Nearly a third of UK AI users hide it from their employer. Put yours on the record first

Deloitte surveyed 25,000 UK workers and found that 63% have used AI tools at work, Bloomberg reported on 16 September in a story syndicated by Claims Journal. TechRadar, covering the same study, reported that nearly a third of generative AI users work with it without their employer's knowledge. Almost one in ten have used a tool their employer has banned or would disapprove of. According to Deloitte's figures as reported by TechRadar, British workers spend an estimated £958 million a year of their own money on generative AI tools for work. And 63% worry their manager may think AI can do their job. That last number explains the rest. People hide the tool because they fear how visible productivity will be used against the case for their role.

The secret protects the wrong thing

The consensus reading treats shadow AI as an IT security problem. The National Cyber Security Centre warned about it the week before the survey landed, and David Chismon of the NCSC said security teams "should not assume they are seeing the full picture." The data backs him. Cyberhaven, which tracks data movement across hundreds of companies, found that a third of employees reach AI through personal accounts and almost 40% of AI interactions involve sensitive data. PagerDuty found in June that 88% of workers had shared work information with public AI systems, including 34% who had entered customer information.

For the individual, the cost is more specific. Gloucestershire Police dismissed a member of staff for using AI against the force's rules and getting round blocked sites with personal accounts, a case first reported by the Financial Times. The Independent Office for Police Conduct is investigating an officer for alleged gross misconduct over AI use. An employment tribunal is hearing a case in which an employer says it dismissed a worker for relying on AI-generated writing; the worker argues the real reason was concerns they raised about the company's handling of data.

The hidden route keeps the downside and gives away the upside. When the output is good, the credit goes to someone who merely appears fast. Nobody sees the judgement that chose what to hand to the model, what to check and what to throw out. When something leaks, the personal account and the unapproved tool become the evidence.

TechRadar quoted a Reddit user who had done "about 2 weeks' work in an hour" and wanted to know whether to tell the boss. Staying quiet makes the job look smaller. The contribution stays invisible.

The opening is the employer's gap

Most of Deloitte's respondents said their leaders have a poor or non-existent understanding of AI. Half of AI users received no formal guidance or training. PagerDuty found that 72% of workers believe they understand AI better than their own tech teams, and 77% think their company's AI restrictions are limiting their professional development.

Read that as a vacancy. Most organisations have nobody who can describe how AI is actually used inside their own workflows. James Longster of the law firm Travers Smith said unauthorised use nearly always runs on the consumer version of a tool, which will ingest the data and use it for training. The person who can say which tasks run on which tool, with what data, under which checks, holds the knowledge the policy will be written from. Chismon's prescription is open dialogue and clear guardrails. Someone on the floor has to supply the facts.

AI does the routine work. You do the thinking. Part of the thinking is deciding where the tool is safe to use — and that is work a manager can see and credit.

Your Next Move

1. Write your own use register this week. One page. List every task where you use AI, the tool, whether the account is personal or company-provided, and what data goes in. Mark anything involving customer information, financials or confidential documents. Those lines move to an approved tool or stop today.

2. Turn your subscription into a proposal. If you pay for a tool yourself, you are running a pilot nobody has measured. Take two tasks from the register, record the time before and after, and note what you check by hand and why. Bring that to your manager with three requests: an approved licence, a rule on what data may go in, and a named reviewer for the output. That is a request for guardrails, made from evidence.

3. Claim the hours before someone else allocates them. The 63% who fear their manager thinks AI can do their job are right to expect the question. Answer it first. At your next one-to-one, state in writing where the saved time now goes: the client conversation, the review, the decision the model cannot own. Hidden productivity reads as a smaller job. Declared productivity, with the judgement attached, reads as a larger one.