What should we tell employees about AI and headcount?

Tell them which decisions a machine may make in the organisation’s name, which it may not, and what you will and will not do with the time it saves. Deloitte’s survey of 25,000 UK workers, published on 16 September 2026, found 31 per cent using generative AI without their employer’s knowledge and 64 per cent of weekly users worried that their manager will conclude AI can do their job. Those two figures are one finding. A worker who believes the tool is a case for their own redundancy hides the tool, and an organisation whose people hide the tool cannot see which decisions are being handed to a machine, cannot check them, and would not know if one went wrong. The headcount question is a control question before it is a communications one.

What the survey measured#

Ipsos UK ran the fieldwork for Deloitte’s UK GenAI Workforce Survey between 7 May and 10 June 2026, online and weighted, among 25,000 workers aged 18 to 70. Deloitte reports that 63 per cent use generative AI for work. Of the users, 31 per cent do so without their employer’s knowledge, 23 per cent perceive a stigma around using it, and 64 per cent of weekly users worry that their manager will think the tool can do their job. About half have had no formal training in using it safely, and 65 per cent say leadership guidance is not convincing. Hayley McKelvey, Deloitte UK’s chief AI officer, is quoted in the release: “Workers who feel stigma around adopting GenAI tools are more likely to conceal use.” Bloomberg, reporting the survey on 16 September, added that almost one in ten had used tools their employer had banned, and quoted David Chismon of the National Cyber Security Centre saying that IT teams “should not assume they are seeing the full picture”. The figures are self-reported and Deloitte advises on AI adoption; what the survey establishes is direction and order of magnitude.

The penalty has been measured, twice#

In four preregistered experiments with 4,439 participants, published in PNAS in 2025, Jessica Reif, Richard Larrick and Jack Soll found that people described as using AI for a task were judged lazier, less competent, less diligent and less independent than people described as getting the same help by other means. Managers who did not use AI themselves preferred candidates who reported no AI use. The penalty shrank among evaluators who used AI weekly or daily, and disappeared when the tool’s usefulness for the task was made explicit. Asked directly, participants said they would be reluctant to disclose their own use. The experiments ran between March 2024 and February 2025, and the authors say the penalty may shift as use becomes ordinary.

Microsoft and LinkedIn’s 2024 Work Trend Index, a survey of 31,000 knowledge workers in 31 countries by a company that sells the tools, found 52 per cent of AI users reluctant to admit using it for their most important tasks and 53 per cent worried that doing so made them look replaceable. Two years apart, different sponsors and different samples, and the same half of the workforce says the same thing.

What concealment removes#

The cost of hidden use is usually described as a data risk: the National Cyber Security Centre wrote on 7 September that “you cannot manage what you do not know” and that the response it recommends is to understand why people reach for the tools and provide secure alternatives. The larger cost is to judgement. Rules Before Tools puts four questions to any deployment: which decisions the machine may make, who can stop each one, what people must remain able to do unaided, and how anyone would know if it went wrong. None of the four can be answered about a decision the organisation does not know is being made. A third of users working unseen means a third of the delegation happening off the map: nobody has decided what may be delegated, nobody is checking it, and the first the organisation hears of a bad output is when a client or a regulator finds it.

Why silence is heard as a plan#

Pew Research Center published on 17 September a survey of 42,151 people in 37 countries, fielded between February and May 2026: in 34 of the 37, majorities expect AI to cause job losses rather than growth, and in high-income countries a median of 55 per cent expect fewer jobs in twenty years, about seven in ten in the United States, Australia and South Korea. Axios reported in August that chief executives who had announced AI-driven cuts in 2025 now describe the same programmes as transformation; what became of several of those cuts is at what happened to the companies that cut staff for AI. Where the leadership supplies no intention, employees supply one from the news, and the one they supply is the worst case. The evidence on actual displacement is more mixed than the expectation, as will AI replace entry-level jobs sets out, but expectation is what drives concealment.

What to say, and what saying it commits you to#

Two things, both of which bind the leadership rather than the worker. The first is the decision rights: which work a machine may draft, which it may decide, which it may never touch, and what a person must check before it leaves. Written at that level, the policy turns “are you allowed to use AI” into “is this decision yours to hand over”, a question people will answer in the open. The second is the headcount position, declared by domain and dated. If time saved by AI will not be used as evidence in a restructuring for a stated period, say so and be held to it. If roles in a function will go, say which and when. Most weekly users already fear the cut; a feared cut with no date drives every user underground, and a stated one at least tells them where they stand. Trust tracks whether the declared position and the business case agree. If the slides say augmentation and the case counts heads, people work out which is real.

Then make declaring safe: a period in which anyone can say which tools and prompts they use, without penalty, produces the map of actual use, and how to run one is at what to do when people work around the AI policy. And train the half who have had nothing. A person who has never been shown how to check an output is not overseeing anything, whatever the policy says; how do you measure AI adoption properly covers how to tell whether that has changed.

The question for a board#

What has been said to employees about AI and headcount, in writing, and does it agree with the operating plan the board approved? And how much of the organisation’s work is now done with tools the organisation cannot see? If the second answer is a guess, the first is the reason. On Deloitte’s figures a UK employer with 10,000 staff should assume that close to 2,000 of them use generative AI on its work without its knowledge, half of them untrained in checking what it produces. A capability audit that ignores hidden use measures the wrong organisation.

What this does not show#

It does not show that a stated headcount position reduces concealment; no study has tested that, and the two commitments above are argued from the mechanism. The Deloitte figures are self-report from an online sample, and the 31 per cent measures use the employer does not know about, which is not the same as use the policy forbids. The PNAS experiments put hypothetical workers to online participants, mostly in the United States, and the penalty may already be smaller. Pew measures what people expect, not what has happened to jobs. What it does show is that close to a fifth of UK workers hand work to a machine where nobody in the organisation can see it, that the fear driving that has been measured twice and is rational, and that the fear is about the leadership’s intentions, which the leadership can state.