The people driving your AI transformation don't have permission.

Most organisations are running two AI transformations at once, and only one of them is on the agenda. The visible one happens in board meetings and strategy decks: a working group is convened, a roadmap is commissioned, the implications are studied, and six months later a report arrives and nobody’s working day looks any different. The other has already happened, quietly, without anyone signing it off.

Somewhere in the same building, a few people have rebuilt how they work around these tools so completely that going back would be like trading a car for a horse. They did it without asking, having worked out early that permission moves slower than forgiveness, and they have kept it to themselves. These people are the second transformation, the one nobody convened, and most of the organisation has not noticed them.

The gap is not between the company and its competitors. It runs inside the building, between the people who have already changed how they work and the people still waiting to be told how, and it widens on its own, because the first group gets better at this every month while the second waits for instructions. Those people are the most valuable thing an organisation has in this transition, and most employers are wasting them.

They are not all in engineering, though some are; the others are in product, design, marketing, operations, client services. The common thread is not technical skill but curiosity and a willingness to act without being asked. They tried the tools, stayed past the first month of frustration, and kept reshaping their own work until the AI became load-bearing. They give themselves away by output: they finish faster than the process expects, and produce more than the role allows. They are also, often, evasive about how, because they have learned that saying “I used AI for that” invites a conversation they would rather avoid.

The evasiveness is rational. In the largest global study of the question, run across forty-seven countries by the University of Melbourne with KPMG, fifty-seven per cent of employees said they hide their use of AI and pass the output off as their own.1 They have reason to. A controlled study published in PNAS found that workers who disclose AI use are judged by colleagues and evaluators as lazier and less competent, and are penalised in hiring and assessment decisions, with the work itself held constant.2 So the people getting the most from these tools conceal it, and the organisation, without ever deciding to, has built an incentive to hide its most productive behaviour.

The instinct, when this surfaces, is to tighten control, and the instinct is not wrong. There are real exposures around client data, intellectual property, security and compliance, and a blanket policy promises a safety that case-by-case judgement cannot.

These tools are not uniformly good. The largest field experiment on knowledge work, run with BCG’s consultants, found large gains on tasks within the AI’s reach, twelve per cent more tasks completed, a quarter faster, at higher quality, but on a task deliberately chosen to sit outside that reach, consultants using AI were about nineteen percentage points more likely to get the answer wrong.3 The frontier is jagged, and someone has to know where its edges run.

The jagged frontier argues for the opposite of a ban. A blanket ban stops the risky use and the productive use with it. And the people it restricts are usually the ones who understand the edges best, because they have spent months finding them, while the policy is written by people who have not. The result protects the organisation from a theoretical harm and guarantees a real cost: the people who could lead the transition either work around the rule, which is worse than no rule, or take their energy elsewhere, to a side project, or to another employer.

The better approach inverts the default. Instead of naming the approved uses and forbidding the rest, name the places AI must not go, and grant broad permission everywhere else. The first builds a whitelist that is permanently out of date and needs sign-off for every new use. The second draws a short list around the things that genuinely need protecting, client data, personal information, security-critical systems, whatever the organisation cannot afford to expose, and leaves people free to change how they work outside it. Define the boundaries, say plainly why each one is there, and step back.

This takes trust, which is uncomfortable, because it means accepting that no one controls the detail of how these tools get used. That control is already gone. Only forty per cent of organisations have any policy on generative AI at all, and the use is happening regardless.1 The choice is between trust with clear boundaries and the appearance of control.

Permission removes the barrier. Visibility does the rest. When the most productive person on the team is openly the heaviest user of these tools, “I used AI for that” stops being a confession and becomes the reason colleagues come to ask how. The stigma the research documented inverts: in the open, heavy use becomes a mark of competence, because the results are sitting right there. The practice spreads sideways, from the people already proving what is possible to the people beside them, on the strength of results their colleagues can see for themselves.

Visibility also answers the objection that broad permission invites. The honest fear is a flood of plausible, unchecked output, the jagged frontier’s failures shipped as finished work. But work done in the open has to stand up in front of someone other than the person who made it. The same exposure that spreads the good practice catches the bad. Visibility is the quality mechanism.

Clamping down carries a cost beyond the lost productivity. The people who have already remade their own work are also the most mobile. They know what they can do, and they know the market has noticed. When they look at the distance between how they work and how the organisation around them works, the friction is real, and an employer that responds by slowing them down has handed them a reason to leave. The evidence runs the same way: in the customer-support study, the agents given a capable AI assistant resolved more issues and were measurably less likely to quit.4 The cheapest way to keep the most adaptive people is to let them be adaptive.

None of this starts with a programme. It starts with noticing that the people who will carry the transformation are already carrying it, without sanction and mostly out of sight. They never asked permission, and they are not going to. The only decision left to the people above them is whether to grant it openly, and get the visible, compounding version of what these people can do, or to withhold it and keep the hidden, hoarded one, while believing the transformation is still theirs to direct. It is not. Someone in the building took it months ago, and has been using it quietly ever since.

---

References

1. Hiding and the policy gap. University of Melbourne and KPMG, “Trust, attitudes and use of artificial intelligence: A global study 2025” (48,340 respondents across 47 countries, surveyed November 2024 to January 2025): 57% of employees said they hide their use of AI and present AI-generated work as their own; 66% use AI with some regularity; only 40% report any workplace policy or guidance on generative AI. https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-attitudes-and-use-of-ai.html . Narrower single-market surveys put unsanctioned-use figures anywhere from roughly a quarter (1Password, ~27%) to around three-quarters (WalkMe, ~78%); the spread reflects differing definitions and samples. ↩ ↩²

2. Reif, Larrick and Soll, “Evidence of a social evaluation penalty for using AI,” Proceedings of the National Academy of Sciences 122(19), 13 May 2025 (four experiments, 4,400+ participants): people who disclose AI use are judged by peers and evaluators as lazier and less competent, and are penalised in evaluation and selection decisions with the work held constant. https://www.pnas.org/doi/10.1073/pnas.2426766122 ↩

3. Dell’Acqua, McFowland, Mollick and colleagues, “Navigating the Jagged Technological Frontier,” Harvard Business School Working Paper 24-013 (2023), a field experiment with 758 BCG consultants: on tasks within the AI’s reach, consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, with measurable quality gains; on a task chosen to fall outside that reach, they were about 19 percentage points less likely to reach the correct answer. https://www.hbs.edu/faculty/Pages/item.aspx?num=64700 ↩

4. Brynjolfsson, Li and Raymond, “Generative AI at Work,” Quarterly Journal of Economics 140(2)

   (2025); NBER Working Paper 31161 (2023). Across 5,179 customer-support agents, access to an AI assistant raised issues resolved per hour by 14% on average and 34% for the least experienced, with little effect on the most experienced; access was also associated with lower staff attrition. https://www.nber.org/papers/w31161 ↩