Delegate the Drafting, Never the Deciding

On Tuesday 8 September 2026, Forbes ran a piece under the headline "Managers Who Use AI Risk Losing Employee Confidence And Trust", syndicated the same day by Yahoo Finance. It rests on two 2026 surveys from Omni Calculator: in one, of 913 US employees, 42 per cent said they view managers who rely on AI as less competent and 36 per cent said they trust such managers' judgement less. Sixty-eight per cent said AI-assisted messages feel less personal. The advice that follows is to keep AI away from performance reviews, strategy and anything sensitive, and to reserve it for "supporting work".

The headline is doing something the evidence does not. It is worth being precise, because the facts here matter more than the noise around them.

What the survey actually measured

Read the questions rather than the headline and a different picture appears. Half of respondents said their manager had used AI to answer a question or make a decision "that they felt the manager should have handled personally". Forty-five per cent said the clearest giveaway is a message that does not sound like the manager's usual voice; 40 per cent pointed to a generic or robotic tone. Twenty-seven per cent had seen AI used to prepare performance reviews or feedback, 26 per cent to set strategy or goals, and 18 per cent to make workplace decisions.

None of that is a reaction to a tool. It is a reaction to a manager who has handed over the part of the job that was theirs to own: the judgement, and the voice it is delivered in. An employee who receives a review in a register they have never heard from their manager is not objecting to a piece of software. They are noticing that nobody decided anything, and that the person accountable for the decision has stepped out of the room.

Two caveats on the source itself, because a figure published under my name should survive a reader checking it. The survey is of US employees, so it says nothing directly about a UK workforce. And at the time of writing I could not locate the underlying report on Omni Calculator's own research pages, where its other 2026 studies are published; the figures above are as reported by Forbes on 8 September and, independently, by HR Katha on 9 September. They are consistent across both, but they are a calculator company's survey, not a peer-reviewed study, and they deserve to be weighed as such.

The peer-reviewed evidence says something more useful

The stronger evidence has been in print for over a year. In May 2025, Jessica Reif, Richard Larrick and Jack Soll of Duke University published "Evidence of a social evaluation penalty for using AI" in the Proceedings of the National Academy of Sciences: four preregistered experiments, 4,439 participants. People who use AI at work "anticipate and receive negative evaluations regarding their competence and motivation", and those evaluations carried through into hiring decisions. So far, so Forbes.

But the same paper contains the finding that changes the conclusion. Managers who rarely used AI themselves were less likely to hire candidates who used it daily, while managers who used AI frequently showed the opposite pattern and favoured candidates who did the same. The penalty is an attribution about effort, and it lives in the evaluator, not the tool. As AI use spreads through a workforce, the population doing the judging changes, and the penalty shrinks with it. What the Omni survey captured is a snapshot of a transition, not a law of management.

There is a second paper every executive should have read before reaching for concealment as the answer. Oliver Schilke and Martin Reimann of the University of Arizona, in Organizational Behavior and Human Decision Processes in 2025, ran thirteen experiments and found that "actors who disclose their AI usage are trusted less than those who do not". That sounds like an argument for keeping quiet. It is the reverse. The same research found that an actor exposed for using AI is trusted less again than one who disclosed it. Disclosure costs something. Being caught costs more. The worst position available is the one the Forbes piece quietly nudges managers towards: use it, say nothing, and hope the voice does not give the game away. Forty-five per cent of employees say it already does.

The UK angle, which is the one that matters here

A UK reader has a sharper data point than a US survey. Careerminds UK asked 600 full-time UK employees in August 2026 who they would rather receive performance feedback from: 74.2 per cent chose their line manager, 14.2 per cent an AI system, and 11.7 per cent had no preference. Three quarters of British employees want the judgement on their work to come from a person who is answerable for it. Not from a person reading out something a system produced.

And the UK regulator has already written the rule for this, in a context close enough to borrow. On 31 March 2026, the Information Commissioner's Office published its report on automated decision-making in recruitment. Its central finding is that "human involvement must be meaningful and active. It cannot be a token gesture or a rubber stamp of an automated outcome." The test the ICO sets is whether a human "can exercise real influence over a decision before it is applied and has the authority, discretion and competence to alter it". That is a legal standard for solely automated decisions under UK GDPR, and it does not formally reach a manager who drafts a review with a chatbot. But as a management standard it is exactly right, and it is simpler than anything in the Forbes piece. The Business and Trade Committee's inquiry into AI, business and the future of the workforce, which took oral evidence through 8 July 2026 and has not yet reported, is circling the same question of where accountability sits when the work is automated.

Meanwhile the adoption gap is real and widening. Gartner's survey of 2,986 employees, published in March 2026, found 46 per cent of managers experimenting with AI to improve their work against 26 per cent of employees. Managers are ahead of their teams. On the Duke evidence, that is precisely the configuration in which the penalty bites hardest: a manager using a tool that the people judging them do not yet use themselves.

The rule that replaces the fear

The Forbes prescription is a list of tasks AI should not touch. That is the wrong shape of rule, because tasks are not the unit of trust. Accountability is. So here is the rule I would give any leadership team, and it fits on one line: AI may draft the evidence, and the manager must own the judgement and deliver it in their own voice.

In practice that means three things. First, a manager may use AI to gather, summarise and structure the material behind a review, a plan or a difficult conversation, and should, because the alternative is a manager doing worse preparation in less time. Second, the conclusion is theirs. If they cannot say why the rating is what it is, in their own words, without the draft in front of them, they have not made a decision, they have forwarded one. That is the ICO's rubber stamp test applied to a person rather than a system. Third, the AI policy is declared once, at team level, as a fact about how the organisation works, so that no individual use is ever a confession. The Arizona research is clear that exposure is the expensive outcome, and a standing policy removes the possibility of exposure entirely.

I have argued elsewhere that a chatbot giving a weak answer is a quality issue, while an agent acting on weak data is a control issue. The manager who forwards an unowned decision has crossed the same line. The output may be polished. The control has gone.

What to do on Monday

Three things, none of them complicated. Ask every people manager to write down which parts of their job they are already using AI for, and publish the combined list to their teams as policy rather than leaving it to be discovered. Adopt the ICO's wording as the internal standard for any AI-assisted decision about a person: meaningful, active, and alterable by the human whose name is on it. And test the voice: if a message to a team member would not survive the question "did you write this", it is not ready to send.

Employees are not asking their managers to stop using AI. Three quarters of them are asking for something older and simpler: that when a judgement is made about their work, a person who is accountable for it is the one making it. That is the heart of it. Delegate the drafting. Never delegate the deciding.

Sources

Forbes via Yahoo Finance UK, Michelle Travis, "Managers Who Use AI Risk Losing Employee Confidence And Trust", 8 September 2026: https://uk.finance.yahoo.com/news/managers-ai-risk-losing-employee-124500218.html. HR Katha, 9 September 2026: https://www.hrkatha.com/news/when-managers-use-ai-42-employees-question-their-competence-survey/. Reif, Larrick and Soll, "Evidence of a social evaluation penalty for using AI", PNAS 122(19), 8 May 2025: https://doi.org/10.1073/pnas.2426766122. Schilke and Reimann, "The transparency dilemma: How AI disclosure erodes trust", Organizational Behavior and Human Decision Processes 188, 2025: https://doi.org/10.1016/j.obhdp.2025.104405. Careerminds UK survey of 600 UK employees, as reported by Workplace Insight, 19 August 2026: https://workplaceinsight.net/third-of-people-think-ai-would-do-a-better-job-than-their-human-manager/ and Onrec, 20 August 2026: https://www.onrec.com/news/statistics/more-than-1-in-3-uk-employees-say-an-ai-manager-would-boost-productivity. Information Commissioner's Office, Recruitment rewired, report on automated decision-making in recruitment, 31 March 2026: https://ico.org.uk/about-the-ico/what-we-do/recruitment-rewired/. House of Commons Business and Trade Committee, Artificial Intelligence, business and the future of the workforce: https://committees.parliament.uk/work/9674/. Gartner press release, 4 March 2026: https://www.gartner.com/en/newsroom/press-releases/2026-3-4-gartner-hr-survey-reveals-45-percent-of-managers-report-ai-has-lived-up-to-their-expectations.