What new UK research reveals about workers, AI and the policies in between

In February 2026, Red Eagle Tech surveyed 200 UK desk-based and hybrid workers about artificial intelligence (AI) at work. The headline finding was about permission. Specifically, who is allowed to use AI for what, and how clearly the company has communicated that to the people doing the work.

54.5 per cent of those workers said their employer has no clear, enabling AI policy. 63 per cent use AI at work at least weekly anyway. About a third reach for products their employer has not approved. The result, across a great many UK businesses, is that decisions about AI use are being made by individual employees on a daily basis, well before the company has decided what it considers acceptable.

For anyone whose job involves choosing technology for a business, this matters in a specific way. Vendor selection is one of the few moments when permission, capability and procurement collide. If the procurement side moves slower than the people it is meant to support, the chosen product enters a workplace that has already routed around it.

A permission problem, not a skills problem

The dominant narrative in 2025 was that UK workers needed AI training. Our research points to something different. 74.5 per cent of UK desk workers already use AI in their personal lives every week. They have already worked out how to draft an email faster, summarise a long document, tidy up a spreadsheet, prepare a first version of a contract clause. The capability is already inside the company. The blocker is whether the company has yet decided to count it as work.

That distinction has real cost. Of the workers in our sample who said they were operating without any guidance at all, around 30 per cent had decided to use AI anyway. The other 70 per cent had decided to wait. The waiting group reported lost productivity, frustration, and a sense that colleagues at competitor firms were getting ahead of them. 37.6 per cent said they were spending time on repetitive work that AI could speed up, and they knew it. 18.3 per cent said they worried about falling behind peers who were less cautious about the rules.

We have been calling this the Conscientious Worker Penalty. Vague rules do not stop the workers who were going to use AI anyway. The price is paid by the careful workers who would have asked first if there had been somewhere to ask. Those are usually the workers a company least wants to lose.

Why bans do not solve it either

13.5 per cent of the sample said their workplace had banned AI outright. You might expect that to make shadow use disappear. It does not. The rate at which people use unapproved AI stays at around 31 to 35 per cent across every policy category, whether the policy is enabling, vague or prohibitive. Microsoft’s UK research found a similar pattern, with 71 per cent of workers admitting to using unapproved AI and just over half doing so at least weekly.

A ban does not count as governance. It moves the behaviour out of view, with all of the risk of the original behaviour and far less of the visibility that would let leadership see what is happening. The company still has employees pasting work content into a chat window. It just no longer knows about it.

The strictness of the policy turned out not to be the deciding variable. Workers operating under an enabling policy reported a much stronger positive view of their employer, and were three and a half times more likely to cite AI policy as a major factor in whether they stayed in their role. 25.5 per cent of the total sample now treat AI policy as a major factor in job decisions. AI has joined pay, flexibility and progression as part of the basket of things a worker weighs when deciding whether to stay.

The buying decision has a hidden second layer

Most product evaluations look at features, integrations, security posture and price. None of those, on their own, answer the question an employee asks on Monday morning: am I allowed to use this for the actual work I do.

Microsoft’s UK study found that 28 per cent of workers said their employer provides no work-approved AI option at all. Those employees do not give up. They use the consumer version on their phone, pasting in fragments of work content while no one is looking. In those workplaces, procurement looked like it was working in the usual sense, with vendors compared and a contract signed. What had not happened was the earlier step of writing down what employees were allowed to use any of this for, and a product purchase cannot fill that gap once it has opened.

For a buyer wanting to avoid that outcome, the questions worth asking before any contract is signed have less to do with the product itself and more to do with the surrounding context. Has the company published an acceptable use policy in plain English? Do employees know which tasks are in scope and which are not? Is there a named person they can go to with a borderline case? Has training been scheduled, or is the assumption that the product will speak for itself? When those questions have answers, the chosen product fits the workplace. When they do not, the product becomes another thing employees work around.

What good looks like

The recommendations in the research write-up are simpler than the framing might suggest. Write down which AI use cases the company is comfortable with, in plain English. Connect that document to the specific tasks people actually do, so it does not sit unread in a folder. And give employees a route for raising new ideas, so the policy can keep up with the work rather than freeze on the day it was published.

None of this needs a compliance department behind it. Regional breakdowns and the open-ended responses are explored further in the UK AI permission gap research write-up.

Where this leaves a buyer in 2026

If you are choosing AI for a business this year, treat the policy work as part of the project rather than something HR will pick up later. The procurement decision and the permission decision affect the same people in the same week. Sequencing them separately is what produces the disconnect this research describes, where workers run ahead on capability and behind on authorisation, and where the chosen product enters a workplace that has already worked out how to live without it.

The permission gap closes the moment somebody in leadership decides to spend half a day on it. Every quarter that the conversation slips down the agenda, the default version of the gap stays in place by inertia. The work itself is shorter than most leaders expect: a written document, a named person to handle edge cases, and a date in the diary to revisit the policy as the use cases grow. The first version usually takes less than a week of focused attention to get in place, after which the procurement decisions that follow stop running against a question nobody had answered.