Why Your AI Pilot Is Stuck: It's Not the Skills Gap

The dominant explanation for stalled enterprise AI — skills gap, training, change management — is a partial story. AI's first-order effect inside an organisation is to dissolve the informational moats that the formal hierarchy was built to defend, and the resistance to it is best read through Crozier, Pfeffer and Mintzberg rather than through another training programme.

---

The data does not match the story

MIT's NANDA initiative, in The GenAI Divide: State of AI in Business 2025, looked at more than three hundred enterprise GenAI initiatives, fifty-two interviews and a survey of one hundred and fifty-three senior leaders. The headline finding was that ninety-five percent of those initiatives produced no measurable P&L impact, despite cumulative enterprise spend that NANDA places somewhere between thirty and forty billion dollars. McKinsey's State of AI 2025 triangulates with the same picture from a different angle: eighty-eight percent of organisations now report using AI in some form, and six percent qualify as high performers. The gap between near-universal adoption and rare measurable impact is the entire phenomenon to be explained.

The standard explanation is some combination of three things. The workforce lacks the skills. The change-management programme is immature. The data isn't ready. Each of those gets cited in roughly every consulting deck on enterprise AI written in the last two years, and each has enough surface plausibility to keep the conversation going. The trouble is that the empirical picture has stopped fitting them.

BCG's AI at Work 2025 surveyed 10,600 employees across eleven countries and found that managers report seventy-eight percent regular use of AI tools versus fifty-one percent for frontline workers. The layer of the company that is supposed to be the gap — the people the training programme was supposed to lift — is not the layer that has the lower fluency. Managers are more fluent than the workers they manage. Rollouts still stall.

McKinsey's Superagency in the Workplace 2025 report, in a quietly remarkable turn for a firm whose business model is selling change-management programmes, reverses years of their own framing. The exact line: the biggest barrier to scaling is not employees, who are ready, but leaders, who are not steering fast enough. When the consultancy that invented "the people side of change" tells you that the people side is no longer the constraint, that is worth pausing on.

What Crozier saw in 1964

Michel Crozier published The Bureaucratic Phenomenon in 1964, in the middle of the post-war French civil service, and the book has been quietly correct about every wave of organisational technology since. Crozier's central observation is the idea of zones of uncertainty. In any rule-bound organisation, there is a residual set of pockets where outcomes are not predetermined by the formal system — places where the rule-book runs out and someone has to make a judgment call using information the rule-book cannot codify. Whoever controls those zones, typically through specialist information that nobody outside the zone can access, holds disproportionate influence regardless of where they sit on the formal chart.

The org chart, in this reading, was never really a map of accountability. It was a map of who controlled which uncertainty zone. The senior person was senior because they could see across two or three departments worth of detail that the people below could not see. The vendor selection went up the chain because the people at the top held supplier history and strategic intent that wasn't accessible elsewhere. The hiring decision needed sign-off because the next level had visibility into headcount plans nobody below had seen.

AI's defining property is that it collapses zones of uncertainty. When everyone in the company can pull a synthesis across the whole codebase, the whole CRM, the whole supplier base or the whole policy stack in thirty seconds, the moat around any one role's informational authority stops doing the work it was built for. Crozier predicted the next move with no difficulty: the people whose authority depended on those zones would resist, and they would rationalise the resistance through the formal rule system itself. Compliance reviews. Governance frameworks. Data quality concerns. "Let's run a pilot first." "We need to understand the risks better." Each of those objections can be entirely sincere on its face and still function, in aggregate, as the immune response of a structure whose internal power-distribution is being attacked.

Pfeffer, in Power in Organizations (1981) and again in his California Management Review essay "Understanding Power in Organizations" (1992), reaches the same place from a different direction. Power flows to those who control critical resources, and information is among the most critical resources inside any large firm. Mintzberg's Power In and Around Organizations (1983) goes further: the informational roles he catalogues — the monitor, the disseminator, the spokesperson — are precisely the roles AI is most efficient at automating away. Connelly and colleagues' "Knowledge Hiding in Organizations" (Journal of Organizational Behavior, 2012) gives the behavioural taxonomy of how this resistance actually shows up — evasive hiding, rationalised hiding, and playing dumb being the three patterns that recur — and every one of them is recognisable in the way pilots get politely smothered inside large companies.

Brynjolfsson's "The Turing Trap" essay (Daedalus, 2022) is the macroeconomic version of the same point. AI's distributive effects depend on whether the technology is used to augment human work or substitute for it, and the choice between those is political, not technical. The architecture of the deployment determines who wins and who loses, which is exactly why the deployment becomes the political fight.

The empirical evidence the prediction was right

NANDA's most striking sub-finding is buried in the body of the report: internally-built AI systems succeed at roughly one-third the rate of external vendor partnerships. Internal builds get politicked through the organisation. External vendors don't, because the vendor is a black box that arrives with the redistributive question already settled. The statistic is the political economy of the rollout in numerical form.

BCG found that forty-three percent of managers expect their job to disappear within ten years, against thirty-six percent of frontline workers. The layer most exposed to the technology is also the layer with the strongest individual incentive to slow it down. The same survey found only twenty-three percent of middle managers feel prepared for AI rollouts, even though they use AI more than the people they manage — a perception gap that fits the territoriality reading better than it fits a pure preparedness reading.

In November 2025, Li, Zhu and Hua published "Overcoming the Organizational Barriers to AI Adoption" in Harvard Business Review. Their finding, drawn from interviews and case work across multiple sectors, names legal, HR, risk and compliance as the most frequent blockers — staff functions with the structural authority to slow projects regardless of executive support. Each function cites real concerns. Liability exposure is real. Change management is real. Regulatory uncertainty is real. The pattern that needs explaining is not whether the concerns are sincere but why the same four functions appear at the same friction point with such consistency, and why the magnitude of the response so often exceeds the magnitude of the actual risk on the file.

Asana's State of AI at Work 2025, surveying 9,236 knowledge workers across five countries, names the resulting condition the two-speed workplace. Fifty-two percent of executives use AI weekly. Thirty-six percent of the broader workforce does. Executives believe transformation is well under way; the layer between them and the frontline absorbs the friction and reports it back upward as "we're making progress on the pilot."

Shadow AI is the behavioural counter-signal

The single most informative statistic in the dataset is the one about employees who use AI without telling anybody. Microsoft's Cyber Pulse research, conducted with Hypothesis Group across 1,725 data security professionals, found twenty-nine percent had used unsanctioned AI agents at work. ISACA's 2025 figures put the same behaviour as high as seventy-one percent in some sector cuts. Slack's Workforce Lab Fall 2024 Workforce Index found that forty-eight percent of employees would not feel comfortable telling their manager they had used AI for a piece of work; the top reasons cited were "feels like cheating" and "fear of being seen as less competent or lazy." KPMG and the University of Melbourne's Trust in AI: Global Study 2025 found that fifty-seven percent of workers globally present AI-assisted output as their own, and roughly half of the US workforce reports using AI at work without knowing whether they are permitted to.

If the binding constraint on enterprise AI were really a skills gap, you would expect employees to be waiting for training. They are not waiting. They are routing around the gatekeepers, quietly, at scale, often hiding the fact from the same managers whose objections are formally holding the rollout up. Slack also found that workers who feel comfortable using AI are sixty-seven percent more likely to actually incorporate it into their workflow — which means the binding constraint is the permission climate, not the capability. The behavioural evidence makes the case the survey data only implies. The constraint is the manager's objection, not the worker's capacity.

The patterns are specific and recognisable

The structure that gets stood up to enable AI adoption — the Centre of Excellence — is in practice the structure most likely to slow it down. CoEs become the bottleneck they were designed to prevent, because they end up holding the gate on behalf of the staff functions they were meant to coordinate. Reported failure rates for CoE-led AI programmes sit roughly where overall enterprise-AI failure rates sit, around the high seventies to mid-eighties depending on the survey — a convergence that stops being a coincidence the moment you read the structure as a control mechanism rather than as an enabler.

Pilot purgatory — the well-documented condition of running successive proofs of concept that never graduate to production — reads more usefully as a redistribution-avoidance pattern than as a technical one. A pilot that stays in pilot does not change the existing decision rights inside the organisation. A pilot that moves into production does. Permanent pilot status preserves optionality for the people whose authority is on the line; production status takes the optionality away. The same three or four objections recur at the production-readiness gate across firms that have nothing else in common, which is the signature of a structural pattern rather than a technical one.

Legal, HR, risk and compliance — the four functions HBR named — are not bad-faith actors. Each cites real and serious concerns, and each is also defending professional territory; the two are not mutually exclusive, which is precisely why the pattern is hard to name out loud. IT departments, often through no fault of their own, end up moving slower than the work they are supposed to enable, and the workforce notices and adapts. The IT-leader surveys that have looked at this find large majorities — well above two-thirds in most cuts — reporting both unsanctioned AI use inside their own perimeter and organisational silos that slow execution down. The shadow-AI numbers are the workforce's read on that gap, expressed in behaviour rather than survey responses.

The middle-manager information-brokerage layer is under direct attack and the public examples are now numerous. Klarna's customer-service-and-headcount story through 2024–2025, Tobi Lütkin's "AI-first" hiring memo at Shopify in April 2025, Duolingo's contractor restructuring around the same period — each is a public version of the private restructuring conversation happening inside many large firms, and each is also a signal to the middle of the organisation that the layer brokering information up and down the chart is the layer the technology is most efficient at compressing.

Where this hypothesis runs out

The hypothesis is real but it is not the whole story. Greenfield AI-native firms still have politics — OpenAI's and Anthropic's well-publicised governance crises are evidence enough that removing the legacy structure does not remove the territorial dynamics. Data quality problems are sometimes genuinely the blocker, particularly in firms whose data infrastructure was built for a different operating model. Regulatory uncertainty is genuinely chilling for legal and compliance functions, and the EU AI Act's Article 4 deadline this August is one good reason that's so. Sectoral variation matters: healthcare, defence and public-sector procurement face genuine constraints that an enterprise software vendor does not face. The skills-gap framing has some truth in it, particularly deeper into the workforce.

The cleanest version of the claim is that the political layer is under-recognised in the dominant narrative, not that it is the only thing that matters. But under-recognised is enough, because it changes what you do next.

The diagnostic question

The question worth asking before another training programme is: if we ran this rollout assuming the resistance is structural rather than cognitive, what would we do differently? That single reframe shifts the operational answer in several places at once.

You stop interpreting "we need a pilot first" as a request and start interpreting it as a position — a position which may also have a real concern attached to it, but which is a position regardless. Once "let's run a pilot first" is read as a position rather than a neutral request, you can negotiate it as such, including the question of what evidence would actually move the position and what evidence is being requested precisely because it cannot be produced. You stop adding governance layers when the evidence says the existing governance layers are already holding the gate, and you start measuring whether judgment is actually being held at each node — which is a different and harder measurement than whether a process exists. You stop pouring training budget into the layer of the workforce that is already adopting (your frontline, who are using AI in the shadow numbers) and start auditing the layer that is slowing the rollout (your managers and your staff functions, who have the structural authority and the strongest individual incentive). You pair every AI deployment with an explicit redistribution-of-decision-rights statement, so the political question is on the table rather than under it. The political question is being negotiated whichever way you go; the only choice is whether the negotiation is visible.

For founders running smaller organisations, the implication is simpler. The reason a fifteen-person company integrates agentic tooling in three weeks while a ten-thousand-person firm has spent eighteen months in pilot is not that the smaller team is more skilled. It is that the smaller team has fewer load-bearing zones of uncertainty for the technology to threaten, and fewer professional functions whose territorial defence runs through the formal rule system. The relative speed advantage of the lean team is not a technology advantage. It is a political-economy advantage, and it has a clock on it — every successful AI-native team that ships product into a market the incumbent is still piloting in is one more piece of evidence that the incumbent's internal architecture is no longer competitive on speed.

For CHROs and AI leads inside larger firms, the implication is more uncomfortable. The standard internal narrative — "we're working on the skills gap" — is unlikely to be challenged in the C-suite, because the alternative narrative ("we have a redistribution fight inside our staff functions and our middle management layer") is harder to deliver and harder to action. The longer the standard narrative holds, the longer the rollout sits in the gap McKinsey's own framing has now shifted underneath. At some point, the gap closes by force — usually because a competitor has crossed it first and the C-suite stops accepting the standard narrative — and the organisations that crossed it earlier did so by treating the political question as a first-class problem, not as a downstream effect of "change management."

What replaces the dissolved structure

Beane's The Skill Code (HarperBusiness, 2024) — and his earlier "Shadow Learning" paper in Administrative Science Quarterly (2019) — describes what happens to apprenticeship inside organisations when a new technology disrupts the expert-to-novice handoff. Novices route around the sanctioned learning paths, and the organisation stops being able to see where competence is being built and where it is quietly atrophying. Beane's frame is the closest thing in the academic literature to the operational problem the post-AI organisation is now living with: the architecture that made competence visible has stopped working, and nobody has built the replacement.

The architecture that replaces the dissolved org chart is not going to be built by the people whose authority depended on the old structure. It will be built by founders, by technical leaders, and by the smaller number of senior executives who have decided to give up the informational moat in exchange for an organisation that can actually move at the speed AI makes possible. That decision is not a training problem and it is not a change-management problem. It is a structural choice about what the company is going to put weight on now that the old foundations have stopped bearing load.

I have written elsewhere about what the replacement structure looks like — the competence net, the load-distributed layer of judgment-bearing humans whose presence is what actually makes "AI at the centre" work rather than break — and an open standard called TwinLadder (twinladder.ai) that we are building as the measurement instrument for it. That is the architecture worth measuring, and the instrument is the architectural drawing of the company you are rebuilding now that the old barriers have stopped doing their work.

The pilot is not stuck because of a skills gap. It is stuck because the rollout is changing who controls which decision, and the people whose control is changing have the structural authority to slow it down. The first useful step, before another training cohort, is to read the resistance for what it is.

---

Sources

1. MIT NANDA, The GenAI Divide: State of AI in Business 2025 — project.nanda.media.mit.edu — 300+ initiatives, 52 interviews, 153 senior-leader survey; ninety-five percent of enterprise GenAI pilots produced no measurable P&L impact; internal builds succeed at one-third the rate of external vendor partnerships.
2. BCG, AI at Work 2025 — bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain — survey of 10,600 employees across eleven countries; manager AI use seventy-eight percent vs frontline fifty-one percent; forty-three percent of managers expect their job to disappear in ten years vs thirty-six percent of frontline.
3. McKinsey, The State of AI 2025 — mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai — eighty-eight percent of organisations using AI; six percent are high performers.
4. McKinsey, Superagency in the Workplace 2025 — mckinsey.com/capabilities/mckinsey-digital/our-insights/superagency-in-the-workplace — "the biggest barrier to scaling is not employees, who are ready, but leaders, who are not steering fast enough."
5. Li, Zhu and Hua, "Overcoming the Organizational Barriers to AI Adoption," Harvard Business Review, November 2025 — hbr.org — names legal, HR, risk and compliance as the most frequent blockers with structural authority to slow projects regardless of executive support.
6. Slack Workforce Lab, Fall 2024 Workforce Index — slack.com/blog/news/fall-workforce-index-2024 — forty-eight percent uncomfortable telling manager about AI use; comfortable workers sixty-seven percent more likely to incorporate AI.
7. KPMG and University of Melbourne, Trust, attitudes and use of AI: Global Study 2025 — kpmg.com/au/en/home/insights/2025/04/trust-attitudes-and-use-of-ai-global-report-2025.html — fifty-seven percent globally present AI-assisted work as their own.
8. Microsoft Cyber Pulse / Hypothesis Group 2025 — microsoft.com/en-us/security/business/security-insider — survey of 1,725 data security professionals; twenty-nine percent reported using unsanctioned AI agents.
9. Asana Work Innovation Lab, State of AI at Work 2025 — asana.com/resources/state-of-ai-at-work — 9,236 knowledge workers across five countries; "two-speed workplace"; fifty-two percent executives use AI weekly vs thirty-six percent broader workforce.
10. Crozier, Michel, The Bureaucratic Phenomenon (University of Chicago Press, 1964) — original statement of the zones-of-uncertainty thesis on which the political-economy reading rests.
11. Pfeffer, Jeffrey, Power in Organizations (Pitman, 1981) and "Understanding Power in Organizations," California Management Review 34(2), 1992 — doi.org/10.2307/41166692 — power flows to those who control critical resources, including information.
12. Mintzberg, Henry, Power In and Around Organizations (Prentice-Hall, 1983) — informational roles (monitor, disseminator, spokesperson) as the substrate of organisational power.
13. Connelly, C. E., Zweig, D., Webster, J. and Trougakos, J. P., "Knowledge Hiding in Organizations," Journal of Organizational Behavior 33(1), 2012, pp. 64–88 — doi.org/10.1002/job.737 — three behavioural patterns: evasive hiding, rationalised hiding, playing dumb.
14. Beane, Matt, The Skill Code: How to Save Human Ability in an Age of Intelligent Machines (HarperBusiness, 2024); and "Shadow Learning: Building Robotic Surgical Skill When Approved Means Fail," Administrative Science Quarterly 64(1), 2019 — doi.org/10.1177/0001839217751692 — what happens to apprenticeship and competence visibility when a new technology disrupts expert-to-novice handoffs.
15. Brynjolfsson, Erik, "The Turing Trap: The Promise and Peril of Human-Like Artificial Intelligence," Daedalus 151(2), 2022 — doi.org/10.1162/daed_a_01915 — augment-vs-substitute as the political-economic choice that determines distributive outcomes.
16. TwinLadder Standard v1.0 — twinladder.ai — open standard (CC BY-SA 4.0) for AI competence measurement; the work-in-progress instrument for the architecture this piece points toward.