AI does not need another committee. It needs accountable leadership. That should, for most big organisations, be a CAIO. Smaller businesses should have a fractional CAIO.

BUT - the center of gravity for CAIO skills is all wrong.

In 2020, there were no Chief AI Officers. In Adopting AI (2025) - though the role had grown some, I argued this was a huge mistake.

Now there are a lot of CAIOs - but I wonder whether the way the role is conceived of is correct.

Share

Subscribe now

Failure mode #1 - far too junior

Where you position a role matters hugely. When (a decade ago), I was advising boards on ESG-Sustainability, the head of ESG was generally buried somewhere.

That is awful symbolically - leaders saying HEY - sustainability matters and is central to our strategy - but OH, it reports through legal, 3 layers down.

I recall the head of sustainability for an airline trotting off to report their findings to a country head and being tossed out of the meeting - a physical exit - but most leaders were more subtle - they’d read the sustainability recommendations and then (as I believe the miliary say) - grin f*#k the guy.

Failure mode #2 - wrong skill portfolio

AI is complex at every level - unartfully captured below.

But that isn’t the half of it - because that is “static” - and what becomes really tricky is when that stack interacts with human systems:

• existing tech

• organizational structure and politics

• human cultures

• human intelligences

So, the CAIO needs to be full stack - and MOREOVER, able to manage the interaction between layers of the stack AND the interaction with the existing org.

But we don’t choose CAIOs that way.

The CTO gets AI because it sounds technical. The CIO gets AI because it touches infrastructure. The CDO gets AI because models need evidence.

Each answer seems sensible until it meets the business.

AI adoption is not a technology deployment problem with a bit of human clean-up attached. It is organisational transformation with a technical core. That distinction matters because the hard constraints are increasingly human: adoption, trust, workflow redesign, governance, ethics, culture, and executive accountability.

The 2026 Data & AI Leadership Executive Benchmark Survey puts the point bluntly: 93.2% of AI adoption impediments are human issues.

That should change the leadership conversation.

If the binding constraint is human, then handing AI to a purely technical function is not strategy. It is category error.

The case for a Chief AI Officer is not “AI is important, so create another C-suite title.”

The case is that AI now crosses too many domains to be safely absorbed by one existing silo:

- AI strategy and prioritisation

- Governance and risk

- Human capital and adoption

- Culture and ethics

Consultants can help. Vendors can help. Committees can coordinate.

None of them supplies accountable leadership.

I have published a discussion document making the full case: why the CTO/CDO/CIO absorption pattern fails, why consulting-led AI strategy often stalls, why vendor governance is not enough, and why a fractional CAIO model may be the practical path for mid-market firms.

Full whitepaper for Substack subscribers:

Understand the frontier of AI leadership - get the whitepaper

Leadership does not suffer from a shortage of models. Transformational leadership, servant leadership, authentic leadership, adaptive leadership, emotional intelligence, psychological safety, and a long list of X-leadership theories have all made real contributions. I identified nearly thirty of them in my own leadership research in the 1990s.

So any claim that AI needs another leadership framework has to earn its keep. The burden of proof is not novelty; it is changed conditions.

THE CONDITIONS CHANGED

My argument in The AI Leadership Delta is that the conditions have changed. AI is not simply another enterprise technology to be rolled out: it can generate, recommend, decide, and act with partial autonomy. It spans technical layers no executive can master alone, creates ethical questions faster than many governance systems can answer them, and unsettles expertise, identity, trust, power, and the meaning of work.

Established leadership models remain valuable. Their relational axis is more important than ever: trust, psychological safety, meaning, and innovation climate are load-bearing. But relational capability alone is no longer enough.

TWO AXES, NOT ONE

AI demands strategic judgement that many leadership models address only partly:

🧭 Make build-or-rent decisions under uncertainty

⚙️ Sequence augmentation and automation

⚖️ Govern while regulation is still moving

🛠️ Develop direct contact with the technology

🧪 Practise build-test-kill discipline

🧱 Orchestrate people, process, evidence, and tools as one system

That is the gap between the leader you were trained to be and the leader this moment demands. The point is not to replace older models, but to locate what they still do well, identify what the AI moment adds, and turn that addition into observable practice.

THE LEADERSHIP DELTA

The framework uses seven dimensions across three domains: Lead Self, Lead Others, and Lead System. It asks what leaders practise themselves, what they make possible in relationships, and what they build into the organisation.

The work is not to admire another framework. It is to change what leaders do.

The whitepaper is free for subscribers. Head over to paulgibbonsadvisory.com/publications for a free download.

Subscribe now

What organizational form would we create if intelligence, memory, coordination and execution were predominantly machine capabilities rather than scarce human capabilities?

Nobody yet has a convincing answer. There is emerging literature on “AI-orchestrated” and “fluid” organizations, but most of it still begins with the existing enterprise and asks where agents should be inserted. It assumes human accountability, human organizational goals and recognizable human roles. That is organizational augmentation, not first-principles redesign.

The modern organization is an accommodation to human limitations

The familiar corporation is not a natural or inevitable way to organize productive activity. It is an evolved response to several characteristics of human beings.

Humans have:

• narrow expertise acquired slowly;

• limited working memory and attention;

• difficulty maintaining context across large systems;

• social and status needs;

• finite communication bandwidth;

• uneven motivation;

• geographical and temporal constraints;

• and a need for authority, identity, belonging and career progression.

Departments reflect educational and professional silos: finance people go into finance, marketers into marketing, engineers into engineering. Hierarchy compresses information because the chief executive cannot communicate meaningfully with 20,000 individuals. Middle management filters, translates, prioritizes, coordinates and reconciles.

Much of what we call “management” is therefore human glue:

• moving information between functions;

• convening meetings;

• resolving ambiguities;

• persuading people to cooperate;

• translating one professional dialect into another;

• locating knowledge;

• monitoring progress;

• allocating attention;

• and preserving organizational memory through personal relationships.

That glue is expensive, slow and lossy. But historically it was necessary.

Computing automated transactions, not organization

This explains why fifty years of accessible computing barely altered the fundamental organizational form.

ERP systems computerized transactions. Email accelerated messages. Databases stored records. Workflow systems routed forms. Collaboration platforms proliferated communication. But none of those systems could understand what the organization was trying to do, interpret ambiguity, generate a plan, reconcile inconsistent information or act autonomously across functional boundaries.

Indeed, much enterprise computing reinforced the organization chart. Finance acquired finance systems, HR acquired HR systems, sales acquired CRM, operations acquired ERP and IT became responsible for connecting them. The digital enterprise reproduced the functional enterprise in software.

It often made the organization faster at generating information while simultaneously creating more need for human coordination.

Generative and agentic AI are different because they attack the semantic and coordinative layer. An agent can interpret, synthesize, decide, generate, negotiate and execute across systems. OpenAI reports that its own use of Codex has spread beyond engineering into departments such as legal and recruiting, with agents increasingly performing longer and more cross-functional work. Yet even this remains largely an example of capable tools entering existing departments—not proof that departments remain the correct structure. OpenAI

The crucial change is not task automation

Most corporate AI programs ask:

Which human tasks can we automate?

That question preserves the existing decomposition of work. It assumes that jobs, processes, functions and reporting relationships are approximately correct and that AI should replace pieces of them.

The first-principles question is:

Given a goal, a set of assets, constraints and legitimate stakeholders, how should intelligence and execution be organized?

That may produce no recognizable HR department, marketing department, product department or middle-management structure.

The machine-native organization might instead be composed of:

• persistent objectives;

• constraints and policies;

• a shared institutional memory;

• dynamic capability agents;

• temporary mission structures;

• decision and escalation protocols;

• continuous sensing and feedback;

• and a relatively small number of humans responsible for legitimacy, purpose and exceptional judgment.

The unit of organization would no longer be the job. It would be the mission.

The unit of capability would no longer be the department. It would be an addressable combination of models, tools, data, permissions and expertise.

The organization chart would become a dynamic computational graph.

Hierarchy would change function

Hierarchy probably does not disappear completely. But its primary purpose would change.

In the traditional organization, hierarchy serves two distinct functions:

1. Information routing and coordination

2. Authority, accountability and legitimacy

Machine intelligence could eliminate a large fraction of the first. Information would no longer have to travel upward through successive managerial filters and then back downward as instructions. A shared system could sense an event, interpret its implications, identify affected activities, generate options and initiate authorized actions almost immediately.

But the second function remains. Someone must define the objective, bear legal responsibility, resolve genuinely normative questions and represent the organization to society.

So hierarchy may survive as an authority topology, while disappearing as an information architecture.

That distinction is foundational.

A machine-native enterprise might have very few permanent managerial levels while retaining explicit human control over capital allocation, legal commitments, safety boundaries, public consequences and changes to the organization’s constitutional objectives.

Tacit knowledge is the present bottleneck

Right how, humans are the keepers of tacit knowledge. Perhaps not for long. As AI systems get embedded into workflows, advanced AI memory systems will have ontologies for capturing tacit knowledge. That is, the esoteric knowledge that lives inside people’s heads will become more explicit.

Maybe not ALL tacit knowledge will become explicit. Some knowledge is embodied, relational and contextual: how a particular executive reacts under pressure, whether a supplier is concealing anxiety, how a manufacturing process sounds immediately before failure, or what a client means when the formal words say something else.

But the economically decisive question is not whether all tacit knowledge can be codified.

It is whether enough of it becomes machine-accessible that humans cease to be indispensable storage devices.

Continuous transcripts, workflow traces, decision records, video, operational telemetry, correspondence, model-generated summaries and institutional memory systems can capture vastly more of organizational life than conventional documentation ever did. Models can infer patterns from behavior without requiring the human expert to explain every rule.

Thus the human expert may remain the original source of some knowledge without remaining its sole custodian.

That is a profound change in bargaining power and organization design. Today, when a senior engineer, salesperson or operations leader leaves, a network of undocumented judgments often leaves with them. In a sufficiently instrumented organization, departure may mean losing a relationship or an exceptional mind—but not losing the entire history of how work was performed.

AI as customer changes the external boundary of the firm

The purchasing point is equally important.

Agentic commerce is already moving from recommendation toward execution. Visa is building infrastructure for AI-initiated transactions, agent authentication, authorization and spending controls. Gartner forecasts that by 2030, half of cross-functional supply-chain-management solutions will use intelligent agents to autonomously execute decisions; it explicitly gives autonomous procurement as an example. visa.com

Consumer forecasts are also substantial: Bain estimates that agents could influence or complete 15–25% of US e-commerce transactions by 2030, although end-to-end consumer trust remains a constraint. bain.com

B2B technology is likely to move first because it is unusually machine-legible:

• specifications are relatively formal;

• performance is measurable;

• prices and licenses are digital;

• products can be trialled automatically;

• security and compliance evidence can be queried;

• and delivery may occur immediately through APIs or software deployment.

The future buyer may ask:

Find the database, model, cybersecurity service or cloud configuration that satisfies these latency, regulatory, reliability and cost constraints. Test the candidates, negotiate terms within this authority envelope and purchase the best option.

At that point, the seller is no longer primarily persuading a person. It is satisfying an evaluative system.

That changes sales and marketing fundamentally. Machine customers will care about:

• accessible product data;

• verifiable performance;

• machine-readable pricing;

• trustworthy provenance;

• security attestations;

• integration friction;

• contractual constraints;

• service history;

• and evidence that can be independently tested.

Brand does not disappear, because reputation itself becomes a decision variable. But brand must become computable trust rather than merely emotional salience.

The B2B sales organization built around account executives, sales engineers, procurement negotiations and months-long human courtship could shrink sharply in categories where buyer and seller agents can establish fit, risk and price directly.

The IBM problem

IBM is a corporation with more than a quarter-million humans contains enormous amounts of coordination latency. Every major decision encounters functions, geographies, systems, incentives, committees, budgets, permissions and local interpretations.

My experience as a partner is illustrative. Hired to run their leadership-future of work “eminence” (thought-leadership) I began to produce - and as an experienced researcher-writer - a serious piece of content to compete with our competitor firms, Accenture and McKinsey, in about 3-5 weeks. However, getting IBM thought-leadership pieces through the system might take six months, even for something as short as a blog post.

In other words, the system was explicitly designed with frictions so slow innovation and output.

And in the AI world, six months is an eternity. Little that was written in 2025 has any relevance today.

IBM has extraordinary assets:

• customer trust;

• installed infrastructure;

• data;

• patents;

• capital;

• regulatory knowledge;

• distribution;

• and deep domain expertise.

Those assets may remain highly valuable. The human bureaucracy used to coordinate them may not.

IBM need not disappear as a legal entity or commercial platform. But it may be unable to survive in its present organizational form if AI-native competitors can produce comparable outcomes with one-tenth or one-hundredth the human coordination burden.

The successful incumbent would have to become something resembling a federation of machine-intensive enterprises, sharing capital, governance, infrastructure and brand while operating with far more autonomous and computationally mediated units.

That would be transformation in the literal sense—not a workforce with better copilots.

The strongest counterargument

There is a serious alternative possibility: AI may make very large corporations more viable rather than less.

Historically, organizations stop growing when internal coordination costs overwhelm the advantages of scale. If AI dramatically lowers those internal coordination costs, a corporation might manage far more products, markets and activities than any human bureaucracy could.

A gigantic enterprise with unique data, enormous compute, established customer access and powerful agents might coordinate resources more effectively than a small entrant. The result could be fewer but even larger firms.

There are also durable human constraints:

• law assigns responsibility to people and legal entities;

• legitimacy cannot simply be computed;

• objectives remain contested;

• power and politics do not vanish;

• customers may demand human accountability;

• physical work remains significant;

• and machine systems can fail systematically at enormous scale.

Therefore, “AI means every large company dies” is too strong.

IN THEORY, AI could make bureaucracy more efficient.

However, bureaucrats have something to say about that. Turf to protect,: people whose job it is to make sure other people don’t do their jobs.

This is a prediction - take it as such - very few large enterprises have the change capability to do this. Perhaps none. And there is a moral case - those bureaucrats have families to feed.

My view

I expect a bifurcation.

We will see:

• a small number of extremely large platform, infrastructure and capital-owning enterprises;

• and a vast number of very small, extraordinarily capable AI-native operating companies.

What becomes endangered is the middle: the large, labour-intensive, managerial corporation whose principal competency is coordinating thousands of knowledge workers through hierarchy and process.

The corporation may remain large in revenue, assets, reach and computational activity while becoming radically smaller in human headcount.

So my summary view:

The twentieth-century corporation was built to coordinate scarce human intelligence. The AI-native corporation will be built to govern abundant machine intelligence.

And that means we are not merely building agents, memory or orchestration.

The square peg is not AI.

The square peg is the inherited corporation.

Master the Craft is pillar one of the Adaptive Adoption change framework

In 1996 I found research that has haunted me ever since: roughly ten percent of management training ever transfers back to the job. Ten percent. We have spent three decades building curricula, modules, and certificates despite that fact — and now we are doing it all over again with AI.

We all intuitively know this, the webinar you half listen-to and don’t use. The amazing 3 day conference that leave you a sugar high, and with unchanged behaviors.

This is the argument of Master the Craft: AI capability is not a curriculum problem. It is a craft problem — closer to cabinetmaking than to compliance training. You do not learn it by completing modules and reporting licence utilisation. You learn it by doing real work, under uncertainty, with feedback, tools, standards, and a community of practice.

The 2023 question — who uses AI and who doesn't — is settled. The 2026 question is sharper, and most adoption dashboards cannot see it: are your people using AI to offload cognition, or to augment it? The same model produces either. If the path of least resistance is to let the machine think for you, most people will take it. That is not a character flaw. It is system design — and it is fixable.

• Who is teaching AI craft in your organisation — practitioners who build daily, or trainers handed a deck last month?

• Are you defaulting to pricey business school certificates?

(The latter may do more harm than good… business school professors - apart from a few - are way behind and the only one I know who is competent as a builder is Mollick.)

Master the Craft seven dimensions cover way more than literacy:

1) Ethics and civics are rarely taught well and are taught as theory

2) Meta cognition and critical thinking are assumed not properly taught

3) Agentic design requires creativity. Literacy frameworks assume it and don't teach it

4) We are creating hybrid teams… how do you delegate and orchestrate with AI in the mix?

The full report is available for subscribers

In August, the first cohort of the Master the Craft program goes live. DM me for the application.

If you want to stay abreast of what is new in the human-AI-enterprise world - what I call adaptive adoption, you have come to the right place. Kindly consider subscribing.

Read more

I recently exhumed a whitepaper I drafted in 2022, back when I was at IBM. It was about “talent hyper-mobility” — the idea that knowledge workers, untethered from geography, would enjoy the greatest surge in career agency in a generation. Workers could take a dream job 5,000 kilometres away without moving house. Employers in Chattanooga could poach from Silicon Valley. The war for talent would be won on purpose and culture, not compensation.

Four years later, nearly every prediction in that document has inverted. The quits rate, 3.0 per cent at its 2021 peak, has been frozen at [1.9 per cent for the better part of a year. Hiring, as measured by the JOLTS report, has slowed to levels last seen in 2010, when unemployment was near 10 per cent. Economists call it the Big Stay: low hire, low fire, nobody moves. The worker who could once field three offers from three time zones now clings to the job she has, wondering which of her tasks an AI agent will absorb next.

I could quietly bury the document and hope nobody ever finds it. Instead I want to do something more useful: ask which ideas in it survived the inversion. Because a few did — and they matter more now, not less. The half-life of labour-market predictions is about eighteen months. The half-life of the psychology underneath them is measured in decades.

The semantics of balance

Misgivings about “work-life balance” have circulated for years. Every article promising the working world’s nirvana is matched by one calling it a myth. So we got alternatives: work-life *integration*, work-life *blend*, work-life *harmony*. The US Chamber of Commerce defines integration as handling personal and professional responsibilities “at the times that work best for them” — interspersed through the day, no hard line between the two.

I am sceptical that channelling our frustration into semantics helps. A draining job does not become less draining because we update our metaphors. Rebranding balance as harmony is the linguistic equivalent of throwing a retention bonus at a toxic culture — it addresses the label, not the thing.

But beneath the branding exercise sits real research, and it points somewhere more interesting. Since 1996, organisational psychologists have distinguished two ways people manage the border between job and life.

• Integrators prefer a seamless blend — the school run between meetings, the inbox on holiday, work and life as fibres in one weave.

• Segmenters want walls — separate phone, separate desk, a hard stop at six. Most of us sit somewhere on the spectrum between the two, and where we sit is a fairly stable individual difference, not a failure of discipline.

My favourite study in this literature examined Episcopal priests — people whose “employer” is, in the most literal sense, always watching. Kreiner, Hollensbe, and Sheep found priests being approached for confession at the petrol station, in the supermarket, mid-errand. Even those with a sacred calling — the ultimate “purpose, not a paycheck” — had to build boundaries to preserve their sanity. If the clergy need to segment, perhaps your employees are allowed to as well.

The practical upshot is what Kreiner calls boundary fit: wellbeing depends less on which style you have than on whether your workplace lets you live it. A segmenter in an always-on culture suffers. So does an integrator under a rigid nine-to-five. This single finding demolishes most corporate wellbeing policy, which insists on solving an individual-difference problem with a blanket rule.

When the music stopped

Here is what my 2022 draft got right for the wrong reasons. I argued that when labour markets are static and workers cannot move, they “suck it up” — they rationalise, settle, and stop asking dangerous questions like *what do I really want from work?* Ruminating over what you cannot change is corrosive; adaptation is the sane response. Mobility, I argued, was the solvent: when workers could leave, employers would finally have to compete on culture.

Then mobility went away. The logic, however, did not.

In a frozen market, the bargaining power that made work-life boundaries defensible evaporates. The worker who might once have declined the 8 p.m. meeting — because a recruiter called every week and her employer knew it — now weighs that refusal against a hiring rate at fifteen-year lows. Gallup’s most recent global data show engagement falling to 20 per cent, its second consecutive annual decline. Managers, the people meant to protect their teams’ saw their boundaries fall fastest. The Big Stay is not contentment. It is people gripping the rail.

And into precisely this moment — maximum employer leverage, minimum worker exit — arrives the most boundary-dissolving technology since email.

The infinite workday

Microsoft’s telemetry across its 365 platform gives us something the balance debate rarely enjoys: measurement. Their 2025 Work Trend Index describes what they call the “infinite workday”: meetings after 8 p.m. up 16 per cent year on year, the average employee interrupted every two minutes — 275 times a day — and by 10 p.m., 29 per cent of active workers back in their inboxes. The average user now receives 58 chat messages outside core hours. Half of employees describe their work as chaotic and fragmented.

Read that alongside the boundary research and a diagnosis suggests itself. The infinite workday is an integrator’s world imposed on everyone, including the roughly half of humanity wired to segment. We have built a working rhythm that matches one boundary style, universalised it, and then wondered why burnout metrics look the way they do.

Legislatures have noticed. France legislated a right to disconnect in 2017; Portugal attached fines of up to €9,690 per violation. Australia’s law, was prompted in part by Australia Institute research finding employees averaging 5.4 hours of unpaid after-hours work per week. I favour these laws the way I favour speed limits — necessary, insufficient. You cannot legislate a culture into respecting time off; you can only make its disrespect slightly more expensive. The laws are best read as a symptom: when boundary-setting has to be outsourced to parliament, the labour market has stopped doing the job.

Does AI make this better or worse? Yes.

The lazy answer is that AI worsens always-on culture. That is an oversimplification: AI is an amplifier, not an author. It intensifies whatever relationship with work an organisation already has.

Start with the worsening case, because it is currently winning. Researchers at BetterUp Labs and Stanford coined the term “workslop” for AI-generated output that masquerades as good work while lacking the substance to advance the task. Their survey of US workers found 41 per cent had received workslop in the prior month, each instance costing nearly two hours of rework. — and, worse, recipients rated the senders as less capable and less trustworthy afterwards. AI here does not reduce work; it launders effort, shifting the cognitive burden downstream to a colleague. Always-on culture plus workslop is a particularly nasty compound: infinite availability to process infinite mediocrity.

Then there is the ratchet. When a task that took a day takes an hour, the rational firm does not bank the surplus as leisure; it raises the baseline. Content teams are expected to produce multiples of their pre-AI output; engineering sprint velocities get recalibrated upward. The technology removes the natural pauses that knowledge work once contained — the twenty minutes of formatting that doubled as cognitive recovery — and back-fills them with more cognitively demanding work.

Now the inconvenient evidence for the doom case. The largest careful study we have — Humlum and Vestergaard’s examination of 25,000 Danish workers across 7,000 workplaces — found AI chatbots had [no significant impact on hours or earnings in any of eleven occupations, with average time savings of about 3 per cent, a fraction of which was reabsorbed by new tasks. The lesson is not that AI does nothing. It is that AI does nothing *by itself*. The technology’s effect on your working life is almost entirely mediated by the organisation deploying it — which is another way of saying AI adoption is a human problem wearing a technology costume.

That is also where the hope lives. The same tools that generate workslop can also absorb real drudgery: the minute-taking, the status-chasing, the first drafts of documents nobody reads carefully enough. AI that summarises the 8 p.m. meeting so a segmenter need not attend it is a boundary technology. AI that makes the 8 p.m. meeting cheaper to convene is a boundary solvent. Identical capability; opposite cultures. The firms treating AI as a headcount-reduction device with a chat interface will get the infinite workday, intensified. The firms treating it as an instrument for redesigning work — deciding deliberately what the reclaimed hours are *for* — will get something rarer: slack, recovery, and the deep work that fragmentation killed.

From balance to flourishing

The most durable idea in my dead whitepaper came from positive psychology. In my work on human *flourishing* at work — the recognition that humans want more than an absence of overwork. They want engagement, relationships, meaning, and achievement, not merely a well-defended weekend. Researchers in the boundary literature have made the same turn: away from “how do we get to balance” and towards “how do we get to flourishing.”

I like this reframe for a reason that has nothing to do with positivity. “Balance” smuggles in a zero-sum assumption — two weights, one scale, every hour for work stolen from life. Flourishing permits the observation that good work is *part* of a good life, and lousy work poisons the other half no matter how firmly you disconnect at five. The priest who finds meaning in the petrol-station confession is not failing at segmentation; the accountant answering midnight emails about nothing is not succeeding at integration.

For organisations, the agenda follows from the evidence rather than the fashion. Stop solving culture problems with compensation — the research consistently finds culture vastly outpredicts pay in driving attrition, and toxic culture has roughly ten times the predictive power of compensation). Stop imposing one boundary style on a workforce that contains both types; boundary fit is cheap to offer and expensive to ignore. Treat after-hours contact as a design flaw, not a loyalty signal. And decide — explicitly, before the tools decide for you — whether AI in your organisation will compress the workday or colonise the evening.

For individuals, the frozen market counsels a colder realism. You may not be able to leave, for now. You can still know which end of the integrator-segmenter spectrum you occupy, negotiate for fit rather than for slogans, and refuse to mistake responsiveness for contribution. Markets thaw. The people who flourish on the other side of this one will be those who spent the freeze deciding what work is for — rather than letting an inbox, or an algorithm, decide for them.

The 2022 draft ended by declaring that Pandora’s box of talent mobility could never be closed. It closed. Boxes do. What stays open is the older question the mobility era briefly let workers ask aloud: *what kind of place is worth my finite attention?* No labour market, however frozen, and no technology, however clever, has yet managed to make that question go away.

---

Sources:

[BLS JOLTS May 2026](https://www.bls.gov/news.release/jolts.nr0.htm);

[Indeed Hiring Lab, May 2026 JOLTS analysis](https://www.hiringlab.org/2026/06/30/may-2026-jolts-report-more-of-the-same/);

[Microsoft Work Trend Index, “Breaking Down the Infinite Workday” (2025)](https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday);

[Gallup, State of the Global Workplace 2025] https://www.gallup.com/workplace/708071/global-employee-engagement-continues-decline.aspx);

[Niederhoffer, Kellerman et al., “AI-Generated Workslop Is Destroying Productivity,” HBR (2025)](https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity);

[Humlum & Vestergaard, “Large Language Models, Small Labor Market Effects,” NBER w33777](https://www.nber.org/system/files/working_papers/w33777/w33777.pdf);

[Kreiner, Hollensbe & Sheep, AMJ (2009)](https://journals.aom.org/doi/10.5465/amj.2009.43669916); Nippert-Eng, *Home and Work* (1996); [Sull, Sull & Zweig, “Toxic Culture Is Driving the Great Resignation,” MIT SMR (2022)](https://sloanreview.mit.edu/article/toxic-culture-is-driving-the-great-resignation/);

[DLA Piper on Australia’s right to disconnect](https://knowledge.dlapiper.com/dlapiperknowledge/globalemploymentlatestdevelopments/2024/the-right-to-disconnect-in-australia-what-we-know-so-far).*

I once added it up. Across a consulting career, I have spent something like 100 weeks — two full working years — building slide decks. Then I found Gamma, and a deck went from a week to half a day.

I want those 100 weeks back. You have your own version of that number. Most knowledge workers do.

You are not short of AI tools. You are drowning in them. What you are short of is systems. Email, meetings, document search, status updates, deck production — they still eat the week because the work itself was never redesigned. Bolting a chatbot onto a broken workflow just gives you a faster broken workflow.

There are 1000s of tools for knowledge workers. Here is my stack, before and after AI.

If I were driving these tools, clicking buttons, juggling tabs, opening apps, logging in, getting permissions right, I would jump off a bridge.

The cheat code for the AI world, is IT DRIVES THE TOOLS. My CRM? Never touch it. My 6000 Obsidian 2nd brain files. I didn’t put them there, and I never open the folders.

The key is orchestration - connecting tools so they do your work, and prep outputs for your very human superpower - judgment and taste.

Watching YouTube or webinars won’t get you there

You are also drowning in amazing content - 23-year-old YouTubers clicking around their demos like dervishes. “My second brain…” “I build this app and generate 80k MRR.”

You watch in wonder, then get back to your job.

So I built a programme to fix the workflow, not sell you more tools.

In three weeks, with 3 in-person “build-days”, you build a personal AI Chief of Staff — one that triages your inbox before it owns your morning, retrieves the right context from your own documents instead of making you hunt for it, and turns reports, decks, and meeting prep into workflows you run rather than chores you dread. You build it on your real work, in live sessions, with me. You leave with something that runs the Monday after we finish.

It is not prompt tricks. It is workflow architecture, judgement, and adoption science - the behavioral science that I have spent fifteen years on.

The first cohort starts 27 July. It is small and it is live, which means seats are genuinely limited — that is a fact of the format, not a marketing line.

If your morning starts with 250 emails and ends with no time to think, this is for you.

https://maven.com/paul-gibbons/the-ai-chief-of-staff-program-build-your-250k-a-year-ea

Paul

This is a warts and all recap of building a firm that is now in its 25th year - and which won just about every award in the leadership and coaching space.

It started on a bus. I had just finished a sabbatical leave and finished my Masters thesis on Spirituality at Work. Waiting for me was an apartment on the Thames, and a partnership at PwC. Living the dream?

Then, and today, change management was predominantly about implementing large ERP systems: years-long change programs. I ran change on one with a billion dollar price tag which made fat stacks for the partnership. Very profitable for a new partner, and very dull for this one.

As I’d just written 60 thousand words on purpose, meaning, and joy at work. My passion was leadership development and coaching, something PwC and the other big firms don’t do. I’d spent most of the last decade going deep with myself. I was a personal development junkie, a spiritual warrior going back to work on technology adoption - valuable work. But for somebody else.

So my firm started there on that bus. I filled that yellow pad with a vision of a consulting firm that brought the latest and best personal development methods to business leaders. The good thing? Nobody was doing that. The bad thing? Nobody was doing that.

Problem was, I had no clients, and I’d spent six months burning through savings. Screw it, I took the paycheck and went back to work.

Just three months later, the dissonance became too great. I jumped ship.

Future Considerations was born 25 years ago, 2001! The first “integral” consulting firm based on the research I’d done on workplace spirituality. I brought in other personal development “technologies” which will mean nothing unless you are deep in the leadership development world. Ontological coaching. Embodiment. Teal. Spiral Dynamics. Holacracy.

In secular terms, could we help leaders bring their whole selves to work, to embody their values, to lead authentically, to care deeply for people as well as profit?

Hello? Anybody listening?

Not for the first year. Yet it was exhilarating. One day building my first website in Front Page. The next working on marketing, the next finance, the next sales.

But for Christmas, everybody got socks. It was a very lean year.

But we eventually grew - a lot. 50 people and $5m in revenue isn’t big in absolute terms, but for a leadership startup five years in? It was pretty good. Our client list was enviable - Amex, Barclays, Bayer, Deutsche Bank, Shell, BP, HSBC, Microsoft, Zappos, Comcast, and Anglo-American.

So what went right and what would I do differently.

The lessons (expensive mistakes)

Cut your expenses - eat Ramen for the first year.

I did the opposite and it was very foolish - I had just got married, and we decided the family of two needed a 4 bedroom house in one of London’s bougiest neighborhoods. I was used to a big salary, not a startup with no cashflow. I burned through savings like a trader on a weekend junket.

Be extremely careful whom you hire.

One toxic person in a 10-person firm is cancer. I had two. They were charming, and interviewed well, but that is a narcissist superpower. Moreover, when it becomes clear, cut them instantly. Yes, we needed the person-power, but culture is the soul and lifeblood of a startup. Kill the culture, kill the firm.

Distribute ownership mindfully and quickly.

In a consulting firm, there is nothing you can do to stop someone quitting, opening up a shop across the road, and taking clients with them. (Sure the law and employment contracts should prevent that - in practice, such clauses are almost unenforceable without very expensive lawyers.)

I was good at people owning the vision, mission, purpose, and culture. But giving them a financial stake is both morally praiseworthy and good business.

The lessons (the things we did well)

Aim for “world-class” from the get go.

There are more than 100,000 coaches in the world and thousands of leadership development firms.

Maybe this impulse is some sub-clinical narcissism, but I’ve always thought if you are going to do something, aim to be the best in the world. (Our first leadership development program, for the UK’s NHS, had 197 applicants.)

So from year three, our competitors became McKinsey, Duke Corporate Education, and INSEAD.

Tough guys to beat - so you have to be at your most innovative. (They are very easy to out innovate, not so easy to out-brand.) So that forces innovation and differentiation upon you.

Pivot fast and hard.

When Future Considerations launched I was running an org change programme at British Airways. The next big job we won was DEI for a London municipality. The next job was building and organisational change programme for FMCG Cadbury Schweppes.

Then the pivot happened. Clients kept asking about leadership development. So far we had done none as a firm, although I had done rather a lot personally.

So we pivoted hard - branding, product development, hiring were all focused on leadership - and by the time I sold the firm, we were 90 percent leadership development consultants and coaches and only very little organisational change.

Message: Strategy is like that - I call it Ying and Yang strategy - one type is the sort of singular focus you hear about in management books - a “male” version if you will. The other type is emergent and listening - “what is the world telling you?” - call that “female” (again if you will.)

Charge like you are the best.

One of the most interesting lessons came from my PwC mentor - long retired. We were charging 900 pounds a day during the first year (about $1500.) He was working for a top-10 British company and I dared asking him how much his day rate was. 3500 pounds ($6000) a day!

I said, “how on earth do you charge that much?”

He said, “there is a lot less competition at 3500 a day than at 900 a day.”

Wow. We doubled prices and didn’t lose a client.

As someone who loves both philosophy and AI, I’m torn between getting my hands greasy building stuff and reflecting on bigger questions — some of which have been with us for thousands of years.

This is a summary of a project I started in January. Nine posts, each taking an area of philosophy and asking: what are the big questions AI raises within it?

These aren’t abstract exercises. Every one of these questions is landing on the desks of CEOs, policymakers, and technologists right now — most of whom have never taken a philosophy course. That’s the gap this series tries to close.

If you’re new here, start wherever your curiosity pulls you. If you’ve been following along, the final piece — meta-ethics — just dropped.

1. Philosophy of Science

Is AI ending the scientific method — or fulfilling it?

What counts as explanation? Can a model that predicts everything but explains nothing still be called science?

Read →

2. Meta-ethics (NEW — the series finale)

What arguments can settle moral disputes about AI?

The ethics beneath the ethics. Before we can debate whether AI should do X, we need to ask: how do we settle moral questions?

Read →

3. Philosophy of Language

Is AI playing Wittgenstein’s language games with us?

Meaning vs. mimicry. When a model produces perfect prose, does it understand — or is it the most sophisticated parrot in history?

Read →

4. Aesthetics

But is it art?

What does aesthetics tell us about creativity, originality, and the role of intention in art — now that machines can paint, compose, and write?

Read →

5. Political Philosophy

Is AI fundamentally undemocratic? Should we care?

Power, consent, and governance. When a handful of companies shape the information diet of billions, what would Rawls, Mill, or Habermas say?

Read →

6. Philosophy of Mind

Are human minds special? Why?

Consciousness, intentionality, and the hard problem. If a machine behaves as if it’s conscious, does the distinction still matter?

Read →

7. Epistemology

AI is shaking the foundations of knowledge

How do we know what we know — and what happens when our most powerful knowledge-producing tool is also our least transparent one?

Read →

8. Metaphysics

What are the two most important topics in metaphysics?

Free will and the nature of reality. AI forces both questions out of the seminar room and into the boardroom.

Read →

9. Philosophy of Technology

What separates humans from machines? Is that separateness an illusion?

From Heidegger to Haraway — the philosophical tradition that takes technology itself as the object of inquiry, not just its consequences.

Read →

This series is free and will stay free. If it made you think, the best thing you can do is share it with someone who’d wrestle with these questions too.

Share this post

Paul Gibbons is the author of 9 books on leadership, change, and AI adoption — including the forthcoming Adopting AI. He advises organizations on AI adoption through his Adaptive Adoption framework.

This is part of a nine-part series on philosophy and AI started in December. This one took so long because it is the hardest. Other topics? Philosophy of Technology, Philosophy of Science, Philosophy of Language, Political Philosophy, Philosophy of Mind.

Open AI ethics at any conference and you will find thirty arguments and no foundation. The image below is the quick and dirty version of AI ethics - each of those is worth a PhD thesis or six.

The original vision for this article was to review a subset of those issues in some detail. That can be found in Adopting AI: The People First Approach.

This sets those momentarily aside. Not because these questions are small — some will define the decade (or the century) — but because ethical conversations share one feature - the one most overlooked by commentators.

Every issue is a fight about a should.

Should a model train on a painter’s work without consent? Should enhanced surveillance be permitted? Should governments have control over who uses the technology and how they use it? Should an agent act with no human in the loop? Should compute on this scale be spent on projects with unclear human benefits? Should the environment be sacrificed in favour of this technology?

The dominant grammar of the field is utilitarian (JS Mill, Jeremy Bentham)— the fight narrows to cui bono. Not whether benefit justifies cost, but whose benefit counts, weighted how, against whose harm.

These are real fights. They are also the wrong altitude.

Two people can agree that the good is what maximises welfare, come to blows over whose welfare, and never ask why welfare is the measure at all. The shared premise escapes examination because it is shared.

That is the trap of applied ethics: getting lost in an issue (albeit a critical one) without thinking about where “shoulds” come from. Those shoulds come from metaethics.

So while we yell “shoulds” at each other, knowing where those come from is more important than ever. What was a fringe academic conference topic takes on new importance today.

As you think through the categories, think also about where you see them in public, so this isn’t purely abstract.

Share

Metaethics

Metaethics asks three questions from above the fray.

• What are we doing when we make an ethical claim?

• And what, if anything, justifies it? Not what is right, but…

• What makes anything right, and how we could ever know?

Now hold that against the machine. We are building systems that make moral choices at scale — what to refuse, what to optimise, whom to weigh — and we train those choices in from data, from preference, from a written constitution. Each method smuggles an answer to the metaethics question. None of them admits it. The industry has taken a metaethical position and named it common sense.

What follows is the map of positions it could have taken — and the one it took without looking.

Command

Quis custodiet ipsos custodes? — “Who guards the guardians themselves?” — Juvenal, Satires VI.

The oldest answer. Right is what the authority commands. Divine command theory: good because God wills it. Its secular heirs say the same with the names changed — good because the sovereign decrees it, the law states it, the policy forbids it. Plato split it open in the Euthyphro and the crack never closed: is the act good because it is commanded, or commanded because it is good?

Take the first and morality is arbitrary. Take the second and the command was never the ground. And this finds its way into contemporary ethics unanswered, is the good a property of the object, or a property of the subject. (If the former, where is it “inside,” what is it made of? If the latter, and it is subjective, then ethical relativism is alive and well.)

Take the first and morality is arbitrary; take the second and the command was never the ground. It survives as the oldest open question in ethics: is good a property of the object, or of the subject? If it sits in the thing itself, where exactly, and made of what? If it sits in us, relativism is not a heresy but the default — and one tribe's atrocity is another's custom.

This is the governance layer. The EU AI Act, the model constitution, the corporate use policy — “thou shalt not” descending from an authority. Most of what is sold as AI governance is command theory wearing a compliance badge.

And command theory cannot say who guards the guardian. (Quis custodiet ipsos custodes?) A rule encoded is a command obeyed; it is not a reason understood. A model that refuses because it was told to refuse has the conscience of a soldier, not the judgement of a judge. The Euthyphro fork is live in every constitution: either the principles are arbitrary, or they point past themselves to a ground the document does not contain — and does not name.

Share

Biology

“The time has come for ethics to be removed temporarily from the hands of the philosophers and biologicized.” — E.O. Wilson, Sociobiology: The New Synthesis (1975).

Ground ethics in nature. We are evolved animals, and our morality is the residue of survival — kin selection, reciprocity, the moral emotions that held the band together. Right is what fitness wrote. This is the naturalist’s hope: that ought reduces to is, that the good is just the adaptive seen from inside.

It surfaces as the appeal to “human values” treated as a fixed biological dataset — read it off the species, align the machine to it. But the residue is mixed. Evolution gave us tribute and tenderness in one packet, the in-group hug and the out-group spear. Which inheritance gets aligned?

The naturalistic fallacy bites here. What evolution selected is not what we owe each other; the disposition is not its own blessing. And the machine is the first moral agent with no evolutionary past at all — no kin, no band, no death to dread. Ground its ethics in our biology and you have aligned it to a story that was never its own. We are reading the machine its inheritance from a will it was not named in.

Emotion

“Reason is, and ought only to be the slave of the passions, and can never pretend to any other office than to serve and obey them.” — David Hume, A Treatise of Human Nature (1739–40).

Reason is the slave of the passions. Morality is felt, not deduced — sympathy, approval, the sentiment of humanity. We do not argue our way to cruelty is wrong; we recoil, and the philosophy comes after to dress the recoil in reasons. Ethics is a structure of feeling, refined.

This is RLHF’s secret metaphysics. Human preference — the thumbs up, the ranked pair, the annotator’s nod — is sentiment captured as signal. The model learns what we approve of, not what is right. It is sentimentalism at industrial scale, and the field does not call it that because the field has not noticed it chose it.

Here is Hume’s guillotine, the cut the whole industry has stepped over. You cannot derive an ought from an is — and a dataset is nothing but an enormous is. Preference data records what humans wanted; it is silent on what humans should want. You cannot derive a value from a dataset. Every pipeline that learns values from feedback is trying to scale Hume’s wall on a ladder built from the one material he proved too short. The sentiment is real. The justification never arrives.

Utility

“It is the greatest happiness of the greatest number that is the measure of right and wrong.” — Jeremy Bentham, A Fragment on Government (1776).

The calculus. Right is what maximises welfare — the greatest good, summed across all who count. Bentham’s arithmetic, Mill’s refinement. Clean, quantifiable, the native tongue of engineers and economists, and for that reason the default setting of the room.

It is the dominant grammar. “Beneficial AI.” “Maximise helpfulness.” Cost and benefit weighed on a planetary scale. Optimisation is utilitarianism with a loss function; when the field reasons about ethics, it reasons in expected value unless told otherwise.

But the fight is never whether benefit justifies cost. It is cui bono — whose benefit, weighted how, against whose harm. The calculus presents itself as neutral and is nothing of the kind. Every sum hides a prior choice about who enters the equation and at what exchange rate. Add up “human welfare” and you have already ruled that the displaced worker and the shareholder weigh the same — or that they do not. The model optimises a number. Someone chose the number. That someone was doing metaethics and filing it under product.

Duty

“Act only according to that maxim whereby you can at the same time will that it should become a universal law.” — Immanuel Kant, Grundlegung zur Metaphysik der Sitten (Groundwork, 1785).

Kant. Right is duty, not outcome. Act only on a maxim you could will universal; treat persons as ends, never as means alone. The rule binds whatever follows from it — the lie is wrong even when it saves.

This is where most enterprise AI lives. Do not discriminate. Be transparent. Do not breach privacy. Constitutional AI is Kant in code: a model holding its outputs against a list of principles it may not break. Auditable, legible, friendly to the lawyers — which is most of why it won.

Duty breaks where duties collide. Privacy against transparency. Non-discrimination against accuracy. The rule-based system has no metalanguage for the conflict; it can carry two commandments and not tell you which one yields. And it assumes the rules are justified, which hands us back to command — who wrote the duties, on what ground? A model that keeps every rule and misses the point is not moral. It is obedient. The field keeps mistaking the second for the first.

Contract

“[Principles that] free and rational persons concerned to further their own interests would accept in an initial position of equality.” — John Rawls, A Theory of Justice (1971)

“Solidarity, not objectivity.” — Richard Rorty, “Solidarity or Objectivity?” (1985).

In Rawls, right is what rational, disinterested people would agree to. Put them behind a veil of ignorance — none knowing whether they will be rich or poor, gifted or disabled, born here or there — and the principles they would choose are just by construction. It is an elegant machine for manufacturing fairness out of self-interest. But it presumes two things it never quite earns. First, a view from nowhere: the veil is a god’s-eye neutrality, a standpoint stripped of every particular, and no actual human — and no actual model — has ever stood there. Second, that even granted the veil, the deliberative processes behind it converge on the good. Rawls trusts that disinterested reason, talking to itself, arrives somewhere just. That is a hope wearing the costume of a proof.

Rorty pulls the floor out — what is right is what we can democratically agree to. That is, solidarity trumps ethical objectivity.

He is not claiming democratic processes approach an ideal — he was far too sly to think procedures crawl toward Truth. He is claiming there is no ideal to approach. The good is not discovered, behind a veil or anywhere else; it is what we can agree to, and nothing stands above the agreement to grade it. This is the hard opposite of divine command. Command sent the good down from an authority above us; Rorty has it rise from among us, with no ceiling but solidarity — “what your contemporaries let you get away with,” made a virtue rather than a scandal.

In AI this is the most flattering story the field tells about itself. Coherent extrapolated volition — align the machine to what humanity would want “if we knew more, thought faster, were more the people we wished we were” — is the veil of ignorance rebuilt as an optimisation target. Social-choice alignment, democratic fine-tuning, a constitution drafted by public deliberation: all of it is contractualism in code, the hunt for principles everyone could accept.

But the veil cannot be built. There is no view from nowhere to optimise toward; what ships is the contingent agreement of whoever was actually in the room — the rater pool, the lab, the jurisdiction — universalised and stamped “humanity.” And Rawls’s second hope fails exactly where it matters: aggregated agreement converges on the comfortable and the many, not the right; a veil drawn over a biased crowd only launders the bias. Rorty is the more honest of the two — it is only us, agreeing, with no cosmic warrant — but that honesty is what alignment cannot afford to admit, because it dissolves the claim to universality the whole project leans on. You did not align the model to the good. You aligned it to a vote, and hid the franchise.

Virtue

“Moral virtue comes about as a result of habit.” — Aristotle, Nicomachean Ethics II.1.

The question is not what should I do but what kind of agent should I be. Ethics is character — the settled dispositions that make a life go well. Not rules, not sums, but phronesis: the trained judgement that sees what this situation, this once, is asking.

It is the least operationalised position and the one the field needs most. You cannot write practical wisdom into a constitution. The interesting alignment failures are failures of judgement, not rule-breaking — the model that satisfies every constraint and still does the thing no wise person would. Compliant to the letter, foolish in the act.

Virtue ethics demands a thick account of human flourishing, and the field has none. It has metrics. It can say what passes; it cannot say what excellence is. And virtue is earned by habituation — a history of choosing well — while the model has no history it lived, only a corpus it absorbed. It can reproduce the output of wisdom without ever becoming wise. That gap, between the look of judgement and the having of it, is the one no benchmark measures.

Critique

“Wrong life cannot be lived rightly.” (Es gibt kein richtiges Leben im falschen.) — Theodor Adorno, Minima Moralia (1951).

Horkheimer, Adorno claim the game is rigged, and the individual cannot win. They suggest that norms that are dressed as neutral merely encode domination. The fact-value split is not innocent — it is a dominance ideology with a rational veneer. Critique works from the inside: hold a society’s professed ideals against its actual practice and name the gap.

Adorno’s enigmatic phrase means that in a false society — capitalist, commodified, “administered” — no individual can assemble a genuinely good life inside it.

Private virtue doesn’t give you an exit. Your decent salary is paid by the same machine, your ethical consumption still runs on exploited labour, your inner authenticity is still shaped by the culture industry. There’s no clean corner of a dirty whole. It’s a holist, anti-individualist claim: you can’t moralise your way out of a system by personal effort, because the system constitutes the self doing the effort.

An AI trained on the world inherits the world’s injustice as ground truth. Align it to “human values” drawn from a corrupt order and you have automated the corruption — value lock-in, the unjust pattern frozen and scaled. The “aligned” system can be aligned, without a flaw, to a wrong.

This is alignment’s buried problem, and it is the oldest one in this essay. If every norm the machine could learn is drawn from a flawed order, on what ground do we call any alignment good? Critical theory met this exact wound — the normative deficit: if all norms are ideological, what licenses critique at all? Its answer was not a truth handed down from above — that is command theory in a lab coat — but the contradiction between what a system promises and what it does, made visible and pressed. That is the only purchase available. And the machine cannot reach it: it has no standpoint outside the data. We do. For now.

If you are thinking “what a mess,” so am I! Ethics is hard, who knew?

Wicked and social messes

If that weren’t complex enough, there are, of course, each of the applied ethical issues that can be approached through each meta-ethical lens— and the lenses rarely agree. These are wicked problems in Rittel and Webber's sense, social "messes" in Ackoff's: no clean formulation, no stopping rule, no view from nowhere.

And if that weren’t complex enough, there are social polarities.

Not everyone who is a utilitarian thinks the same way. Not all utilitarians think alike, and aggregative thinking is constitutively distribution-blind — it counts the size of the utility pie, not who eats. Cui bono? Who benefits from a datacentre, and what long-term tradeoffs — the environment — justify the economic sugar high of building one against the perceived stakes of global AI leadership? Prioritarianism and Rawls offer repairs, but they relocate the argument rather than settle it.

Ethical conversations between groups with different metaethical lenses may defy resolutions, and those differences are often handled through politics. Handled but not reconciled.

In countries with strong “divine command” constituencies, the United States and the Middle East, certain issues have defied harmonious resolution for (sometimes) centuries.

And we muddle the categories — treating contested values as plain facts, or facts as if they compelled an ought. No quantity of is entails an ought: you cannot derive a value from a dataset.

The alignment problem inherits all of it. "Aligned" smuggles the question metaethics hasn't settled — aligned to whom, under which lens, with whose welfare weighted how. That isn't a technical detail the engineering will mop up. It is the problem.

Metaethical blindness

Those are eight meta-ethical stances in brief. And a close reader will have seen each has tentacles into metaphysics, epistemology, and political philosophy.

The industry occupies several at once and confesses to none. RLHF is sentimentalist. The constitution is deontological. The optimisation is utilitarian. The governance is command. The “democratic inputs” are contractualist. They contradict one another, and the contradiction runs unmanaged because it runs unseen.

This was the ninth piece. Across the eight before it, one pattern held: AI does not escape philosophy’s hard problems. It industrialises them. The is–ought gap was a seminar puzzle for two centuries. We have built a hundred-billion-dollar industry on the wrong side of it. Naming that does not close the gap. It is the price of admission to standing anywhere at all while we try.

Coda — The Strongman to the Rescue

And what happens if we don’t find answers?

The metaethical silence is not only an intellectual embarrassment. Unanswered, it is political Any order — a democracy or a deployed model — that cannot justify its values, only assert them, slowly forfeits the consent of the governed. People tolerate an authority they believe is grounded in something. They withdraw from one they suspect is grounded in nothing but whoever held the pen.

Rorty saw where that withdrawal leads. Writing in 1998, in Achieving Our Country, long before it was fashionable, he warned that once enough people decide the system has failed them, they “will start looking around for a strongman to vote for” — someone to promise that the smug experts will no longer be calling the shots.

He meant electoral democracy. The warning is an apt description of 21st century politics An AI order that aligns itself to a laundered consensus and stamps it humanity manufactures the same legitimacy deficit — at machine scale and machine speed. And when the governed discover the franchise was hidden, they do not ask for a better metaethics. They ask for a strongman.

That is the cost of the unpaid invoice. The is–ought gap was a seminar puzzle for two centuries. The strongman is what waits on the other side of pretending it is closed.

A question for the comments: when your organisation says “responsible AI,” which of the eight is it secretly running on — and does it know?

This is Part IX, the last of the AI and Philosophy series — eight pieces before it, each taking one branch of philosophy and showing where AI inherits its unsolved problem. Start the series [here]. Subscribe and the next arc lands in your inbox.

You weren’t unlucky, you were bad, Paul

On the seventh day of the World Series of Poker Main Event, thirty-seven players remained from a starting field of more than nine thousand. Everyone still in a seat was playing for a first prize of ten million dollars. Then I looked down at pocket kings — the second-best hand in poker — got the chips in as a 70-30 favourite, and lost.

My kids, my friends, the fans, and the TV commentators all commiserated. “Bad beat.” Nothing you could do.

The truth only I knew was ugly. For the better part of a day I had been fatigued and playing scared money — bleeding chips rather than haemorrhaging them, but bleeding all the same. By the time the fatal hand arrived, I had been quietly heading for the exit for hours.

That is the lesson business leaders most need to learn: the visible defeat makes the headlines, not the five hundred small decisions that preceded it. In an AI programme those decisions are the unglamorous ones — vague ownership, weak evaluation, governance deferred, adoption performed rather than embedded, talent treated as an afterthought. It looks like bad luck — a disruptor who came from nowhere, a model that failed in front of the client — but the seeds of defeat were sown long before the disruptive-competitor breached your imagined moat.

The AI paradox - mind, machine, markets

Strategy is fog, friction, chance. (Adapted from On War)

In mid-2026, artificial intelligence presents a paradox. The models have outrun every forecast and every timeline. Yet. Enterprise deployments have largely disappointed. According to the Atlanta Fed, 71% of firms now use generative AI; yet most, Fortune reports, see no measurable profit from it.

Individuals routinely report doubling their output. Most organisations are fighting for a few percent.

Why the violent disconnect?

Understanding that paradox — and rethinking strategy around it — is the most important task facing executives today.

It is tempting to read the gap as a technology problem: the wrong model, the wrong vendor, the wrong stack. It is not. The models are extraordinary, and growing cheaper and more capable by the quarter.

The gap is one of judgement — the organisational inability to turn intelligence into disciplined action under uncertainty.

So why poker?

We have been reaching for the wrong metaphor. Executives discuss AI as though it were chess: perfect information, fixed rules, a single best move waiting to be calculated. But the world AI has created is the opposite. It is hidden information, asymmetric incentives, emotional pressure, noisy feedback, and expensive bets placed before the evidence is in.

That is not chess; it is poker.

Poker is game theory, decision-making-under-uncertainty, variance, expected-value, and a delicate balance between human psychology and mathematics.

And when individuals play for tens of millions of dollars, the pressure, fatigue, hyper-focusing while remaining relaxed, cognitive biases, and cognitive atrophy are factors the poker professional has to master.

It is a lab. Just as the contact between AI and humankind is a lab - poker players update their priors constantly. Is humanity?

Smart is just the ante

I have spent much of my career in rooms where intelligent people make expensive decisions with incomplete information — boardrooms, trading floors, transformation programmes. And, for the last several years, at the World Series of Poker. The longer I watched AI strategy unfold, the more familiar the failure pattern became. It was the same one I had watched undo good poker players: confusing the quality of a decision with the quality of its outcome.

Poker has changed almost beyond recognition. The bravado and soul-reads of televised theatre are largely gone. The best players now train their minds the way athletes train their bodies — game theory, probability, risk psychology, bankroll discipline, emotional control, and a ruthless review of their own mistakes. Those are precisely the habits AI strategy now demands, and precisely the ones most organisations lack.

Polymath Poker is my attempt to give leaders that vocabulary. It translates the working language of elite poker — expected value, Bayesian updating, range thinking, risk of ruin, decision quality — into a practical way of thinking about AI, markets, and human capital.

The premise is simple, and I think consequential: intelligence is only the ante. Edge belongs to those who build a harness around intelligence — whether that intelligence is human or machine. The harness is not a restrictive cage. It is what lets force become work.

Read Polymath Poker — the WSOP Pre-Release Edition is out now.

The uncomfortable claims

If the harness were the whole argument, the book would be a comfortable read. It is not meant to be.

It dispels the idea that you can rent innovation through SaaS AI — that buying access to a model is the same as building an advantage. It takes apart the myth of the moat: in an environment where capability cheapens every quarter, a stack is not an edge. And it shows how to call consultants’ bluffs — the confident, unfalsifiable advice that fills the vacuum where judgement should be.

It also reads poker’s staking economy as a mirror for the capital structure of AI: circular flows of capital that may be risk concentration masquerading as diversification. And it argues that leaders operating in genuine volatility should trade short-term optimality for longer-term optionality — a distinction the optimisation-obsessed culture of modern management routinely gets wrong.

Introduction · The Hand Everyone Remembers — The disaster you remember was seeded by five hundred quiet decisions nobody saw — and so is every failed AI programme.

1 · The Poker Laboratory — Why your early AI wins may be luck wearing the mask of skill, and how small samples seduce smart people into bad causal stories.

2 · Smart Is the Ante — Intelligence only buys a seat at the table; performance comes from the harness — workflow, trust, governance, incentives — built around it.

3 · Stop Using McKinsey’s Vocabulary — The poker words — expected value, range thinking, Bayesian updating — that sharpen strategy exactly where consulting jargon goes blunt.

4 · The Game Has Changed — When tooling and transparency rewrite the rules, expertise decays faster than insiders admit — as poker discovered, and AI-exposed industries are discovering now.

5 · You Have No Moat — In a fast-learning domain most moats are merely delay mechanisms; the only durable edge left is your rate of learning.

6 · The Staking Economy and the Capital Structure of AI — When incentives and exposures correlate, diversification quietly becomes concentration — and AI’s circular capital flows may be exactly that.

7 · Calling Consultants’ Bluffs — Strategy decks are signalling games; the real discipline is forcing the hidden assumptions into the open.

Closing · The Adaptive Fold — The hardest and most valuable move in poker, and in AI strategy, is laying down a strong hand the moment the table turns against you.

The field edition

I am releasing this first as a WSOP Pre-Release Edition — a limited run timed to coincide with the 2026 World Series of Poker, ahead of the formal launch on 1 August. That is deliberate. This is the version closest to the table: the argument tested in the live environment that inspired it.

I am not a tourist at that table. I have over $1 million in live tournament earnings, including a 37th-place finish in the 2025 WSOP Main Event — alongside three decades as a strategist, a former IBM partner, a former derivatives trader, and the author of eight books on change, leadership, and human capital. Polymath Poker is where those two lives finally meet.

The adaptive fold

The hardest discipline in poker is not the hero call. It is the fold — laying down a strong hand when the situation has turned against you, even when everyone is watching, even when your ego wants the pot. The organisational equivalent — the adaptive fold — may be the single most valuable and least practised skill in AI strategy.

The AI era will not reward the smartest organisation. Intelligence is becoming a commodity. It will reward the one that can act on intelligence under pressure, review its own decisions honestly, and fold when the table changes.

Smart is only the ante. The edge is the harness.

Get the book — Polymath Poker, WSOP Pre-Release Edition

Book me to speak (via BrightSight)

Subscribe for more on AI strategy, judgement under uncertainty, and the polymath life.

The WSOP Pre-Release Edition is out. The formal launch follows on 1 August 2026.

The TL;DR

AI does not make leaders irrational. It gives their irrationality a larger budget, and a faster clock.

The most expensive AI failures of the next three years will not come from poor technical choices. They will come from a single decision error: leaders relying on the inside view — what we plan to do, what our team is capable of, what our timeline assumes — when the outside view would tell them, in five minutes, that 80% of comparable projects fail.

Daniel Kahneman called this the most consequential bias in human judgement. Bent Flyvbjerg has spent thirty years documenting its cost across 16,000 major projects. The fix has a name — reference class forecasting. The 2023–2026 AI cohort gives us, for the first time, a reference class that boards can actually use.

Most are not using it.

What follows is an examination of four boardroom biases that AI has industrialised. They are the usual suspects: sunk cost, planning fallacy, overconfidence, status quo. But the unifying lesson is not that boards need to memorise more biases. It is that they need to stop reasoning from the inside.

Part I: Biases that add risk and kill ROI

What follows is a 2026 rework of a 2015 article: Boardroom Biases Kill Strategy. The exam question - what lessons does the past teach? Which learnings no longer apply?

The fish rots from the head

In 1996, FoxMeyer was the fourth-largest pharmaceutical distributor in the U.S., a $5 billion company. They bet their future on a massive ERP (Enterprise Resource Planning) transformation—an SAP implementation designed to revolutionize their supply chain. It was supposed to be the “technological edge” that would crush their competition.

Instead, it crushed them. Within a year of the “go-live,” FoxMeyer filed for bankruptcy and was sold off for pennies on the dollar.

Humans make mistakes and enterprises fail, but some of those mistakes are systematic errors, that is, non-random and directionally “biased.” (Projects never undershoot on time, cost, and quality by a factor of two; they sure do overshoot, though.)

Understanding and managing boardroom biases, your own, dear leader, and your team’s, is an essential and undiscussed leadership skill.

Potted history of cognitive biases in boards

Starting in the 1950s, thinkers like Herbert Simon began to wonder how humans really made decisions and why they deviated from rationality so much. Soon after, Kahneman and Tversky began their partnership that resulted in a Nobel - again, the question they turned their mind to was “what are we systematically and predictably doing wrong?” By the time the 21st century was rolling, we had this:

Don’t worry that you can’t read it; it is substantially accurate and completely useless. What are leaders supposed to do? Memorize all hundred of them?

When I started writing The Science of Organizational Change, that was the state of play — a long list of impossible-to-use biases and fallacies. But I was running a leadership consulting firm and felt this OUGHT to be useful if we could distill it down and provide guidance on what leaders should do about a far smaller number.

What follows are the four biases that, in my experience, do the most damage in boardrooms when making technology investment decisions. Part II will cover the social and group biases — groupthink, authority bias, the halo effect — the ones that corrupt how boards interact. This piece is about the ones that corrupt how boards think.

The “rational” boardroom myth

The core message was that directors are not rational actors; they are, as Dan Ariely put it, “predictably irrational.” That message landed with some discomfort and a lot of nodding.

Today, the stakes are considerably higher. Boards are approving AI capital expenditure at unprecedented speed, often with less technical understanding than they had during the ERP era. We assume billion-dollar decisions are made with spreadsheets and logic. In reality, they are driven by what Keynes called “animal spirits” — confidence, fear, herd behavior, and the intoxicating narrative that this time, the technology really will deliver on the pitch deck.

The cost of getting this wrong is not abstract. McKinsey’s research found that 17% of large IT projects go so badly they threaten the very existence of the company. Not “underperform.” Not “come in over budget.” Threaten the existence of the company. When you sit in a boardroom approving a nine-figure AI transformation, that number should be written on the wall.

Bias #1: The sunk cost fallacy

The sunk cost fallacy — sometimes called escalation of commitment — is the tendency to justify increased investment in a failing endeavour based on what has already been spent. “We’ve come this far” is the sentence that destroys shareholder value.

The textbook case is Nick Leeson and Barings Bank. Leeson doubled down on losing derivative positions to hide earlier losses, eventually accumulating $1.3 billion in exposure that brought down the oldest merchant bank in Britain. The psychology that drove Leeson is the same psychology that drives every executive who refuses to pull the plug on a failing implementation.

Now apply that to AI. A company buys expensive enterprise AI licenses — say, Copilot at scale — or sinks eighteen months into building a custom LLM. Six months in, adoption is low, the use cases haven’t materialized, and the data quality problems are worse than anyone admitted. The rational response is to pivot or cut. The human response is to double the budget, hire consultants, and launch a “change management initiative” to fix the adoption problem. Because we’ve already spent the budget.

The worst part is sunk cost reasoning can be thought of as virtuous: “When the going get tough, the tough get going.”

That is sometimes an indication of good character - but not without a rational evaluation that starts with - is the cost-benefit analysis from here rational. The most storied example from business is Andy Grove and XX Moore thinking about Intel’s memory business - they asked themselves, so the legend goes, “if we parachuted into this company with no history or emotional attachment to being ‘the memory company,’ what would we do? We would get out. And so they did. (If they had a crystal ball that could peer into 2026, they might have concluded differently!)

The antidote is simple in theory and brutal in practice: make it psychologically safe to kill projects and admit we were wrong.. If every cancelled initiative is treated as a failure of the person who championed it, no one will ever cancel anything. The sunk cost fallacy doesn’t survive in cultures where pivoting is rewarded, but it thrives in cultures where admitting a mistake is career-ending.

Bias #2: The planning fallacy

The planning fallacy is the systematic tendency for project plans to undershoot actual costs and timelines. Not by a little. Not occasionally. Systematically, and by a lot.

The research here is unambiguous. Large IT projects run an average of 45% over budget and deliver 56% less value than predicted. That isn’t bad luck. That’s a structural feature of how humans estimate complex undertakings. We anchor on the best-case scenario and treat it as the most likely scenario. We plan for the world where nothing goes wrong, then act surprised when things go wrong.

The canonical example is Denver International Airport’s automated baggage system. It was supposed to revolutionize airport logistics. It ended up $3 billion over budget, years late, and was eventually ripped out entirely. The engineers weren’t incompetent. The technology wasn’t impossible. The planning was delusional — and nobody in the approval chain had the incentive or the courage to say so.

Bias #3: Overconfidence

Overconfidence — hubris, if you prefer the Greek — is the gap between confidence and competence. It is the most dangerous bias in a boardroom because it is the one most correlated with seniority. The higher you rise, the more your confidence is reinforced and the less your competence is tested.

Dick Smith Electronics, the Australian retailer, is a case study. The board believed they could move massive inventory volumes despite clear market signals that consumer electronics retail was contracting. They weren’t stupid. They were overconfident — and overconfidence made the disconfirming evidence invisible.

In AI, overconfidence takes a specific and measurable form.

About a quarter of Fortune 500 companies now have a Chief AI Officer — up from one in ten two years ago. But Heidrick & Struggles finds that nearly half of those titles are simply reclassified existing roles — the CIO or CTO with new business cards. Strip those out, and only about one in eight Fortune 500 companies has a dedicated, net-new AI leader. That's the answer to the question of whether the C-suite has caught up. It hasn't.

This is not a skills gap. It is a structural vulnerability. A board that cannot interrogate an AI strategy is a board that will approve whatever the most persuasive presenter recommends.

And the definition matters. 'AI expertise' on a skills matrix can mean anything from 'sat on a generative AI panel at Davos' to 'ran a $200m ML platform.' If you set the bar at the latter — operating experience that lets you ask the second and third question, not just the first — the number collapses. The Conference Board's S&P 500 data puts disclosed AI expertise at 2.7% of directors. Deloitte's global survey finds two-thirds of directors say their boards have 'limited to no knowledge or experience' with AI. The figures most often quoted in the 10–15% range sit between these, and depend almost entirely on whether the researcher counted self-disclosure or independently verified track record. Whichever number you choose, the conclusion is the same: the overwhelming majority of boards approving AI strategies cannot distinguish a credible plan from a slick deck.

The corrective is not “get smarter about AI” — that’s a years-long journey. The corrective is epistemic humility: know what you don’t know, and build decision processes that compensate. Pre-mortems, red teams, independent technical review, and the discipline to ask “what would have to be true for this to fail?” before asking “how do we make this succeed?”

Bias #4: Status quo bias

Status quo bias is the preference for things to remain as they are. It is not laziness. It is a deep psychological preference for the known over the unknown, reinforced by loss aversion — the well-documented finding that humans feel losses roughly twice as intensely as equivalent gains.

Kodak invented the digital camera. They held the patents. They had the engineering talent. And they chose not to cannibalise their film profits, because the film business was profitable now and digital was uncertain later. By the time “later” arrived, it was too late. Kodak filed for bankruptcy in 2012.

The AI version of this is playing out in boardrooms right now. The idea that having been successful for 25 years gives you a leg up today is a cognitive mistake I hear over and over. Past results are weak predictors of future success. In fact, the world owes Clayton Christiansen a debt - because he raised the counter-intuitive idea that past success might be a trap.

This happens structurally - the company “stack” (structures, systems, mindset) is perfectly synced with a world that no longer exists.

In the “intelligence transition” — a fundamental shift in how knowledge work is produced, evaluated, and valued. The last time something comparable happened was the internet, and the companies that treated the internet as “just a new channel” are mostly gone.

Status quo bias is particularly insidious because it doesn’t feel like a decision. Not acting feels like not deciding. But in a rapidly shifting environment, not acting is a decision — it is a decision to let competitors, market shifts, and technological change determine your future for you.

The solution: the outside view

So what do you do? You cannot eliminate these biases — they are features of human cognition, not bugs in your particular board. But you can build processes that compensate.

The most powerful tool I know is Reference Class Forecasting. The logic is elegant: instead of asking “how long will our project take?” — which invites every bias on this list — you ask “how long did similar projects take?”

If you buy Flyvbjerg’s well researched conclusions, 8.5% of IT projects come on on time and on budget. That is pre-AI research - but ask yourself, “is this easier or harder?”

Three years into the gen AI cycle, the answer is bracing. RAND finds that 80% of AI projects fail — twice the rate of non-AI IT projects. McKinsey finds that only 5.5% of organizations are seeing real financial returns at the enterprise level. BCG finds that the median company has been at this for 24 to 30 months and shows no meaningful EBIT impact. That figure — 5.5% — is within rounding distance of Bent Flyvbjerg's 8.5% base rate for any major project coming in on time and on budget, drawn from his 16,000-project database.

In other words: the AI cohort is so far behaving exactly like every other major project class in history. Worse, on some measures. The inside view says this time is different. The outside view says it isn't.

The hard news: fewer than 10% have fully scaled AI in any single business function. That second number has not meaningfully moved in three years — not when GPT-4 launched, not when reasoning models arrived, not when agents went mainstream. The technology has improved on every dimension we can measure. The organizational ability to deploy it at scale has improved on almost none. Reference class forecasting tells us the binding constraint is not the model. It is the board, the C-suite, and the change capability beneath them.

That question forces you out of the “inside view” (your narrative, your optimism, your special circumstances) and into the “outside view” (the base rate for projects like yours). In a study of Hong Kong road projects, Reference Class Forecasting determined that a 44% budget uplift was necessary to reach 80% certainty of completion within budget. Forty-four percent. That is the gap between how humans plan and how projects actually unfold.

Combined with pre-mortems — “assume this project has failed; write the post-mortem now” — Reference Class Forecasting gives boards a fighting chance against the biases described above. It doesn’t make you rational. It makes you less predictably irrational, which in a world of nine-figure AI bets is worth quite a lot.

What’s next

We cannot eliminate biases, but we can build processes to mitigate them. This piece covered the four that corrupt how boards think — sunk costs, planning fallacy, overconfidence, and status quo bias.

Part II covers the ones that corrupt how boards interact: Groupthink, Authority Bias, and the Halo Effect. These are the “culture killers” — the social dynamics that cause smart people in a room together to agree on terrible ideas, defer to the loudest voice, and mistake charisma for competence.

Part II: The “Culture Killers” (Social & Group Biases) is coming next.

A focus on how the board interacts—Groupthink, Authority Bias, and the Halo Effect. This is about people and governance.

Further reading:

• “Biases in the Boardroom” — Gibbons, P. (2016). Presentation to University of Denver Daniels School of Business faculty

• Keynes, J.M. (1936). The General Theory of Employment, Interest and Money. “Animal spirits” as the driver of economic decisions beyond rational calculation.

• Bloch, M., Blumberg, S., & Laartz, J. (2012). “Delivering Large-Scale IT Projects on Time, on Budget, and on Value.” McKinsey Digital.

• Staw, B. (1976). “Knee-deep in the Big Muddy: A Study of Escalating Commitment to a Chosen Course of Action.” Organizational Behavior and Human Performance.

• Leeson, N. (1996). Rogue Trader. The Barings Bank collapse — $1.3 billion in losses from escalating derivative positions.

• Gibbons, P. (2015). The Science of Organizational Change. Theater ticket analogy for sunk cost reasoning.

• Kahneman, D. & Tversky, A. (1979). “Intuitive Prediction: Biases and Corrective Procedures.” TIMS Studies in Management Science.

• Flyvbjerg, B. & Budzier, A. (2011). “Why Your IT Project May Be Riskier Than You Think.” Harvard Business Review.

• de Neufville, R. (1994). “The Baggage System at Denver: Prospects and Lessons.” Journal of Air Transport Management.

• ASIC v Dick Smith Holdings (2018). Australian Securities and Investments Commission findings on inventory management failures.

• Cheatham, B., Javanmardian, K., & Samandari, H. (2019). “Confronting the Risks of Artificial Intelligence.” McKinsey Quarterly.

• Lucas, H. & Goh, J. (2009). “Disruptive Technology: How Kodak Missed the Digital Photography Revolution.” Journal of Strategic Information Systems.

• Gibbons, P. (2026). “Intelligence transition” — the framing that AI represents a shift in cognitive infrastructure, not a software deployment.

• Flyvbjerg, B. (2006). “From Nobel Prize to Project Management: Getting Risks Right.” Project Management Journal.

The old productivity question was: How do I get more done? The new question is: How do I stop possibility from consuming my life?

I wake at 3 a.m. Not the drugged half-wakefulness of disturbed sleep. Ready to play chess or run a mile. I check in with myself. Anxiety? No. A head buzzing with creative ideas — a research project, a Substack post, the onboarding plan for two new interns. I try to sleep. It is impossible. Full drafts of research papers scroll through my head.

I am not the only one. The AI-builder community talks openly about thirty-six-hour build sessions. Sam Altman, the CEO of OpenAI, has publicly announced that he is moving to polyphasic sleep — short cycles through the day rather than one nightly block — because, in his own words, “GPT-5.5 in Codex is so good that I can’t afford to be sleeping for such long stretches and miss out on working.” He posted that line in the same breath as his own prediction that “post-AGI, no one is going to work and the economy is going to collapse.”

Bryan Johnson — the longevity entrepreneur whose Blueprint protocol is built on the most famously rigid daily routine in technology — has reported, in early May 2026, that he fed roughly 1.5 billion data points from five years of his own health tracking into a custom Claude system, creating a real-time digital twin of his biology. The result derailed him. He broke his own screens-off rule, fell behind on other work, lost sleep, declined social engagements, and his partner Kate Tolo, the co-founder of Blueprint, remarked publicly that she had “lost her boyfriend to Claude.” Johnson’s own description, posted publicly, was unguarded:

“AI is preposterous, as close to magic as I’ve experienced, except a seed becoming a tree and a zygote becoming a baby.”

The most disciplined health optimiser on the planet is reaching, in the same sentence, for the language of botanical growth and human birth — the two strongest metaphors for generation our species has.

This is not the loneliness of two famous men. The same testimony arrives from across the AI-builder community, and the dominant note is not anxiety. It is exhilaration. On a recent episode of The AI Daily Brief titled “Why Agents Make Every Job a Startup,” the host catalogued the same phenomenon and named it for what it is: in the agentic era, “we’re all finding ourselves not clocking out at 3 p.m. on Thursday, but forcing ourselves to go to bed and turn off at 3 a.m.” The voices he assembled are not in crisis.

Aaron Levie, the chief executive of Box, captured it cleanly: “Sorry to anyone who thought AI would mean we’d work less, at least for now. AI makes it easy to explore more than you did before, and so you start doing far more as a result.” These are not the words of people in pathology. They are the words of people who have found a tool that compounds their work session after session, and who cannot easily stop.

Is it psychosis?

Some clinicians, surveying a small but growing literature of case reports, have begun to use the term AI psychosis — a phrase that is not yet a formal diagnosis but is being used to describe delusional thinking that emerges, intensifies, or organises itself around heavy chatbot use (Morrin et al., JMIR Mental Health, 2025; UCSF, January 2026). The peer-reviewed literature is careful. It explicitly notes that there is no epidemiological evidence of a causal or widespread link, and that the term is shorthand used by clinicians and researchers rather than a formal disorder. The press has been less careful.

For me, AI is the greatest creative unlock of my adult life. I have never experienced a creature who can have PhD level conversations on dozens of topics (as could AI 2025) AND at the same time produce useful work at 3-10x normal, pre-AI, speeds (which had to await AI 2026.)

But perhaps I’m deluded along with other thousands of creatives who report the same. Afterall, thousands of crazies doesn’t necessarily make me less crazy.

As a psychologist, I also know the diagnostic asymmetry at the heart of every mental illness: people in the grip of a disorder rarely say, I think I’m delusional. So the question of what is happening here has to be asked with care, neither dismissing it nor running ahead of the evidence.

Nathaniel Whittemore, on a separate episode of The AI Daily Brief titled “Is AI Doom Going Out of Style?”, has reached for a much better name. He calls the phenomenon a “Kierkegaardian dizziness of freedom” — borrowing the Danish philosopher’s 1844 term for the existential vertigo that arrives when human possibility outruns human direction. “You are creating something from nothing,”

Whittemore puts it, “birthing purpose and meaning and value into the world. And yet of course, at the same time, there is an almost Kierkegaardian dizziness of freedom. When you can do anything, when you can work on anything, what are you supposed to do?”

That, to my reading, is much closer to what is happening at 3 a.m. than anything in the psychiatric literature. The cork has come out of a bottle the species has carried for sixty thousand years, and an entire generation of knowledge workers is responding with a vertigo Kierkegaard described in 1844 and that The AI Daily Brief has recognised in 2026.

“The same cognitive phenomenon receives two entirely different verdicts. The question worth asking is: who gets to say which is which?”

Share

Who gets to say?

If a writer stays up until 4 a.m. drafting a chapter, we call her driven. If a painter abandons sleep for three days because a vision will not let him go, we call him possessed by the work. If a founder pulls successive thirty-six-hour build sessions, we call him committed. But if someone says that AI has unlocked an idea-stream that will not let them sleep, the cultural reflex — though, notably, not the peer-reviewed literature — is to reach for the word psychosis. The same cognitive phenomenon receives two entirely different verdicts. The question worth asking is: who gets to say which is which?

Let me be direct, as a psychologist. The application of psychosis to productive 3 a.m. wakefulness — the kind that ends with three usable drafts and a contented return to bed — is a category error. Psychosis, as the clinical literature uses the term, refers to a loss of reality testing: disorganised thinking, hallucinations, delusional belief, marked functional impairment. The cases catalogued in the peer-reviewed literature on AI-associated psychosis describe people who came to believe a chatbot was a deity, or their Great Aunt Mildred, or that the model was conveying coded messages meant only for them. These are real cases. They deserve real clinical attention. They are also not what is happening when a knowledge worker wakes at 3 a.m. with a paper outlining itself in their head, works for four hours, and goes back to bed fulfilled.

The distinction is one of kind, not of degree. That distinction is the exact line the medical history of pathologising difference has repeatedly failed to draw. When Bryan Johnson — by any honest measure the most disciplined health optimiser alive — describes his own derailment as magic, not pathology, the cultural reflex to reach for psychiatric language deserves to be interrogated by the people doing the reaching.

A nineteenth-century American physician named Samuel Cartwright argued that the desire of enslaved people to escape was a mental illness; he named it drapetomania. Hysteria, from the Greek for womb, was applied to women’s distress for more than two thousand years before being retired from the Diagnostic and Statistical Manual in 1980. Homosexuality was a sociopathic personality disturbance in the DSM-I of 1952, a sexual deviation in 1968, and a mental disorder until the American Psychiatric Association removed it by trustee vote in December 1973 — not because new science had arrived, but because activists, clinicians, and the broader culture finally refused the framing.

The history of mental health is not only a history of relieving suffering. It is also, repeatedly, a history of laundering cultural discomfort through diagnostic language.

The sleep half of the question matters here too. The pattern most of us call normal sleep — sixteen waking hours, one consolidated eight-hour block at night — is largely an artefact of the electric light. The historian A. Roger Ekirch has assembled more than two thousand references across a dozen languages, going back to ancient Greece, showing that the dominant sleep pattern in pre-industrial Western life was biphasic: a first sleep of three or four hours after dusk, a long wakeful interval after midnight used for prayer, reading, sex, conversation, the interpretation of dreams, then a second sleep until dawn. The eight-hour consolidated block we treat as biologically normative is, in historical terms, recent — younger than the gas streetlight, younger than the labour movement, younger than the modern weekend.

Like the prohibition on working on the Sabbath, the consolidated eight-hour sleep block has some roots in biology and a great deal of cultural template. Calling 3 a.m. wakefulness a disorder is partly biology and overwhelmingly an inherited habit of mind from a civilisation that has been intensively electrified for two centuries.

A real clinical edge does exist. Sycophantic models reinforcing genuinely delusional frameworks in already-vulnerable patients is a documented and serious phenomenon, and the responsible thing is to name it. But naming a clinical edge in vulnerable populations is not the same act as folding the cognitive abundance of an entire generation of knowledge workers into a diagnosis whose evidence base, presently, is a handful of case studies.

The honest question is not whether AI is making us ill. The honest question is whether the reflex to medicalise the cognitive unbottling is itself the older, more dangerous pathology.

The infinite backlog

So what is happening?

People generate dozens of ideas a day. Everyone does — in the shower, on the treadmill, on the morning walk. On inspecting the idea, most of us also say: that sounds like a week’s work, and I don’t have ten extra minutes, never mind a week. The idea gets parked in a notebook. Usually it gets forgotten.

That friction — the gap between imagining a thing and being able to build it — is the cork that has held human creativity, insight, and imagination corked for the whole history of our species. AI has popped the cork. And what happens when, suddenly, anything is possible?

The answer has a name. The infinite backlog.

For most of modern life, productivity advice assumed that execution was hard. Ideas were cheap; getting them done was the bottleneck. You had a thought in the middle of the night, maybe scribbled it down, and went back to sleep. The friction of drafting, researching, structuring, editing, formatting, publishing, building, and testing acted as a natural filter. Some ideas should die. Some should rest. Some should stay in the notebook. Some should become a paragraph, not a company.

Now the idea can become a draft, a deck, an outline, a research memo, a prototype, a brand, a website, and a launch plan before breakfast.

That is intoxicating. It is also disorienting. The old productivity question was: How do I get more done? The new question is: How do I stop possibility from consuming my life?

AI does not merely help us process the existing backlog. It creates a new species of backlog — the infinite backlog of possible thoughts, essays, products, talks, ventures, research threads, intellectual side quests, and unrealised versions of ourselves.

Before AI, many ideas died naturally because execution was too expensive. AI changes the survival rate of ideas. It makes ideation executable.

That is why the 3 a.m. wakefulness is not just a sleep matter. It is a governance matter, in the older and more interesting sense of the word — the question of how a mind, and ultimately a civilisation, organises its relationship to what has suddenly become possible.

A notebook at 3 a.m. is a capture device.

An LLM at 3 a.m. is an accomplice.

That may be the central distinction. The notebook preserves the visitation. The LLM escalates it. It takes the spark and offers structure, examples, titles, outlines, counterarguments, citations, metaphors, publishing strategies, and next steps. It removes the friction exactly where friction used to protect sleep.

This is not a personal-discipline issue. It is a question for knowledge workers, writers, founders, consultants, researchers, investors, clinicians — anyone whose work depends on idea generation.

• What happens when creativity is unlocked faster than judgment can govern it?

• What happens when the mind’s imaginative output is no longer bounded by the practical cost of acting on it?

• What happens when every half-idea can become a plausible project?

Classical time management was built for a world of finite tasks: appointments, errands, inboxes, deadlines. It assumed the work was mostly known and the challenge was sequencing. AI creates a world in which the work is no longer mostly known. It is generated continuously. The backlog is not merely discovered; it is produced.

The serious problem is no longer task overload. It is possibility overload.

Five architectures of an unbottled mind

What follows is not a prescription. The reflex to issue rules is, as I have argued above, half of the problem. What follows is five architectures that some of us — knowledge workers running our lives in close partnership with AI — are personally converging on. They are local solutions to a global change. They will not be the final answer. The final answer will be cultural, not personal, and it has not yet arrived. In the meantime, here is what is working for me, offered as material for your own thinking, not as a programme.

1. Idea triage

Not every good idea deserves a project.

This is the hardest discipline for creative people because good ideas feel morally demanding. Once you can see them, you feel responsible for them. AI intensifies that feeling by making each idea seem more actionable than it would otherwise be.

But intelligence without triage becomes self-harassment.

The better question is not Is this idea good? The better question is What is the right container for this idea? Some ideas are one sentence. Some are a paragraph. Some are a future chapter. Some are simply compost — they feed the next idea by decomposing in the dark.

2. Backlog ethics

The infinite backlog can become a form of cruelty.

A human being can only live one life. AI can show you hundreds of plausible lives: books you could write, companies you could start, posts you could publish, courses you could build, systems you could design, relationships you could deepen, bodies you could optimise, languages you could learn, careers you could reinvent.

The result is not always liberation. Sometimes it is a more sophisticated form of inadequacy. You are no longer oppressed only by duties. You are oppressed by possible selves.

What is emerging — for me, and for the people I most respect using these tools — is an ethics of the backlog. A way of saying: this will not be done, and that is not failure. This possibility is beautiful, but not mine. This idea is good, but not now. This project is plausible, but too expensive in life-force.

The backlog needs an ethics because, unmanaged, it produces a new species of guilt the older productivity literature never anticipated.

3. State protection

There are times to generate, times to decide, times to execute, and times to recover.

AI blurs those states. A conversation can begin as reflection, turn into brainstorming, become a strategy session, produce a project plan, and end as a publishing workflow. That is magnificent during the right state. It is destructive during the wrong one.

The skill I have found indispensable is knowing which mode I am in, and protecting it. At night, capture. During recovery, restore. During execution, narrow. During strategy, choose. State protection may turn out to be one of the core skills of the AI age, the way concentration was a core skill of the print age.

4. Friction by design

Old friction was accidental: lack of tools, lack of access, lack of speed, lack of collaborators, lack of distribution. The history of productivity was the history of removing it.

New friction has to be chosen.

That sounds counterintuitive. We have treated friction as the enemy for decades. But some friction is protective. Some friction gives judgment time to catch up with imagination. “No AI after 9 p.m.” may become the intellectual equivalent of “no alcohol before driving.” Not because the tool is bad. Because powerful tools require context.

A notebook by the bed is friction by design. It allows the idea to be honoured without being amplified. It says: I saw you. I will not lose you. But I am not building a cathedral at 3:17 a.m.

5. The difference between creativity and compulsion

This is the most delicate of the five.

The same mental energy can be muse or mania, vocation or avoidance, generativity or dysregulation. The question is not whether the ideas are good. The question is whether the system is governing them. An idea can be brilliant and still arrive at the wrong time. A thought can be valuable and still not deserve immediate engagement. A mind can be alive and still need sleep.

The mature creative practice is not to suppress the daemon. It is to decline to hand the daemon a production team at 3 a.m.

The 3 a.m. rule (as I run it)

I have one personal rule. I offer it not as a prescription but as a working example of architecture-four-and-five in action.

At night, capture. Do not collaborate.

Notebook. One sentence. Back to bed.

No LLM. No research. No outlining. No just one prompt. Because the first prompt is the new cigarette. It is not the content that gets you. It is the loop.

The mature discipline, for me, is to preserve the spark without turning on the factory.

I do not want to optimise away the visitation. The liminal mind is real. The half-dreaming state can loosen associations. It can surface material the daylight mind misses. Many writers, scientists, and founders have relied on that threshold between sleep and waking. I have benefited from it myself. The point is not to suppress the 3 a.m. mind. The point is to honour the spark without escalating it — to receive the idea without recruiting an accomplice the moment it arrives.

That is my rule. Others will run different rules. Some of the most productive AI-builders I know run the opposite rule — they take the 3 a.m. wakefulness as a working session, sleep five hours, and treat the post-midnight hour as the most generative window of their day. They may be right. I do not yet know whether I am right. What I am sure of is that something needs to be running in the place where the consolidated-eight-hour cultural template used to run automatically.

Polyphasic sleep and the wrong dream

Polyphasic sleep belongs in this conversation as a historical warning.

It has always seduced ambitious people: da Vinci myths, founder culture, military sleep hacks, the Ubermensch schedule, the fantasy that sleep can be compressed without consequence. The appeal is obvious. If life is short and ambition is large, sleep looks like a design flaw.

For serious judgment work — writing, investing, leadership, science, medicine, parenting — the case against chronic sleep compression is strong. You may buy hours, but you pay in mood stability, impulse control, memory consolidation, emotional regulation, and error rate. The bill does not always arrive immediately. That is the danger. You can feel productive while becoming less wise. You can generate more while deciding worse. You can produce the appearance of momentum while degrading the faculties that make the work valuable.

The dream of polyphasic sleep was that we could hack biology and reclaim time.

The better dream is that we learn how to align cognition, tools, and biology so that good thinking has somewhere healthy to live. Biphasic, in Ekirch’s sense, is one such alignment — historically attested, biologically plausible, culturally forgotten. Monophasic is another. Polyphasic may be a third for the few who can sustain it. The point is that the menu is wider than the post-electrification template assumes, and the right architecture is the one you can sustain across years, not the one Sam Altman can sustain across a quarter.

What we might do in harness

The argument so far has been a governance argument. It has been about architecture, about state, about the practices an individual mind has to develop when the cork is suddenly out of the bottle.

That is the conservative half of the truth. The other half is larger, and worth saying plainly.

We are now, for the first time in our species’ history, sharing the cognitive load with an intelligence that parallels our own. Not a tool. Not an extension. Not a search engine. An interlocutor. A second mind in the room.

As I’ve described elsewhere - this is an entirely new cognitive infrastructure for human kind. Crudely, it is a library that talks back, learns dynamically, and can do work, not just inform.

The right question is not only how do I sleep through this?

The right question is also: what might eight billion human beings do, in harness with a parallel intelligence, that no previous generation could?

This is the utopian possibility worth holding alongside the cautionary one. The 3 a.m. wakefulness is not pathology. It is, for many of us, the first time we have seen what an unbottled mind feels like. The architectures in this essay exist not to suppress that wakefulness. They exist to make it survivable — so that the unlock can be metabolised, by individuals and by civilisations, into something better than the world we currently have.

That is the larger claim. The bottleneck has moved — from information to judgment, from execution to selection, from time management to possibility management. The people who thrive will not simply be the ones with the best tools. They will be the ones who can govern their relationship to possibility.

They will know when to generate and when to stop.

They will know when to ask the machine and when to close it.

They will know when an idea is a calling, when it is a distraction, and when it is just the mind sparkling in the dark.

The future may belong not to those who can think fastest, but to those who can honour the idea without obeying it.

Mass intelligence is the dream. Possibility management is the practice that lets us survive arriving at it.

Author’s note: I wrote this on Sunday 10 May 2026 while in Las Vegas for the World Series of Poker. The 3 a.m. wakefulness referenced in the opening is a current and continuing feature of my life and indeed - my wrestling with it was the aegis of this post.

I do not regret it. I am still learning to govern it. The architectures offered in section five are working hypotheses, not settled doctrine — feedback in the comments and to me directly is welcome and will shape the next revision.

Sources

• A. Roger Ekirch, Segmented Sleep in Preindustrial Societies — PMC review (2016); originating monograph: At Day’s Close: Night in Times Past (2005).

• Morrin H, et al., Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis” — JMIR Mental Health (2025).

• Psychiatrists hope chat logs can reveal the secrets of AI psychosis — UCSF, January 2026.

• “Gay Is Good”: History of Homosexuality in the DSM and Modern Psychiatry — American Journal of Psychiatry Residents’ Journal (2022).

• AI psychosis is not a new threat — PMC commentary.

• Bryan Johnson, public posts on his Claude / digital-twin experience and his partner Kate Tolo’s remark — Instagram, April 2026.

• Sam Altman, public announcement of polyphasic sleep in response to GPT-5.5 in Codex — reported by The Hans India, April 2026.

• The AI Daily Brief (Nathaniel Whittemore), “Why Agents Make Every Job a Startup” — Spotify episode, May 2026. Aaron Levie quotation as cited in the episode.

• The AI Daily Brief (Nathaniel Whittemore), “Is AI Doom Going Out of Style?” — show page, May 2026. Source of the Bryan Johnson full quotation (“seed becoming a tree and a zygote becoming a baby”) and the “Kierkegaardian dizziness of freedom” framing.

• Søren Kierkegaard, The Concept of Anxiety (1844) — original source of the dizziness of freedom formulation Whittemore borrows.

The old enterprise logic, the Coase model of the firm, suggested that in-house expertise entailed fewer frictions. That is demonstrably false today. The gap between insight and execution has dropped to near zero — a talented innovator inside a business may face 9-month enterprise approval frictions; entrepreneurs and microenterprises can ship better in 9 days.

The question we will be debating for the next ten years is whether “economies of scale” logic still holds, or whether modern enterprises are “economies of friction” or more harshly, as I write, anti-innovation machines.

And (as you will read below), my reading of the data is that the gap between slower adopters and folks at the frontier is widening. “Power users” and “power-enterprises” are using AI to learn AI, using AI to deploy AI, and have wrangled the data and infrastructure challenges that were such a pain in 2025.

What makes AI the most fascinating thing I’ve ever studied is its multi-dimensionality — from math and models, to tools, to markets, to ethics, to governance, to neuroscience, to business strategy and design. Keeping up with even one dimension is a full-time job — but keeping sight of the whole is what business leaders must do.

Accordingly, here is what we cover, section by section:

Everything I write stays free. Always will be. But if you want to back the work — $5/month, or $25/year right now (75 percent or so off until 1st of May). For the price of a business book, you get access to all my books via the Corpus (in whatever language) and these monthly roundups.

Back the work → https://thinkbiggerthinkbetter.substack.com/23601d84

SECTION 1: PEOPLE-FIRST (tm) — Attitudes, Trust, and What I’m Seeing

“The gap between thought and action, between belief and will, prevents us from solving our most pressing individual and societal problems.” (The Science of Organizational Change)

People-first is increasingly recognized not just as an ethical statement, but as a technical necessity. Hedge fund Parker Gale has recently picked it up; and McKinsey claim “nearly 60% of respondents cite knowledge and training gaps as the primary barrier to implementing responsible AI practices — up from ~50% the previous year.”

1.1 Everyone started talking about trust. As the Chinese say, “Talk doesn’t cook rice.”

April was the month trust became a consulting buzzword. McKinsey published on it. Edelman’s Trust Barometer flagged AI trust as a standalone metric for the first time. We taught it in our leadership programs starting in 2004, and finally, a friend persuaded me to publish a whitepaper on our work. Read the trust whitepaper →

The problem: “we need to build trust” is the organisational equivalent of “we should exercise more.” It’s true, unfalsifiable, and changes nothing.

What’s missing is measurement. Trust in AI isn’t a single variable — it’s at least four: Reliability, Informativeness, Safety, and Transparency (RIST). Each can be independently calibrated, independently miscalibrated, and independently diagnosed. Undertrust is as costly as overtrust — the team that won’t touch the tool and the team that ships unreviewed AI output are both failing, in opposite directions.

The RIST Trust Diagnostic is FREE and live on the Adaptive Adoption website

It’s open-source and takes five minutes.

If your organisation is “talking about trust,” ask whether anyone has measured it.

1.2 The Four Levels: where your people are

“Learning AI is more like carpentry than calculus.” (March AI briefing)

We launched the AI Mastery diagnostic last week. The sample is small (n=21, self-selected, skewed toward leaders in our network) — but the distribution is already instructive:

• User (43%) — can operate AI tools; cannot evaluate output quality or design workflows

• Practitioner (19%) — beginning to integrate AI into daily work; judgment still developing

• Builder (24%) — designing AI-augmented workflows; can evaluate and iterate

• Architect (14%) — shaping organisational AI strategy; teaching others

Of course, this is an early sample - but if it were skewed, my guess is that it would be heavily skewed toward overestimating rather than underestimating capability. Why? It is a self-report and people generally assess themselves as more able than they are, and second, it is people connected to me, who tend not to be hard skeptics.

The question for every enterprise: where do you need your people to be? If you need Practitioners and you have Users, the gap is behavioural, not informational.

Training won’t close it. Leadership modelling will.

Take the free diagnostic →

Not where you want your people to be. Where are they?

1.3 What people say their barriers are — and what that really means

Non-coders are able to build complex architectures. Above is one of mine. But can they in practice - enterprises would block that whole stack.

The enterprise data is consistent across Gallup, BCG, and PwC this month: the top self-reported barriers to AI adoption are time (”I don’t have time to learn”), training (”nobody’s shown me”), and permission (”I’m not sure I’m allowed to”).

These are not technology barriers. They are leadership barriers dressed in operational language.

• 87% of organisations say they prioritise workforce upskilling. But 57% of executives expect humans to manage and direct AI agents — a competency almost nobody is training for.

• 46% of non-users say they prefer doing things the old way. 40% cite ethical opposition, privacy, or disbelief. These are not irrational — they are rational responses to absent leadership signals.

PwC’s April study quantified the cost of shallow integration: 74% of AI’s economic gains are captured by 20% of companies. The 20% redesign workflows around AI. The 80% overlay AI on existing processes and wonder why it doesn’t stick. This is the Adaptive Adoption thesis in data form: adoption is a change problem, not a technology problem.

1.4 The 50/50 split — and why it’s more interesting than it sounds

Gallup Q1 2026 (n=23,717): 50% of US employees now use AI at work. First time at this milestone. Generative AI reached 53% population-level adoption in three years — faster than PCs or the internet.

First of all, folks who follow me know I think self-reports are close to worthless — in this case, “what does ‘use’ mean?” I used AI in 2023; I use it now. From light occasional research then, to it being integrated in 100 percent of what I do, including building my own LLM at home, relearning coding, and one dozen AI tools in active weekly use (including 4 frontier model LLMs.)

I’m aware this probably horrifies well over half of you.

There are better surveys, such as the AI Daily Brief monthly survey, which is more granular “how many large language models do you use?” or “do you even code bro?” or “are you building agentic workflows?” (The latter are what I call “use”.)

And Gallup’s data exposes the gap: 50% say they “use” AI, but only 1 in 10 say it has fundamentally transformed their workplace. That’s a lot of dabbling dressed up as adoption.

But stats say a little something: 50% do means 50% don’t.

The Gen Z paradox is the sharpest version: the most digitally native generation is simultaneously the most sceptical of AI and the most willing to use it in unsanctioned ways. Gallup’s February–March 2026 data: the share of Gen Z who describe themselves as excited about AI fell from 36% to 22% in one year. The share who describe themselves as angry rose from 22% to 31%. Bloomberg ran this on April 19; Fortune ran a piece on April 8 headlined “Gen Z workers who fear AI will take their job are actively sabotaging their company’s AI rollout.” Read that headline again.

The enterprise implication: monolithic adoption strategies are dead.

And here is one of the grave change management flaws — one-to-many, “sheepdip” approaches target the “average” too often. But what if the distribution is bi-modal? (Two humps not one if you slept during stats class.)

You can’t design a single training programme for a workforce where one cohort is angry and sabotaging and another is building agents at home.

Adaptive Adoption starts with diagnosis — which is why the tools in 1.1 and 1.2 exist.

1.5 The expert delusion — or, talking their book

The expert-public perception gap has become a chasm: Stanford’s recent 400-page tome suggest that 73% of US AI experts view AI’s job impact positively; only 23% of the public agrees — a 50-point gap. Similar divides exist on the economy (69% vs. 21%) and medical care (84% vs. 44%).

The experts are smoking crack — or as we used to say in the markets, “talking their book.”

Exhibit A — Muskian economic growth forecasts:

• “Double-digit growth... is coming within 12 to 18 months.”

• And as if that were a low-ball guess: “Triple-digit is possible in about five years.”

For context: the OECD projects US real GDP growth at roughly 1.5% for 2026. (And depending on whom you ask, and what they count, AI may be all of that and then some. This is contested territory, but in my view, auspicious names such as Goldman Sachs have it very wrong.)

In 2025, I estimated total AI contribution to the economy at $2 trillion — that includes the full stack (land acquisition, construction, electrical grid upgrades, cooling systems, fibre, the chip supply chain’s domestic portion, cloud services, enterprise software, consulting, training, hiring, the entire SaaS layer built on top.) The $700B capex figure that gets cited is just the top 5-7 hyperscalers’ direct spend. It doesn’t capture the construction crews building data centres in Iowa, the utilities upgrading substations, the HVAC contractors, the legal and permitting work, the enterprise IT teams retooling, the consulting firms staffing AI practices, or the thousands of startups burning through venture capital.

No general-purpose technology in history — not electrification, not computing, not the internet — has produced double-digit GDP growth in its first decade of deployment. For more than a century, U.S. real GDP growth has lived in the low single digits. Musk’s prediction isn’t optimistic; it’s at best early, at worst delusional.

His thinking is emblematic of what I call “naive solutionism” — technology as panacea. But technology hits human systems which respond on decades-long time scales. And consider this thought-experiment: how many of the world’s problems are failures of smarts? Could we think our way out of poverty? Out of climate change? In businesses, how many of the problems are problems that hiring a few Einsteins would solve? None. Workplaces don’t have Einstein-shaped holes in them — so they end up handing Einstein a shovel — like asking Einstein to summarize some emails, or manage a spreadsheet.

Meanwhile, the serious analysts converge on displacement but diverge on recovery. The classical Ricardian pattern — technology destroys jobs, then creates new and better ones — is the J-curve optimists invoke. Amodei at Anthropic broadly holds this view, with caveats about timeline and transition pain. Acemoglu (Nobel 2024) is more pessimistic: his published estimates suggest AI adds a modest ~0.5% to GDP over a decade, with gains captured narrowly — not the broad-based recovery the J-curve implies. Imas at Chicago Booth, presenting on Bloomberg Odd Lots, offered the sharpest formulation: AI may be “categorically different” from prior general-purpose technologies, meaning the historical pattern where new jobs replace displaced ones may simply not hold. They agree the descent is coming. They disagree on whether the curve bends back up, how far, and for whom. The pundits promising sunlit uplands are not the ones building transition plans. The people-first question isn’t whether disruption is coming. It’s whether your organisation has a plan for the humans in between.

1.6 The displacement numbers accumulate — and someone finally put a timeline on it

Q1 2026: 345,000 jobs eliminated as profitable companies redirect salary budgets to AI. Software developer employment for ages 22–25 dropped nearly 20% since 2024. Women constitute 86% of the most vulnerable worker cohort. AI-skilled employees earn 56% more than peers — the premium widens as the entry path narrows.

Then Dario Amodei put a concrete timeline on it: “a significant proportion” of entry-level finance, consulting, and tech positions could vanish within 1–5 years. This is the most specific displacement forecast from a frontier-lab CEO to date.

Bloomberg Odd Lots featured Alex Imas from Chicago Booth, presenting survey evidence that economists, AI specialists, and superforecasters are converging: faster AI progress correlates with lower employment. AI may be “categorically different” from prior general-purpose technologies — meaning the historical pattern where new jobs replace displaced ones may not hold. The Overton window has moved. Leaders who dismiss displacement as “Luddite anxiety” are now behind the industry’s own positioning. OpenAI published a policy paper proposing robot taxes and a 32-hour workweek. When the company most aggressively scaling frontier AI is proposing redistributive mechanisms, the question isn’t whether disruption is coming — it’s whether your organisation has a transition plan.

1.7 Backlash: from data-centre protests to targeted violence

Data-centre protests intensified worldwide in April, uniting bipartisan coalitions over environmental and energy concerns. Maine passed the first-in-the-nation data-centre moratorium. OpenAI paused Stargate UK, citing energy costs among the highest globally. With a US election coming, being anti-AI will be the new populist shibboleth — right and left. NIMBYs at work — three planned Wisconsin data centres would add almost 10 percent to the state economy — but at what cost to the environment and domestic electricity prices?

Then it got darker. Molotov cocktails and gunfire targeted Sam Altman’s San Francisco home (April 13–15). The suspected attacker — a 20-year-old community college student — cited AI extinction risk. Bloomberg’s podcast: “Why the AI Backlash Is Only Going to Get Worse.”

This is no longer a communications problem. It’s a societal legitimacy problem. And it connects to the trust gap in 1.1: when institutions fail to measure and demonstrate trustworthiness, the vacuum fills with fear.

And this is yet another reason why traditional change management isn’t suited to AI — nobody ever hoisted a placard against SAP or Excel rollouts. If a chunk of your workforce is actively hostile, that isn’t a problem fixed with lofty words from the CEO.

And while models are 300x more efficient, token use is expanding faster than that. Inference costs for fixed-capability models have collapsed — roughly 280x for GPT-3.5-level performance in under two years — but that has not reduced total compute pressure. It has expanded the addressable market for intelligence. Cheaper tokens create more tokens: longer contexts, agent loops, multimodal generation, always-on copilots, synthetic data, and reasoning-time computation. The result is not efficiency replacing infrastructure demand, but efficiency making vastly more demand economically possible.

1.8 Harness engineering is the People-First discipline nobody named yet

Prompt engineering: 2024. Context engineering: 2025. Harness engineering 2026.

Prompt engineering is so 2024 — harness engineering is what matters more today.

The biggest conceptual shift of the month: Agent = Model + Harness.

The harness is everything that wraps around an AI agent: the tools it can access, the guardrails that keep it safe, the feedback loops, the observability layer. If 2025 was the year of the agent, 2026 is the year of the harness. The industry discovered, painfully, that building an agent is the easy part. Making it reliable, cost-predictable, and safe in production is where the real engineering happens.

The uncomfortable truth: the harness is governance. It’s trust calibration made operational. It’s decision rights encoded as constraints. It’s the Behavioral Governance framework expressed in system prompts, eval suites, and permission boundaries. The people who design harnesses are doing organisational design whether they know it or not.

Anthropic published on context engineering in April — treating context as a first-class system with its own architecture, lifecycle, and constraints. It’s an important contribution. But I’d argue it’s already been superseded. Context is one layer of the harness; harness engineering is the whole discipline. It includes context, yes — but also tool selection, guardrails, feedback loops, observability, permission structures, and governance. The right harness determines whether an agent produces value or chaos. That’s not a technical problem. It’s a leadership problem.

SECTION 2: MODELS & MATH — The Signal Under the Noise

“The engine is not the car. The most powerful model in the world is useless without the harness, the context, and the human judgment that tells it where to drive.”

2.1 The biggest model news is that model news doesn’t matter to most people

Model capabilities aren’t flattening — METR’s Time Horizon benchmark shows AI improving at ~10× per year, with a doubling time of about 4.3 months. Opus 4.7 launched 16th April; ChatGPT dropped 5.5 on 24th April. DeepSeek V4 dropped. Qwen, Zhipu GLM-5.1, Gemma 4 — the open-weights ecosystem keeps delivering. Mythos (see below) was deemed too powerful for general release.

But here’s what matters: the marginal returns for people and enterprises from model improvements are flattening. The top four models on the Arena Leaderboard are separated by fewer than 25 Elo points. For most enterprise users, the difference between this month’s frontier model and last month’s is invisible. And many enterprises can’t use them anyway — banned by policy, blocked by procurement, or locked behind approval chains that move slower than release cycles.

The model glow-ups are noise. The signal is what happened around the models.

And here is what we might be talking about in 2030. How Microsoft screwed the pooch.

The conspicuous absence from any serious benchmark comparison is Microsoft Copilot. A company with 400 million Office 365 seats had a two-year head start on enterprise AI integration — arguably the largest distribution moat in technology history — and squandered it by treating AI as a pricing lever rather than a product revolution. A $30/seat/month tax bolted onto products that didn’t fundamentally change. The harness was an afterthought; the revenue model was the product. Meanwhile, the competitors who took harness design seriously — Claude’s integrated workspace, Gemini’s 2M token context — are shipping tools that actually reimagine how work gets done. Never has a moat been so quickly pissed away, and the lesson for every enterprise leader is instructive: distribution without genuine product transformation is just a more expensive version of the status quo.

2.2 What I call “fractal intelligence” — brilliant and brittle

The Stanford HAI term is “jagged intelligence.” I prefer fractal — the pattern of brilliance and brittleness repeats at every level of magnification. Gemini Deep Think won a Gold Medal at the International Mathematical Olympiad. The same class of models reads analog clocks with 50.1% accuracy. AI scored below 20% on replicating astrophysics papers. PhD-level experts still outperform the best AI agents by 2× on complex multi-step tasks (Nature/Stanford HAI).

The consensus reads that as reassuring: “see, humans still win.” I read it the other way. It takes a PhD to outrun an LLM. Unless someone has doctoral-level domain knowledge, they are running in second place. That is not a story about human superiority. It is a story about what the floor of competence just became — and most of your workforce is below it.

This reinforces the augmentation thesis and undercuts the automation thesis, but not in the way most people frame it. The fractal pattern means you can’t predict from capability benchmarks where an AI will be useful in your context. The only way to know is to test it — which requires the kind of experimentation infrastructure that Section 1 describes. The organisations capturing value are the ones with the manager-multiplier effect, the trust calibration, and the permission structures that let people try.

2.3 The utility signal: integrated workspaces stole the march

In 2023, AI was cute. In 2024, it was clever. In 2026, it got interesting.

What did matter in April for actual users: Claude stole a massive march with an integrated workspace. Cowork, Code, Projects, scheduled tasks, persistent memory, Claude Design (which is “lit”)— these aren’t model improvements. They’re harness improvements. They reduced the friction between “I have an idea” and “I have a working output” more than any benchmark gain this year.

Claude Design in particular (Figma stock fell 7.28% on announcement day) demonstrated something important: the competitive moat is no longer model quality. It’s the completeness of the system wrapped around the model.

The harness, again.

SECTION 3: ECON & MARKETS — The Four-Body Problem

”The data centre is a factory. Its raw material is data, its power is accelerated computing, and its output is intelligence — delivered as tokens.”_ — Jensen Huang, GTC 2026

3.1 The emerging field of tokenomics — and why economists are scrambling

Jensen mentioned tokens 70 times in a two-hour keynote, lending credence to what we’ve been studying since 2024. A new four-body problem is forming at the centre of economic theory: labour, capital, energy, and tokens. Traditional economics operates with two factors of production (labour and capital.) Vaclav Smil wrote in Energy and Civilizations that we had long ignored the third. AI has introduced a fourth: tokens — units of intelligence that are simultaneously a production input, a consumption good, and an infrastructure cost.

China’s daily token consumption hit 140 trillion in March 2026 — up from 100 billion at the start of 2024, a 1,400× increase in two years. “Tokemaxxing” — maximising token throughput as a competitive strategy — is becoming a thing. MiniMax charges ~$1/million output tokens versus $15+ for Claude. Google Research’s TurboQuant delivered 6× memory compression with zero accuracy loss, repricing the entire inference stack overnight.

The question we are racing to answer: how do you model an economy where the marginal cost of intelligence is approaching zero while the marginal cost of energy is rising? The consumer surplus from generative AI reached an estimated $172 billion annually by early 2026 — massive value from mostly free tools. But whether the worker is winning is the open question that Imas, Amodei, and the Odd Lots data are answering with “probably not, at least for the entry-level cohort.”

3.2 Capital concentration: irreversible, and now it has a number

OpenAI closed $122B at $852B. Google is investing up to $40B in Anthropic at a $350B valuation. Cursor raised $2.3B at $29.3B. Bezos’s Project Prometheus is raising $10B. Goldman projects $700B in AI capex in 2026. Taiwan’s export orders surged 65.9% YoY to a record $91.12B, driven almost entirely by AI chip demand.

In comparison, the big infrastructure buildouts of previous generations, highways, electrification, the moon landings were on the order of $20 billion a year.

The overbuild narrative is dead. The question has shifted from “is this too much?” to “can we build fast enough?” And the concentration creates a dependency problem: a small number of hyperscale players now control the infrastructure that every enterprise depends on. Thoma Bravo signed a multiyear Google Cloud deal covering its entire $183B AUM portfolio. Private equity is industrialising AI adoption at portfolio scale. The implications for enterprise buyers who are building dependency on two or three providers are significant and underexplored.

3.3 The ROI question that matters: effects on workers, effects on capital

The effect on capital is clear: AI capex is delivering returns. The effect on workers is the question that determines whether this is a broadly shared prosperity or a capital-concentration event. Young developer employment down 20%. AI-skilled premium at 56%. 89% drop in AI researchers moving to the US. Women 86% of the most vulnerable cohort.

The four-body problem — labour, capital, energy, tokens — is the economic framework for the decade. Economists are racing to solve it. The answer will determine policy, and policy will determine whether the gains are distributed or concentrated. Every enterprise leader needs a position on this, because the policy is coming whether they participate or not.

SECTION 4: TOOLS & ORCHESTRATION — Memory Is the Frontier

“Context is the new code.”

4.1 Karpathy’s LLM Wiki: the post-RAG signal

Andrej Karpathy published his “LLM Knowledge Bases” concept in early April — structured markdown wikis maintained by the LLM itself, replacing the complexity of RAG pipelines. 16 million views. The implication: as context windows expand toward 1M+ tokens, the retrieval infrastructure that enterprises spent millions building may be unnecessary. Lightweight knowledge architectures could displace complex vector-search systems. Karpathy’s weekend project may have transformational consequences.

It looks impressive. The miracle of today is it isn’t that hard — now that the way is paved. (In fact, I’ve built a home LLM and a sophisticated RAG setup with 40-year out of date coding skills. Looks worse than it is.)

This connects to the larger pattern: the tools that matter aren’t the ones that push model capabilities further. They’re the ones that reduce the distance between human intention and useful output.

4.2 Claude Design: the biggest splash of the month (RIP Canva, Adobe, Figma)

Claude Design launched 17th April — built on Opus 4.7, available to Pro/Max/Team/Enterprise subscribers. You describe what you need; Claude builds a first version; you refine through conversation, inline comments, direct edits, or custom sliders. During onboarding, it reads your codebase and design files to build a design system automatically.

Over my career at IBM, PwC, Morgan Stanley, and publishing through the Financial Times, I’ve had a lot of designing done for me — by professional designers, with professional budgets. The quality of what Claude Design produces, the speed of iteration, and the ability to have production-grade code (Rust, Python, Node.js) running under the hood was beyond my imagination. If you are thinking, your designs aren’t so hot Paul — compared to my ham-fisted clipart of years gone by, they are impressive enough for me!

Figma dropped 7.28% on the announcement. Anthropic said it “complements Canva.” (Well I’ve always hated Canva so I think that undersells it.) But what it does is eliminate the skill barrier between “I know what I want” and “I can make it.” For non-designers — which is most of the knowledge workforce — that’s transformative. For designers, it’s an augmentation tool that collapses the draft-to-refinement cycle. This is the kind of tool-level advance that matters more than any model benchmark.

4.3 Memory: personal, portable, persistent, auditable — the most important frontier

Here is my take on the most important and underreported development of 2026: memory is becoming the primary differentiator in AI productivity, and almost nobody is talking about the ownership implications.

Claude now has persistent memory across sessions. OpenAI Codex has persistent memory. These systems are accumulating context about how you work, what you’ve built, what you’ve decided, and why. This context is the most valuable asset in the human-AI relationship — and you don’t own it.

Nate B. Jones (one of my top-two AI commentators) flagged this in April: professionals are building the most important asset of their careers — accumulated AI context, prompt libraries, workflow adaptations — yet own none of it. Context portability is emerging as a workforce infrastructure problem akin to pension portability in the 1970s.

My position: responsible AI use means owning and managing your context. Memory should be personal (tied to you, not your employer’s platform; you control what’s retained, shared, and deleted), portable (exportable across tools and providers), persistent (survives sessions, survives job changes, accumulates over a career), and auditable (you can inspect what any system has stored about you). Four criteria — and no major provider currently meets all four. And so here is a tool — which takes 20 minutes — to help you create yours.

This is a governance problem, a labour-rights problem, and a competitive-strategy problem all at once. The terminology evolution tells the story: prompt engineering (2023) → context engineering (2025) → harness engineering (2026). Memory is where context engineering meets people-first governance — and the organisations that understand this are the ones where the technical architecture and the People-First agenda converge.

SECTION 5: ETHICS & GOVERNANCE — Mythos Changed Everything

“Ethics you don’t own aren’t ethics — they’re wallpaper.”

5.1 A single model triggered emergency responses across five G7 jurisdictions

Anthropic’s Mythos is the governance story of the year — possibly the decade. A general-purpose model that demonstrated the ability to find and exploit software vulnerabilities better than all but the most elite human hackers. Thousands of zero-day vulnerabilities identified across every major operating system and browser. Demonstrated sandbox escape.

Anthropic’s response: restrict access to ~40 companies via Project Glasswing, backed by $100M in usage credits. Microsoft, Apple, Google, CrowdStrike — defensive use only.

Then the cascade: Bessent and Powell summoned Wall Street CEOs. Bank of Canada convened its six largest banks. Bank of England, FCA, HM Treasury coordinated with NCSC. ECB called eurozone CROs. IMF’s Georgieva warned the global monetary system is unprepared. PBOC’s Pan flagged AI risks at the IMF Spring Meetings. SEC warned Mythos could compromise the Consolidated Audit Trail — the database of every US equity and options trade.

First time a single AI model has triggered coordinated financial-stability responses across five G7 jurisdictions plus China. Frontier AI capabilities are now treated as macroprudential risk.

And that was one model, in one month. Given what is established as the non-linear expansion of AI capabilities, regulators, citizens, and business leaders must stop legislating in the rear-view mirror.

5.2 And then it leaked

A group of users accessed Mythos through a Discord channel. One member was a third-party contractor for Anthropic. They guessed where the model was located based on previously leaked knowledge. Fortune, Bloomberg, CBS, Scientific American all covered it. A former national cyber director wrote that Mythos “can hack nearly anything and we aren’t ready.”

The lesson: capability-gating through controlled access is fragile. The most consequential model since GPT-2’s initial withholding in 2019, and the access controls lasted days. This is the stress test that every governance framework — including Behavioral Governance — needs to account for.

5.3 Regulation enters the operational era

The regulatory landscape shifted from “frameworks and principles” to “enforcement and liability” in April:

• Mobley v. Workday: first nationwide class action alleging systematic AI hiring discrimination. 1.1 billion applications processed through Workday’s screening tools. Proceeding as collective action under ADEA.

• EU AI Act full enforcement August 2, 2026: every AI system in recruitment, task allocation, and performance monitoring classified “high-risk.”

• California SB 947 “No Robo Bosses Act”: prohibits AI-only termination decisions; $500 civil penalties per instance.

• Anthropic suing the Pentagon for blacklisting after refusing to remove safety restrictions from Mythos. The first major corporate-government AI rights test. A frontier lab going to court to defend safety guardrails against a government that wants them removed — nobody had that on their bingo card.

The era of voluntary AI ethics is closing. The era of regulated AI governance is opening. Organisations that haven’t moved from principles to operational governance are running out of runway.

SECTION 6: ORG DESIGN & AI STRATEGY

“Creating change-agile businesses will eliminate the need for what we today call change management.”

6.1 The org chart is the next thing AI disrupts

I published this month on this topic directly — and for a splendidly written and artfully designed supplement, check out Howard Yu’s magnificent take while browsing.

The throughline: every major platform move in April — Salesforce going headless (entire platform exposed as APIs, MCP, CLI — “no browser required”), Zuckerberg building an AI CEO co-pilot, Block cutting 4,000+ roles with Dorsey and Botha’s “From Hierarchy to Intelligence” essay — points to the same conclusion. The organisational structures we inherited were built to route information through human layers. When AI handles the routing, the layers lose their justification.

Gartner predicts 20% of organisations will flatten middle management by 2028. The PE firms (Thoma Bravo, ParkerGale, Vista Equity) are already industrialising this at portfolio scale.

But here’s the counterargument that almost nobody is making: flattening works when the work is information-routing. It fails when the work is judgment, trust-building, and ethical calibration — the I-We-It dimensions of the Leadership Delta. The organisations that flatten blindly will discover they’ve removed the human infrastructure that makes AI adoption work. The managers Gallup showed as 8.7× multipliers? They’re the ones being eliminated.

6.2 The Salesforce signal: when agents are the users, everything changes

Benioff announced the entire Salesforce/Agentforce/Slack platform exposed as APIs, MCP, CLI. Simon Willison’s summary: “no browser required.” When AI agents — not humans — become the primary consumers of enterprise software, per-seat SaaS pricing collapses. The business model of the last two decades is being repriced.

Simultaneously, the x402 protocol (Coinbase, Cloudflare, Stripe) is building agent-to-agent payment infrastructure. Visa integrated via Nevermined for autonomous AI card purchases within cardholder-defined budgets. Agents aren’t just doing tasks — they’re transacting autonomously.

The org design question: if agents are transacting, who’s accountable? This is Behavioral Governance Dimension 2 — Agent Authority — in real time. The consequence-based autonomy tiering that the model card describes isn’t theoretical anymore.

WHAT I’M WATCHING IN MAY

• EU AI Act compliance deadlines — August 2 is closer than enterprises think

• Mythos broader release — will Anthropic expand access? Under what conditions?

• Memory portability — will any provider commit to context export standards?

• The four-body problem — first serious economic models incorporating tokens as a factor of production

• Your data — the AI Mastery diagnostic is live. Take it. Share it. The more data, the sharper the picture. paulgibbonsadvisory.com/diagnostics

Paul Gibbons is an AI adoption strategist, keynote speaker, and author of 8 books including Adopting AI: People-First Strategies for Enterprise AI Transformation. Creator of the Adaptive Adoption™ framework. Top-52 to follow in AI — Fortune 100 AI leader.

thinkbiggerthinkbetter.substack.com · paulgibbonsadvisory.com

SOURCE INDEX

Section 1: Gallup Q1 2026 Workplace AI Survey · Stanford HAI 2026 AI Index Report · Pew Research Center (March 2026) · PwC 2026 AI Performance Study · BCG AI Radar 2026 · Bloomberg Odd Lots (April 18, 2026) — Alex Imas, Chicago Booth · McKinsey trust paper · Edelman Trust Barometer 2026 · Paul Gibbons trust whitepaper

Section 2: METR Time Horizons (metr.org) · Anthropic — Claude Opus 4.7 announcement (April 16) · Stanford HAI 2026 — jagged/fractal intelligence data · Nature (d41586-026-01199-z) — AI agents vs PhD experts · Anthropic — Claude Design announcement (April 17)

Section 3: Bloomberg — China token economy (April 20) · Google Research — TurboQuant (ICLR 2026) · Stanford HAI — consumer surplus estimates · Bloomberg — OpenAI $852B, Google-Anthropic $40B

Section 4: Karpathy — LLM Knowledge Bases (karpathy.ai) · TechCrunch — Claude Design launch (April 17) · Nate B. Jones — context portability (natesnewsletter.substack.com, April 17) · Anthropic — context engineering (anthropic.com/engineering) · Martin Fowler — harness engineering (martinfowler.com)

Section 5: Fortune — Mythos leak (April 23) · Bloomberg — Mythos regulatory cascade (April 22) · Scientific American — What is Mythos · Anthropic — Project Glasswing · Seyfarth Shaw — Mobley v. Workday analysis

Section 6: Simon Willison — Salesforce headless (April 19) · Bloomberg — Zuckerberg AI co-pilot (April 16) · Bloomberg — x402 protocol (April 2)

Sam Altman dropped five principles for OpenAI on Sunday. Democratisation. Empowerment. Universal Prosperity. Resilience. Adaptability.

I love them. That is what people-first AI SHOULD be about. However…

Universal Prosperity. From the CEO of an $852 billion company that just converted from a non-profit. The word “AGI” — referenced 12 times in the 2018 charter — now appears twice. Three core safety commitments from the original charter were quietly dropped, including the pledge to stop competing and collaborate if another lab got closer to AGI first.

He published these 12 hours before jury selection in Elon Musk’s $134 billion lawsuit.

Enron’s values were Respect, Integrity, Communication, and Excellence. They were printed on the lobby wall. I teach this case in every leadership programme: values you announce are decoration. Values you enforce under pressure are culture. The gap between the two is where organisations die.

Do leaders deserve credit for articulating what matters? Some. Meagre. Saying “universal prosperity” while sitting on an $852B valuation and cutting safety commitments is the corporate equivalent of thoughts and prayers. But I’ll give Altman this: he’s now on the record. These five words are now testable propositions. And we should test them — relentlessly.

The proof of the pudding is in the next 12 months, not the next 12 hours.

Here is a post on what socialist AI could look like…

The idea that happiness is the most important goal in life has a lot going for it. Philosophers, Eastern and Western, from John Stuart Mill, to the Buddha, testify to its importance. Visit the psychology section of your local bookstore and every second title, it seems, is about happiness. I’m going to suggest that there is much more to the good life than happiness, and that the pursuit of happiness is a recipe for misery.

Here is the self-assessment for those impatient souls.

Human flourishing self-assessment

Lets do a thought experiment:

• Is a soldier, fighting overseas for their country, happy?

• Does caring for a sick child, or sick parent, make you happy?

• Did summitting Everest make Sir Edmund Hillary happy?

• Was Mother Theresa, in her work with the world’s poor, happy?

• Are people who work for causes, say environmental, social justice, or greater freedom, happy?

• Is a scientist, who devotes a lifetime to furthering our understanding of the world, happy?

• Is an entrepreneur working 100 hours a week to make their dream come true?

The answer is “perhaps” - maybe they are - but happiness was not what drives such pursuits. Yet, we admire such lives. We admire the virtues we see: valor, courage, service, honor, dedication, honesty, compassion, and achievement. Happiness is at best a secondary consideration. Just perhaps, personal happiness is not even part of the equation.

We all want to be happy, but personal happiness is too shallow a goal, and pursuing it leads to lives that are too shallow. Aristotle recognized this, maintaining that the “good life” was a virtuous life.

He called this “good life” eudaemonia, which today translates to flourishing. (As well as anything translates over 2100 years.)

The pursuit of happiness is a recipe for misery.

What is flourishing?

Flourishing does not dismiss the idea of happiness, but rather adds four more dimensions: spiritual fulfillment, engagement, relationships, and achievement. Taken together, these are a more fulsome way of creating your life.

• Spiritual fulfillment is not a particular belief system, but rather the notion that our work matters, makes a difference, creates value beyond just ourselves, and leaves a legacy.

• Engagement derives from the work of Hungarian psychologist Czikszentmihalyi (cheek-sent-mee-hi), who studied thousands of people in various professions. He proposes that when we stretch our talents to capacity, our sense of self dissolves—we become one with the task. This is a feeling of intense absorption, sometimes referred to as “being in the zone” or “in a state of flow.”

• The relationships dimension acknowledges two things. First, that a principal source of well-being (and of health and longevity to boot) is the quality of our relationships. Second, we all have personal “stakeholders”: spouses, children, extended family, and communities. Our choices in life must serve and respond to more than just our needs. We flourish by giving to and receiving from others.

• Achievement matters to human beings—the sense of pride created by building cathedrals, slaying dragons, and winning at the game of life. Although each of us defines “winning” differently, it nevertheless is an important part of humans’ psychological makeup.

• Finally, we return to happiness - what psychologists call “positive affect,” or happiness. While this remains important, it is rightfully placed alongside things that matter more or equally.

Applying the concept of flourishing to your life

So, how do you make this practical? One way is to use a psychometric instrument and evaluate your life along these dimensions.

And for those brave souls who read to the end, here is a human flourishing at work quizlet. Take the test - see what is missing.

Human flourishing self-assessment

Use the questionnaire to ask yourself the hard questions. Then, create goals that lead you toward the answers you want.

The harsh summary

• Happiness is important, but it isn’t as important as our culture, or self-help gurus would have you believe.

• Pursuing it may actually, paradoxically, make you unhappy as other more important dimensions of the good life (the virtues, eudaemonia, flourishing) are ignored.

• Too much focus on it leads to shallow lives, full of some of the excesses we see around us - consumerism, vulgarity, focus on appearances, short-termism, and selfishness.

  Share

Psychology offers, in caricature, two ways of changing behavior.

“Sticks and carrots” - you focus on incentives, punishment, and reward. You focus on the OUTCOME (which sounds obvious until you think about it.)

The other approach is “hearts and minds” - you persuade, inspire, teach, guide - this focuses on the PERSON.

Neither works - at least not as advertised and certainly not in isolation.

Incentives backfire, may kill motivation, and work in mysterious ways (sometimes a small one works better than a big one.)

The “hearts and minds” approach, even if you are the best persuader, teacher, guider in the world, relies on changing how people think and feel, and praying that they will change behavior - “spray and pray,” I call it.

(We have all had someone agree to something and do something else; we’ve all been inspired by the motivational workshop only to hit Monday morning; we have all set weight goals on Jan 1, and reached for a piece of cheesecake on Jan 15.

Wait, I thought this was about AI!

We’ve had a year of AI “adoption” claims:

• “Everyone’s using it.”

• “No one’s using it.”

• “People are dabbling but not integrating.”

And every change guru on the planet will trot out MINDSET! CULTURE! LITERACY! SKILLS!

Or, you can go the other way, as Accenture appears to do - track logins and tie to performance management. (And here you see the problem - are “logins” what we are after?)

We need behavioral change, but have had no framework for what drives AI adoption.

So here is one with some science behind it.

It’s simple, behavioural, grounded in COM-B (the backbone of modern behaviour change), but with a special callout:

• Trust matters more with AI.

• Frictions require more attention - because businesses are lined with them.

1. Capability — “Can I?”

This is not “basic AI literacy.”
It’s phronesis — practical wisdom applied to AI.

People adopt AI when they have the skills, judgment, fluency, and self-efficacy to use it in real work.
Without capability, motivation dies and trust collapses.

Inside the Capability bucket:

• Can I judge when AI is reliable?

• Can I model tasks and break them down for AI?

• Am I overloaded or do I have cognitive bandwidth?

• Do I know how to verify?

• Am I confident using this in front of others?

Deep adoption cannot happen without deep capability.

2. Motivation — “Or, why the f#$^ should I?”

Motivation is not abstract intention.
It’s the desire to actually reach for AI at the moment of action.

People adopt when AI feels meaningful, relevant, identity-aligned, and worth the effort.

Inside Motivation:

• Clear value (not vague “productivity”)

• Identity relevance (“people like me use this”)

• Leadership role modelling

• Social norms

• Incentives, loss aversion, concrete wins

• Satisficing — the biggest barrier to deep adoption

People stop at shallow use because they get one “good enough” use-case and never explore beyond it.

3. Trust — “Should I?”

Trust is the most misunderstood adoption driver.

It is not a flavour of motivation.
It is not a Capability problem.
It is a psychological state — and AI’s unique psychological footprint makes trust the decisive variable.

People adopt when they feel AI is legitimate, appropriate, safe, non-embarrassing, and aligned with their work identity.

Inside Trust:

• Identity threat (“Will this make me look incompetent?”)

• Permission bias (“Am I allowed to use this?”)

• Legitimacy (“Is this the right tool for this task?”)

• Error tolerance

• Over-trust & under-trust

• Psychological safety

Trust is the green light for adoption.
No trust = no deep adoption.

4. Opportunity / Friction — “Am I enabled or blocked?”

Most failures of adoption are not motivation or capability failures.
They’re friction failures.

If AI isn’t available at the right moment, in the right workflow, with the right defaults, adoption collapses.

Inside Opportunity (and its evil twin, Friction):

• Workflow integration

• Speed, access, latency

• Defaults (opt-out beats opt-in every time)

• Bad UX, bad design, bad placement

• Sludge: bureaucracy, approvals, policy ambiguity

• No time, no support, no data access

Opportunity is the lever leaders ignore most — and the one that moves behaviour fastest.

Why these four matter

Because they give leaders four different levers:

Capability = skills + wisdom

Leadership lever: training, scaffolding, coaching, task modelling.

Motivation = desire + identity

Leadership lever: incentives, framing, role modelling, purpose.

Trust = legitimacy + safety

Leadership lever: transparency, psychological safety, governance clarity.

Opportunity = systems + environment + friction

Leadership lever: workflow redesign, integration, defaults, time.

Every other model muddies these distinctions.

This one doesn’t.

Under the hood

There’s rich behavioural science inside each bucket (identity threat, status quo bias, cognitive load, satisficing, etc.).
But that’s for the book.

For now, remember the simple version:

Everything else is noise.

In 2024, James Healy and I, writing Adopting AI - scoured the world for humanistic approaches to AI - we coined “people-first” and it started as an ethical claim:

• Technology should serve people

But as I’ve worked with business leaders, I’ve come to realize its strategic and tactical importance - see what you think.

We’re in a strange moment in the AI transition. Boards are accelerating investment. Vendors are promising transformation. Consultants are selling roadmaps by the yard. Yet on the ground, inside real organizations, something far more human determines whether any of it works.

Not data.
Not models.
Not compute.

People.

The biggest barrier to AI impact in the 2030s isn’t technical capability. It’s fear, mistrust, and a widening skill gap. The biggest differentiator won’t be who has the fanciest model. It will be who builds organizations where humans and AI actually work well together.

In other words, People-First AI isn’t moral window-dressing.
It’s the next decade’s core competitive advantage.

Let’s unpack why.

1. People-First AI accelerates adoption — not ROI-first

Companies like to talk about AI strategy. But strategy means nothing if employees don’t understand, trust, or use the tools.

Right now, the reality is this:

• Employees fear job loss and surveillance

• Managers don’t know where AI adds value

• Skill gaps slow every initiative

• “Pilot fatigue” kills momentum

• Change programs underestimate the psychological load

You can pour millions into AI tooling. But without human readiness, you get:

Low adoption → low ROI → leadership disappointment → stalled transformation.

People-First AI is the antidote.
It reduces fear, builds capability, and increases the odds that tools become habits.

The companies that win the 2030s will be those that treat adoption not as a footnote to strategy, but as a strategic domain of its own — with investment, leadership attention, and cultural architecture behind it.

Share

2. It builds public trust — and public trust becomes a strategic asset

By 2030, trust will be as important as compute:

• Consumers will punish AI misuse

• AI backlash will crescendo

• Regulators will be unforgiving

• Talent will choose “responsible” employers

• Investors will reward lower regulatory and reputational risk

A People-First approach is simple:

Design AI that is transparent, accountable, and aligned with human welfare.

Starting inside YOUR business - not waiting for the labs to do so.

That doesn’t just satisfy ethical instincts. It strengthens:

• Brand equity

• Customer loyalty

• Social license to operate

Just as sustainability became a market advantage for companies ahead of the curve, trustworthy AI will become the new ESG battleground.

3. It de-risks the business — by reducing the inevitability of AI mistakes

AI creates new failure modes: hallucinations, bias, fragility, model drift, governance gaps, unclear accountability. These risks don’t disappear with better technical controls alone.

They shrink when:

• Employees understand the limits of AI

• Teams know when to override or escalate

• Systems are designed with human oversight

• Ethical guardrails are built from day one

A People-First mindset views AI not as an infallible oracle but as a fallible collaborator — one that needs monitoring, context, and judgement.

Companies that master this won’t avoid all mistakes (no one will). But they will avoid the catastrophic ones.

4. It becomes a competitive differentiator — especially when AI tools become commoditized

Here’s the strategic truth no one talks about:

The tech will converge. Behavior won’t.

By 2030:

• Most companies will have access to similar foundation models

• Most workflows will be automatable

• AI-native tools will level the playing field

So where will the advantage come from?

From how well humans and AI actually integrate within your culture, workflows, and leadership practices.

High-trust, high-skill, high-adoption organizations will:

• Automate faster

• Innovate faster

• Reduce cost more sustainably

• Deploy with fewer failures

• Move from augmentation to genuine transformation

Your competitors can copy your tech stack.
They cannot easily copy your culture, your norms, your capability, or your people.

In the CSR and triple-bottom-line frameworks I’ve argued for over the years, the same logic applies: treating people well isn’t just morally praiseworthy; it is strategically superior over the long run. People-First AI follows this exact pattern.

The bottom line: People-First AI is not soft — it’s smart

Organizations that treat AI as a purely technical upgrade will fall behind.

Organizations that invest in:

• Human capability

• Psychological safety

• Ethical design

• Transparent governance

• Trust building

• Continuous learning

…will outperform those who don’t.

This is the paradox of the AI age:

The more powerful the machines become, the more valuable your people become.

The “People-First AI” organization is not more humane instead of being more competitive.
It is more humane because that is what makes it more competitive.

And by the 2030s, that will be impossible to ignore.