings in the constantly shifting tides. While this post lists out just a few of the many platforms it is likely to look different a year from now. A good example of this is the difference in tone between Microsoft investing $10b into OpenAi in 2023 and the recent quote from Microsoft CEO Satya Nadella that states "models are becoming commoditized".

As with many things, the only constant in Ai is change.

Where to Start

So what does this mean for companies trying to select the right vendor, model, or strategy? With many models, evolving platforms, pressure from leadership and FOMO, implementation decisions are complicated. Add to this the hype from vendors are “experts” alike that talk about how easy it is to build and deploy Ai. A word of caution - there is no easy button to integrate Ai into an organization - as this post discusses.

There are many considerations when building a strategy to approach and implement Agents in an organization. Below are a few to get started.

• Navigate the Hype
  Read up and find partners to advise on what’s real and what’s not.

• Define High-Impact Use Cases

  Start by identifying tasks that are repetitive, decision-heavy, or span multiple systems—these are ripe for agentic automation.

• Map Existing Data and Tool Ecosystem

  Inventory the APIs, databases, and platforms your agents will need to interact with to ensure compatibility and avoid integration roadblocks.

• Start with Low-Risk Pilots

  Build agents to assist internal teams (e.g., research, operations, IT support) before deploying customer-facing or compliance-critical use cases.

• Choose the Right Long-Term Platform

  Evaluate agentic platforms for fit based on control, scalability, and integration needs.

• Establish Human Oversight Protocols

  Design workflows where agents are supervised, audited, or prompted to pause for review when confidence is low or consequences are high.

• Design for Observability and Feedback Loops

  Implement clear logs, dashboards, and user feedback channels to continuously monitor agent behavior and refine performance.

• Do the Math
  Consider all costs - implementation, transactions, unanticipated cost model changes by providers (consider that agents may cost 3-5x in the coming years).

• Create a Governance and Ethics Framework

  Define principles around autonomy, transparency, and responsible use to ensure trust and alignment with corporate values.

• Upskill Teams on Prompting and Orchestration

  Train staff on prompt engineering, agent chaining, and oversight—skills that will be foundational to working with AI collaborators.

Onward

In the next Installment, we explore the challenges and ethics involved with Agentic Ai and take a deeper look into the hype and realities of this rapidly emerging technology.

Today, Ai helps us in two key areas (among others) First, in reducing the time and effort to get work done, and second in significantly expanding creativity and output. While AI is currently deployed at a task level, its macroeconomic implications are profound. A brief exploration of industry trends reveals that one of AI’s most disruptive impacts will be a fundamental shift in business models. These shifts will likely unfold in waves, and we now stand at the transition between one wave—the rise of automation—and another: the emergence of sophisticated digital labor.

Digital Labor

The emergence of AI agents is ushering in a new paradigm of "digital labor," where automated systems can autonomously handle complex tasks traditionally performed by human workers. A recent Hitachi Ventures article discusses the concept of "digital labor" and potential implications on how we work. Unlike traditional software tools that simply augment human capabilities, these AI agents function more like virtual employees – they can understand context, interpret intentions, and adapt to new situations with minimal human oversight. This shift in capabilities represents more than just an advancement in automation; it signals a fundamental change in how we think about work and workforce composition. Companies are beginning to treat these AI systems not as mere tools to be licensed, but as digital workers to be "hired" at hourly or outcome-based rates. The distinction is crucial: while traditional software requires human operators to achieve results, these AI agents can independently drive end-to-end processes, from understanding tasks to executing solutions. This transformation is particularly evident in high-skill domains like healthcare, legal work, and software development, where AI agents can perform increasingly complex tasks that previously required significant human expertise and judgment.

Hourly Agents

This shift in delivery from human powered SaaS platforms to Ai powered digital labor is opening up a significant opportunity for SaaS companies, namely, hourly billable agents. The article by Sequoia Capital referenced in part 1 of this series further lays out the case how reasoning agents will drive a shift to the hourly billable agent. The new form of "slow thinking" powers models to reason with more real-time data and act with more sound decisions. It also allows companies to expand their business model from one of monthly SaaS seat licenses to a "software + labor" model to create hourly billable agents.

To date, sound decisions were primarily the domain of humans with Ai occasionally producing easily attacked errant judgement. As models slow down, find more information and improve their judgement, comparison with existing human decision making will increasingly come closer to some degree of parity (or appear to). As such, hiring decisions of digital labor vs. human labor becomes an increasingly commoditized decision process. Agents could be an attractive option for companies as they seek to build more, reduce cost or likely both in the coming years.

This shift in perspective, effectively opens up the $8.5 Trillion US labor market and make clear why VC's are excited about the opportunity to tap into it. It's a welcome new source of revenue to the big tech firms which may find themselves hitting a revenue plateau.

Use Cases

Healthcare

Healthcare for example is an industry facing massive challenges. Enflamed by Covid, clinician burnout and dropping rates of people going to medical school means we will not have enough physicians to take care of the population. Left unchanged this will continue to get worse and worse as the years pass. Companies like Hippocratic.ai are hoping to fill the gaps in various aspects of care from patient check-ins to adherence and more. The company is not just selling a SaaS solution, it is selling Healthcare agents at an hourly rate to help meet the demand of these clinicians.

Software Development

Software engineering has companies like Devin.ai, cursor, replit, and others working to automate the software development process. Today developers can use these platforms to develop software in much less time. These firms and more are no doubt working on expanding these human run platforms to agent run platforms that can be billed at a fraction of the cost of a traditional software developer.

Manufacturing

Manufacturing has long embraced automation, but Agentic AI has the potential to enable systems to make real-time decisions, enhancing human-machine collaboration. Companies like Juna.ai are at the forefront, developing AI-driven solutions that optimize production processes and improve operational efficiency. The Abu Dhabi National Oil Company (ADNOC) has also embraced Agentic AI, deploying autonomous systems to monitor and manage complex oil extraction operations, thereby enhancing safety and productivity. In a recent report, the World Economic Forum highlighted the transformative potential of AI agents in manufacturing, emphasizing their role in creating near-autonomous systems and seamless human-machine collaboration. Other notable companies contributing to this evolution include Covariant, which specializes in AI-powered robotic systems for warehouses, and Figure AI, developing humanoid robots designed to perform manual labor in industrial settings. These advancements underscore a significant shift towards more intelligent, responsive, and efficient manufacturing environments.

Legal

In legal services, Harvey, Norm.ai, Westlaw Edge, and others are deploying platforms that take the highly structured field of legal and regulatory and integrate generative and reasoning models to conduct legal research, draft contracts, summarize case law, and even generate arguments in a fraction of the time it would take a junior associate. These companies are receiving great interest from VC firms as can be seen in the recent $300m Series D raise by Harvey.

The Billable Hour

Unlike previous advancements in automation, which typically optimized tasks within existing workflows, AI agents have the potential to completely reframe how labor is valued, structured, and billed. On the one hand while digital labor companies fight to justify their billable hour, the other side, the human billable hour, will be fighting to justify itself.

With Ai as equalizer, the traditional equation of effort to compensation comes into question. From the seller perspective, if one can do four times the work in an hour one could speculate that the hourly rate would have to increase accordingly. From the buyer perspective, in a world where Ai reduces the amount of time to do a task, why pay the same amount it used to cost? The former answer could expand productivity while maintaining (or increasing) wages. The latter dynamic could lead to a situation where firms use AI-driven efficiencies to undercut competitors, delivering the same work at a fraction of the cost. Initially, this might seem like a positive shift—reducing expenses for clients and increasing accessibility to high-quality services, but history says otherwise.

The Legal Field: A Prime Example of AI-Driven Disruption

Few industries are as deeply entrenched in the billable hour model as the legal profession. Traditionally, firms have operated on a time-based structure, where associates and partners log hours for tasks ranging from research and document review to contract drafting and litigation support. However, the advent of AI-powered legal tools is challenging this model in profound ways.

While the democratization of legal services creates cost efficiencies for clients, it also means that firms can no longer rely on these once-premium services as primary revenue drivers. The legal professionals who thrive will be those who focus on high-value advisory roles, complex litigation, and strategic negotiation, where human expertise remains indispensable. Alternatively, instead of competing purely on price, some lawyers will differentiate themselves as AI-enhanced legal strategists—leveraging AI not just for efficiency, but to deliver deeper insights, faster decision-making, and better case outcomes. Boutique firms and solo practitioners may benefit the most from AI, as they can now handle more cases with fewer resources, competing more effectively with larger firms. In-house legal teams may also increasingly rely on AI-driven tools to reduce reliance on external counsel, forcing law firms to rethink their pricing structures.

The Ai-gency

The advertising industry has long been structured around a mix of billable hours, retainers, and project-based pricing. Agencies charge for strategic planning, creative development, media buying, and campaign execution. But as AI continues to disrupt the creative and marketing landscape, the traditional business models of advertising agencies are exposed to change.

Today, AI-generated content & design tools like Midjourney, DALL·E, and Runway can produce ad creatives, social media visuals, and even entire video ads in minutes—work that previously took days or weeks. Automated copywriting with ChatGPT, Jasper, and Copy.ai can generate headlines, ad copy, and blog posts in seconds, reducing the need for copywriters to spend hours crafting content. AI-driven media buying & optimization: Google Performance Max, Meta’s AI ad platforms, and programmatic advertising engines can automatically optimize campaigns in real-time, making traditional media planning roles less critical.

Beyond single purpose tech we are also beginning to see full Agentic Pipelines being developed. Tools like LTX Studio are chaining multiple functions together into a single video production pipeline. Video production consists of idea generation, casting, scripting, still imagery, video concepts, visual effects, editing, etc... LTX studio packages these all into a single pipeline in which creators can manipulate output throughout the pipeline, condensing what may have taken a team of 10 several weeks can now be done in 1/10th of the resources and time.

Redefining Value

Without question, Ai is forcing the re-evaluation of value. As in all traditional market dynamics, the market will set the price and determine value. Firms will inevitably run into project bids and billing discussions about the value of the billable hour or project delivery in an Ai powered economy. They will have several strategic options when it comes to pricing and service models. One approach is to maintain current hourly rates while leveraging AI to complete more work in less time, increasing margins but risking client pushback as they expect costs to reflect reduced effort. Another option is to lower rates to stay competitive, using AI to offset the reduced revenue per project. The risk here is of course a race to the bottom, where firms continuously undercut each other, ultimately eroding profit margins, devaluing services, and making high-quality work unsustainable.

Aside from these reactive strategies, companies have options to rethink their business models to something old, new, or different.

1. Outcome-Based Pricing – Instead of charging for time or deliverables, companies could shift to performance-based models, where compensation is tied to measurable success (e.g., increased sales, customer acquisition, or brand engagement).

2. AI-Enhanced Strategy Consulting – Instead of selling execution, companies could position themselves as AI-powered strategic partners, charging for deep insights, positioning, and Ai-powered strategic direction.

3. Subscription-Based Creative Services – Some companies could move to a SaaS-like model, where clients pay a fixed monthly fee for AI-powered services that leverage digital and human labor.

4. Licensing & IP Monetization – companies could develop on-demand proprietary AI models and products that are licensed to brands rather than selling services.

Regardless of which business models are chosen, in an AI-driven world, companies that cling to traditional pricing models and labor-intensive processes may struggle. Those that embrace AI as a force multiplier—enhancing their capabilities rather than replacing them—will thrive.

Also important to note is the reality that established players and relationships will not go gently into that good night. It is unlikely that Accenture, Deloitte, and many more will sit by as Ai destroys their business models. "Follow the money" it is said and this wisdom no doubt will hold true for Ai. While small, nimble startups will be able to innovate faster and take market share from bigger players, the established businesses will continue to hold value in the age of Ai, particularly the ones with data and resources to train big models of their own.

Tangible vs. Intangible

The distinction between tangible output vs. intangible human factors (relationships, communication, politics, responsibility, etc…) are increasingly coming into focus as Ai evolves to address the work product side, but lacks the human aspects of how the world operates. As much as Ai can produce, there is a lot it still can't do. Finesse, intuition, leadership, trust, emotional intelligence, politics, egos are all part of the human condition that Ai is not intentionally being trained on (yet). For example. An Ai agent can pull together a report on the latest trends in Ai and how it might apply to a business, but it cannot understand that the CEO has no interest in hearing anything about it, especially not from Dave who insulted him at the company party. An Ai agent is also not going to go to the bar with you to discuss why you lost the project bid. These intangibles, sometimes difficult to clearly see and understand in other humans, are actually quite important to how the world operates.

Safe...for now

With this in mind, humans still hold the upper hand in many critical aspects of work. While AI can generate compelling content, it lacks true emotional intelligence and the lived experiences that allow humans to craft stories that resonate deeply with audiences. It lacks cultural nuance, tapping into shared experiences, and evoking emotions in a way that builds authentic connections.

Ethical and legal considerations also remain a key area where human oversight is essential. AI lacks the moral reasoning and contextual awareness required to navigate complex ethical dilemmas, avoid bias, and ensure compliance with ever-evolving regulations. While AI can assist with pattern recognition and optimization, true creative differentiation—the ability to push boundaries, challenge conventions, and do so in an ethical framework—remains an inherently human strength.

For these reasons and more, Ai Agents will indeed require human oversight for years to come. One need not look further than the lawsuits and headlines of Ai being trusted too much. The risk of full autonomy in agents is simply too great and the tech still early. For the foreseeable future, humans will still be the ones telling the stories, responding to crisis, managing office politics and reading the room. But they will be doing it more and more with their digital coworkers.

Adoption

If there were a shift in SaaS models toward hourly billable AI agents, it would represent a tectonic shift in traditional work—both in its scale and speed. While adoption across different industries and companies will vary, mass adoption of AI agents—regardless of their effectiveness—will likely be a slow and uneven process. History has shown that transformative technologies rarely achieve overnight ubiquity due to structural resistance, cultural adaptation, and economic realignments. For example, The Personal Computer began to emerge in the 1970s, but It wasn’t until the 1990s that PCs became indispensable across all industries, driven by cost reductions, improved software, and generational shifts in the workforce. The Cloud Computing Revolution started in the early 2000s and is arguably still under way with a lot of companies and industries, especially big older ones. Despite the clear efficiency gains of cloud-based software, many enterprises have hesitated to move away from on-premise servers due to security concerns and legacy systems. Widespread cloud adoption took over a decade, accelerating only after companies like Amazon (AWS), Microsoft (Azure), and Google (GCP) proved the business case at scale. Adoption is also dependent on external pressures such as economic downturns, regulatory landscapes, and industry-specific needs

Such is the way with technology adoption. These same challenges faced by new technology adoption for many years will indeed be faced by Agentic Ai and digital labor. It will likely follow the typical "adoption curve".

Source: https://www.business-to-you.com/crossing-the-chasm-technology-adoption-life-cycle/

Actual adoption will be interesting to watch. Due to the complexity and important nature of some of the increasingly cognitive tasks “results may vary”. A notable factor here too is how fast Ai is developing and making tech that was released and tested 3-6 months ago outdated.

Challenges in Digital Labor

Beyond Agentic Ai simply lacking a number of important human traits, somewhere in the fine print is the complexity and challenges of integrating digital labor. Starting with a traditionally basic one - service level agreements. How will these evolve when Microsoft, AWS, Open Ai and others provide decision making? Today, they all have disclaimers stating “Ai is experimental and subject to hallucinations”. This is unlikely to pass too many governance discussions when it comes to automating workflows with Agents that randomly lie or make terrible judgement calls. Add to this the legal questions about liability, complexities in vendor evaluation and contract negotiations before any deal gets done.

Beyond these usual suspects a glaring challenge for those selling Agentic systems will be the unending number of outstanding questions that will take time for both the tech and the humans to grapple with. While blog posts and commercials can inspire thoughts about the potential for Agentic Ai and enumerate the features, the truth is that human-based systems have many questions to be answered before we will see substantial adoption.

1. How Will Businesses Justify AI-Generated Billable Hours?

   1. If an AI agent can complete a task in seconds, does it still justify hourly billing?

   2. Will companies shift from per-hour pricing to per-outcome or per-task billing models?

2. How Will Companies Structure Human-AI Collaboration?

   1. Will AI agents work independently, or will they require human oversight for approval?

   2. How will businesses handle accountability and liability when AI makes decisions?

3. What Regulatory and Ethical Barriers Will AI Face?

   1. Will governments impose restrictions on AI labor models (e.g., taxation, licensing)?

   2. How will AI agents comply with labor laws and industry regulations?

4. Will AI Disrupt Traditional Employment Models?

   1. Will AI agents replace human contractors, or will they augment human work?

   2. Will companies use AI as a way to reduce workforce costs, leading to new economic inequalities?

5. How Will Market Forces Respond?

   1. Will early adopters gain a competitive advantage, forcing laggards to adopt AI out of necessity?

   2. Will industries resist AI agents the way some unions resisted automation?

Many questions remain for organizations, industries, governments, and co-workers of Ai agents. While these questions evolve, one thing remains clear: as agents evolve and these questions are answered, the future will look quite different.

Whats Next

As the saying goes, we tend to overestimate in the short-run and underestimate in the long-run. As with previous technologies, Agentic Ai will have to navigate the hype cycle that we are only at the beginning of today. Some of today's marketing promises will frustrate explorers as the headlines ride the rollercoaster of hype. But there is a degree of inevitability here. The question is not if this shift will happen—but rather when and how industries will adapt. There will no doubt be a pendulum swinging for years to come. Much like offshoring gave way to reshoring, companies will inevitably have some false starts, especially with a technology in its infancy. But just as computers, the internet, cloud computing, automation eventually became industry norms, AI agents will likely redefine how work is structured, priced, and valued. The key for all of us will be to understand the impact of this transformation as it does so.

In the next post we will dive into some of the tech platforms (Azure, AWS, others) that offer Ai Agents as well as the considerations when shifting work from human-based workflows to machine-based ones.

Stay tuned and subscribe.

What is Agentic Ai

Agentic Ai is storming the headlines and is quickly becoming a staple of Ai in 2025. Microsoft, AWS, Google, Deloitte, Accenture, and many more are racing to describe how their LLM-based agents will be fundamentally changing the future of work. But with all the hype comes confusion of exactly what Agentic Ai is (and what it is not). Much like the marketing engines behind Spatial Computing, Web3, and other early stage terms, definitions of Agentic Ai are "in development". This article is part 1 of 5 in which we will explore Agentic Ai by breaking down definitions, use cases, business implications, challenges, and attempt to separate hype from reality. This article will focus on the what and the how of Agentic Ai.

To start a discussion about what Agentic Ai is, it may be best to ask the question we are trying to build? What do we need? What's missing? What are the benefits of building it? Do we want to build a racecar driver, a waiter, a financial analyst, an executive, a chef, a politician, a CEO, a logistics coordinator, event planner, etc...? Roles are a common way to categorize AI functions by mirroring human professions, but as we’ll discuss, the reality is more complex. We must also consider the industry and other lenses through which we make this definition. While racecar driver implies both role and industry, others get pretty generic. Take a project manager for instance. The PM role has some overarching similarities, however the difference between PM in construction is vastly different than a PM in software, which is substantially different than a PM in biomechanical engineering. Once one peels back the veneer of a headline on Forbes saying "agents are here" to think about what we want/need/build they can begin the process of critical thinking about the immensely complex topic packaged up in neat headlines.

Before diving deeper, let’s establish a working definition of Agentic AI based on existing discussions. Researchers and industry leaders have proposed various definitions. Some examples include: “the degree to which a system can adaptably achieve complex goals in complex environments with limited direct supervision” (Shavit et al. 2023) and “an agent defines any entity that perceives its environment through sensors and acts upon it through actuators” (Russel and Norvig 2021). For simplicity, we can define Agentic AI as incorporating autonomy, contextual reasoning, and adaptability.

Next, let's determine what "type" of agent we are talking about. There are traditionally 7 types of Ai agents (sometimes described as anywhere between 5 and 9), each designed to interact with their environment in unique ways.

1. Simple Reflex Agents operate reactively, using predefined rules to respond to immediate stimuli.

2. Model-Based Reflex Agents add an internal state, enabling them to consider past information when deciding actions.

3. Goal-Oriented Agents focus on achieving specific objectives by planning and selecting steps to reach a desired outcome.

4. Utility-Based Agents optimize decisions by balancing trade-offs to maximize a utility function, such as safety or efficiency.

5. Learning Agents improve their performance over time by adapting to feedback or new data. Cognitive/Deliberative Agents simulate human reasoning, using logical thinking and internal models for complex problem-solving.

6. Hierarchical agents structure decision-making into multiple levels, where higher layers manage abstract goals and strategies while lower layers handle detailed execution, enabling efficient, scalable, and adaptive AI systems.

7. Collaborative (Multi-Agent) Systems involve multiple agents working together, either cooperatively or competitively, to solve distributed problems.

We are now adding some new ones:

8. Autonomous (Agentic) Agents take this further, acting independently with minimal human input to pursue goals, often creating subgoals dynamically.

9. LLM agents are AI systems powered by a large language model that processes and generates human-like text, reasoning through prompts, retrieving knowledge, and executing tasks autonomously or interactively.

Note that these newer agents may actually be a combination of several of the traditional types of agents at once. Suffice it to say, Ai agents have been around for a long time and have continuously been upgrading their skills (much like us humans).

How We Got Here - The 3 Waves

With that background in mind, it can be helpful to help inform the "what" agents are by understanding how we got here. It is indeed an interesting history that goes back to WWII. A recent Forbes article discusses the 3 waves of Ai as the culmination of decades of Ai research that began with traditional Ai focused on prediction, Generative Ai that helps us generate content and ideas and the new wave that incorporates goal oriented behavior, action, decision making and autonomy. The concepts of human level intelligence in machines began in the post-war with big dreams but significant hardware constraints. One could argue we have been trying to make Agentic Ai for 80 years and we just now are beginning to get the tools to really do so.

Even though "strong Ai" was defined early on, it is quite complex and the technology was not available. The first wave of Ai focused on narrow or "weak" Ai that used algorithms focused on very specific predictive tasks. This wave is responsible for for Machine Learning algorithms that today are ubiquitous to the point where they make many decisions about our lives. The decades that followed allowed researchers the technological and Ai advancements to bring humanity into the next wave - that of Generative Ai. Google pioneered much of this early work and OpenAi stunned the world with ChatGPT in 2022 with the release of GPT 3.5. With the advancements in LLM's ability to do many things, including what initially appeared to be it's "reasoning" capabilities, the door opened for the third wave, now being called Agentic Ai that uses "reasoning models". Early movers in this space include open frameworks like BabyAgi, AutoGPT, or LangChain, in which one can breaking down more complex requests into steps and then execute those tasks in some kind of order to produce the final result. This concept forced the LLM to “Stop and Think”.

LLM Limitations

Recent articles on LLMs reaching their limitations highlight the challenges of enhancing logical reasoning in pre-trained models. You can only train an LLM on so much data before the point of diminishing returns. Models are effectively "baked" in the oven and launched with limitations in feeding it real-time data (see context windows, RAG, etc..). In order to expand reasoning capabilities the way we need it - in a more real-time manner, new approaches and architectures were explored and tested.

This shift in design of reasoning Ai systems is laid out in a good article from Sequoia Capital - one of the biggest VC firms investing in tech in recent decades. The article lays out a shift from System 1, pre-trained models to System 2, reasoning models. While LLMs will be a big part of the future, they are evolving into something new, models that focus more on real-time reasoning instead of just referencing massive pre-trained data. The OpenAi and Anthropics of the world have been enhancing the models to ultimately “stop and think” before they respond.

This evolution has a number of implications, primarily related to when and where the "thinking" happens. We are now moving from today's “thinking fast” LLMs (system 1), toward more deliberate reasoning during inference, known as “thinking slow" (system 2). This transition emphasizes the importance of “inference-time (or test-time) compute,” where AI models allocate additional computational resources during inference to enhance their reasoning capabilities. A prime example is OpenAI’s o1 model, which demonstrates advanced reasoning by pausing to “think” before responding, thereby improving performance in complex tasks like coding and mathematics. This approach mirrors the strategy employed by AlphaGo, which, beyond its initial training, simulated numerous potential future scenarios during inference to determine the most optimal move (at run-time).

As AI research progresses, leveraging inference-time compute is becoming pivotal in developing systems capable of deeper reasoning and problem-solving. An interesting recent development is the release of deepseek R1 that jolted the stock market with it's announcement of the model being trained on only $6m of compute. While that is indeed a significant discount to work that OpenAi, Anthropic, and Google are doing, it is also important to note that the compute needed to train giant LLMs is moving from training to test-time compute. More compute will be needed at run-time because more of the reasoning is happened there instead of in the training model. This is evident in their response time. Pre-trained LLMs respond quickly because they already have the answers whereas reasoning models require a slower response because it is working out the answer. Beyond the amount of time this can be seen in the amount of tokens required to return a final answer.

As will be discussed, the "stop and think" is a good next step in the evolutionary path of Ai, but the question remains "is slowing down LLMs to process new information repeatedly equivalent to human-level intelligence, or do we need entirely new architectures?" More on this later.

How Agents Work

Before getting into anything too technical, let's take a look at some of the characteristics of "agentic" behavior. First and foremost, Agentic Ai should include some form of Goal-oriented Behavior. This is the agents north star and everything it does should effectively align with that. In the case of ChatGPT it has a goal - that goal is provide an answer, even if it’s wrong. Your agent's goal could be to book a trip, finish a grant application, research a business deal, etc...

Once the agent is imbued with a goal, it can essentially be described with 3 key elements: perception, cognition, and action. This Article breaks these factors down nicely.

Source: https://markovate.com/blog/agentic-ai-architecture/

Perception

Machines perceive similar to how humans perceive. Through language, visuals, sound, sensors, and environment. LLMS take real-time information in through In-Context Learning - their prompts (aka context windows). Multimodal takes this a step further and allows users to upload images and documents to provide additional context. LLMs can also be connected to datasets through RAG (Retrieval Augmented Generation) for additional contextual information. If you plug them into File Systems and the mechanisms to do some searching they can look for files and do a search similar to how a human would do it. Increasingly these systems will be able to take in more real-time data from cameras, sensors, microphones, and other data sources to add much needed information in the reasoning process. As will be discussed later, the hyperscalers (AWS, Microsoft, Google, and others) are well positioned for Agentic Ai as this is where may folks already store this information.

Action

Actions can be defined as what the agent "does". This can include Content Generation, Communication (emails, teams, etc..), Selling a Stock, Rating a Product, Etc… Theoretically actions are anything that you give permissions and functionally for the agent to do. Agents do this through system access, integrations, API’s IoT devices and more. For example, an Agent in AWS can make an API call to a web-search like Tavily or SerpAi to crawl the web. It can then take the results it finds and call another API to do something with that information.

Cognition

Where perception and actions are more straightforward, cognition is where things gets tricky. What exactly is meant by "cognition" or "reasoning"? Because we are building systems in our likeness, the top benchmark we measure against is human capabilities. This cognition can be described in a variety of ways.

Cognition may include but is not limited to:

• Contextual Reasoning

• Deliberation

• Decision-making

• Problem solving

• Planning

• Collaboration

• Adaptation

• Ethical Considerations

As to how these are measured, researchers have scores of benchmark tests to prove and disprove reasoning capabilities (along with the companies building the LLMs). These benchmarks include ARC, HELM, and many more, and recently have been joined by the somewhat ominous humanitys last exam. Beyond measurable cognition, one can dive down the rabbit holes of exploring deeper human questions such as "what is consciousness?" or "what is sentience". A simple definition can be found on the internet or at your local LLM, but in reality, humans are still searching to fully understand what consciousness is. A more complex topic for another post. Suffice it to say, as we go about trying to ascribe our human language to machines, we may end up furthering understandings of ourselves.

But back to the task at hand. How do LLMs actually reason? At the heart of "Agentic Ai" agents are LLMs. While there are likely some more complex elements involved, an LLM (transformer architecture, attention mechanisms, etc..) lies at the center. Reasoning agents like ChatGPT o1 are essentially LLMs that are designed to 'slow down and think' as described above. For those familiar with prompt engineering, this is similar to Chain of Thought prompting. CoT breaks down complex problems into a series of logical steps and encourages the model to think through intermediate reasoning before arriving at a final answer.

As impressive as they are, LLMs aren't the only show in town. While LLMs are good at some things, they are not great at everything. Each model also has it's strengths and weaknesses. Machine Cognition is also comprised of Machine Learning Algorithms that have been used for decades. Tasks like Fraud detection can be best aligned to algorithms and not a text-based LLM. Beyond traditional ML models, LLMS can leverage approaches like PAL (Program Aided Language Models) that defer to other top shelf tools and frameworks for the task at hand. This approach takes in a request and will write SQL or Python or other code to best execute and respond with much improved accuracy. LLMs can also use or integrate rules. These rules can be embedded in the request itself (write 3 paragraphs in the style of Seinfeld) or can plug into complex rules systems that are more deterministic in nature.

Deterministic vs. Probabilistic Systems

The overlap or integration of traditional deterministic rules systems with probabilistic Ai systems is a topic to be explored unto itself as we march towards AGI. Amid the complexity of developing an Ai based system (or integrate Ai into an existing workflow) is distinguishing between where we program a system to do what we say only (deterministic rules), and where do we give it power to make decisions (probabilistic). Agentic Ai doesn’t solve this problem, it actually brings greater scrutiny to a challenge debated for decades now. The looming question I see with Agentic Ai is who do you blame when it goes off the rails? This risk is likely to increasingly be mitigated by the marriage of deterministic and probabilistic systems. This can be seen today through "guardrail" frameworks and prompt engineering in which provide guidance and rules of what systems can and cannot do.

So what does all this look like? If we take a couple of examples from the hyperscalers (Microsoft and AWS) we can see these behaviors correlated to systems.

• Perception = Data

• Apps = Actions

• Cognition = LLMs

The following shows how LLMs are dropped into the massive ecosystems of Microsoft and AWS, are given access similar to a human and are able to leverage data and take action in apps to achieve their objectives.

With cognition comes the ability to plan and orchestrate tasks and information. Armed with the tools described above, a central agent can take in a request (or create one) and make decisions on what program/framework/algorithm/rule is best to execute on a given task. This brings us to Multi-Agent Systems.

Multi Agent Systems

Much more than the use of a single agent, one can employ many agents in what is known as Multi-Agent Systems (or could also be considered swarms). A central agent can orchestrate tasks and spin up as many agents as needed (and is cost effective) to complete a task. For example I want an agent to write me a grant proposal (a big task, costing many thousands of dollars today). That agent will have or find instructions for the grant application and spin up a number of writing agents for each aspect of the application, an agent that uses PAL to do financial projections, a style agent, an editor to review everything and a grant QC reviewer to evaluate quality of the submission.

Mixture of Experts

This brings us to the concept of Mixture of Experts. MoE uses the concept of sparsity, an approach of activating only a subset of specialized expert models for any given task, reducing computational overhead while maintaining efficiency and accuracy. Much like us humans call a plumber to fix the sink and general contractor to build a house, this approach centralizes management of a goal, but defers to the experts to execute.

Beyond the hyperscalers building this into their platforms, startups like Ema are building their own proprietary Mixture of Experts model in which they select the right model or agent for the right job. For example a math problem might best use a ChatGPT model that employs PAL (program aided language models) to calculate things on a spreadsheet whereas Claude may be used for a writing assignment. Note that this model maintains a design in which we have LLMs, Apps, and data as the primary pillars of the agent.

The concept of Mixture of Agents is prescient of the increasing level of complexity in reproducing human intelligence and labor. As LLMs and reasoning models evolve, so to do expectations on what Agents can do. This expectation is pushing forward on concepts old and new that move us in the direction of architectures more cognitive in nature.

Cognitive Architecture

Multi-agent systems and mixture of experts represent a movement towards diversity of thought and a step closer to more complex systems of Cognitive Computing such as Cognitive Architecture, Neuromorphic Computing, and Neuro-Symbolic AI. Cognitive Computing is inspired by cognitive science and human thought processes. It focuses on high-level abstraction of human reasoning, language understanding, and learning. We will cover more on cognitive computing in a future post.

As described, Agentic AI represents a significant evolution in artificial intelligence, moving beyond pre-trained models toward real-time reasoning, adaptability, and autonomous action. By integrating perception, cognition, and action, these systems can dynamically interact with their environment, execute complex workflows, and even collaborate within multi-agent ecosystems. The rapid development of reasoning models, the integration of deterministic and probabilistic approaches, and the rise of mixture-of-expert frameworks signal a fundamental shift in how AI systems are designed and deployed. However, this transformation is not without its challenges—business leaders, technologists, and policymakers must navigate the implications of these advanced agents in the workplace, from automation and augmentation of jobs to governance, security, and ethical concerns.

In the next post, we’ll dive into the business implications of Agentic Ai, exploring how enterprises are adopting and monetizing these intelligent systems.

While some embrace Ai and are quickly able to find meaningful ways to use it, others struggle to understand and adopt. This adoption curve applies to much more than Ai, as can be understood through books like Crossing the Chasm. Ai, particularly Gen Ai like ChatGPT, comes with some additional complexities like knowing the language of LLMs and how to get creative with them that contribute to the phenomena of “cognitive load”.

In 1957 Social scientist Herbert Simon developed a Nobel Prize-winning concept called bounded rationality that examines how people make decisions with limited cognitive resources and incomplete information. His work revealed that, rather than acting with perfect rationality, individuals often simplify complex realities to make them more manageable, which impacts their ability to process new information effectively. In 1988 Australian psychologist John Sweller developed cognitive load theory, which explores how the brain’s working memory has a limited capacity, making it susceptible to overload when faced with complex tasks. Sweller’s theory highlights that when cognitive load is high, learning becomes difficult, as individuals struggle to process and retain new information.

Bounded rationality and cognitive load theory help shed some light on the very human cognitive challenges that arise when learning about new and, at times, complex concepts like Ai. Our brains are only wired to handle so much, and each of our brains work a bit differently depending on our exposure to the concepts we are learning. Technology adoption is an area of high potential for cognitive overload on a good day and can be even higher when the adoption requires us to change our behavior or think differently, which ultimately, Ai does.

So how do we break through the cognitive “load block”?

In a word, intentionally.

Our brains are great at helping us avoid the thing we need to do. Intentionality means becoming aware of these limiting behaviors and having a plan that includes openness, engagement, and consistency to push forward through and around them.

Let’s take a look at some of these human behavioral tendencies, strategies to counter these behaviors, and examples of how to engage with LLMs to practice.

Fear of the Unknown → Cultivate Curiosity

A natural response to new technology (or more appropriately, new ways of thinking) is apprehension. By fostering curiosity, individuals can transform AI from a source of anxiety into a pathway for exploration and growth. Viewing AI as an opportunity to enhance life instead of something to fear allows for a smoother, less intimidating learning and adoption process.

Action: Use an LLM to explore “What are some creative uses of AI in my field?” and let it suggest unique applications to spark your curiosity and broaden your understanding.

Resistance to Change → Embrace Change as a Constant

The world is constantly changing, yet people often resist change due to a desire for stability. Flip this by adopting a mindset that views change as an ongoing, inevitable part of progress. This approach enables individuals to see adaptation as a core competency, rather than a burden, and to remain flexible in the face of rapid advancements.

Action: Ask an LLM "Please provide an example of some prompts that can help me with [insert curiousity here]"

Perfectionism → Normalizing the Learning Process

The drive for immediate mastery, in ourselves and in the products we use, can lead to frustration and cognitive overload. By normalizing that learning takes time, individuals can release the pressure to perform perfectly from the start. This applies not only to ourselves but to the Ai. Remember this is the worst Ai will ever be. Emphasizing growth and learning at one’s own pace and appreciation of evolving platforms fosters a culture where experimentation is valued and errors are seen as stepping stones to improvement.

Action: Look at resources like promptengineering.com to learn the language of talking to LLMs and ask the LLM on things like style, details, constraints, and more to get the best prompt.

One-Size-Fits-All → Personalized Learning

People often struggle with the unknown. Prompts are great, but what do you put in them? Personalized learning addresses this by allowing individuals to engage with AI on their terms. Asking yourself questions like what takes up your time and what would you like help with are good ways to build your personalized learning journey. Once those are established you can begin trying prompts, doing searches online or even asking an LLM how to learn about using the LLM for your situation.

Action: Take a problem you are interested in solving, break it down into steps,  and experiment with Ai to see where it can (and can't) be helpful. Example "I'm researching ways to make travel experience better for airline passengers, please provide a plan for better understanding what would make for a better flight experience."

Complexity → Leveraging AI to Optimize Learning

We tend to avoid the complex, sometimes in subtle ways. The nice thing about Ai is that we can leverage Ai to learn about how to use Ai. AI-powered learning tools can break down complex ideas into more manageable segments, making the learning process more accessible. Using AI to learn about AI helps reduce the mental burden and accelerates familiarity with these tools.

Action: Ask an LLM "Please list out the top 10 prompting techniques I can use for formulating a business plan" or "how does prompt engineering help me get better results?"

Overcommitment → Incremental Improvement

Attempting to master everything at once can lead to burnout. Starting small and incrementally learning how Ai can be incorporated into one’s workflow, builds confidence through small, manageable steps. By focusing on simple tasks initially, individuals can adjust at a comfortable pace, reducing cognitive overload and providing the small wins to keep us going.

Action: Use the LLM to help automate a single repetitive task you perform regularly, such as drafting emails, and gradually expand to other tasks as you become more comfortable.

Inconsistency → Steady Engagement

Sporadic use of new tools often results in fragmented learning and discarding of valuable tools. Much like building a muscle, the brain needs reinforcement to use a tool and improve results, bringing further value to the tool. Consistent interaction with AI helps reinforce knowledge and transforms new skills into ingrained habits. By integrating AI into daily routines, users can develop a strong foundation that supports long-term mastery.

Action: Set a daily reminder to use the LLM for quick tasks, like brainstorming ideas or summarizing articles, to build consistency and reinforce learning.

Hype → Healthy Skepticism

The Hype Cycle exists for a reason. Marketing and collective hype tend to over inflate expectations before crashing into disillusionment. Approaching Ai with realistic expectations is a practical way to adopt technology over time. By approaching with healthy skepticism and a practical approach, one can adopt what's valuable now and build the muscle of adopting the tech as it evolves.

Action: Try a variety of different prompts to learn what it does well and what it does not do well. Read up on the latest in Ai trends, including where Ai falls short.

Disconnection → Track Progress and Measure Impact

Many shy away from measuring progress or impact for a variety of reasons, but as with much of life we need mile markers. Tracking the impact of AI on work/life provides tangible evidence of benefits, reinforcing motivation and highlighting areas for further improvement. By maintaining a record of accomplishments, individuals can stay engaged and make informed decisions about AI integration.

Action: Estimate how much time the Ai platform saves you each time you work with it. Add that up to appreciate the benefits it provides

Stagnation → Iteration

The fear of mistakes or getting poor results can discourage experimentation. LLM's often fail to produce the desired results. Much of the time this is due to the user not providing the proper level of detail, context, constraints, and instruction. Approaching Ai with an iterative mindset can help flip a "one and done" mindset to one of continuous progression. This mindset encourages ongoing adaptation and reduces cognitive load by breaking tasks into manageable phases.

Action: Use Prompt Chaining to iterate through a task with the LLM and improve the prompts and results as you go. Ask the LLM for feedback on completed tasks or ideas for further improvement.

By recognizing our own human tendencies and becoming more self aware we are able to challenge our own limitations and embrace change. An Ai powered world is inevitable and as with the industrial revolution, computers, and the internet it will be important to adapt to a changing landscape. Using any of the above points of understanding and actions can put you one step closer to pushing through the cognitive load block and find ways to leverage the power of Ai.

Follow for more.