British workers are spending an estimated £958 million a year from their own pockets on AI for work, according to a new Deloitte UK survey of 25,000 workers. The same research found that 31% of people using generative AI at work do so without their employer’s knowledge, putting a striking price tag on a workplace shift that is happening partly outside official IT budgets.
AI for work has moved beyond the company software budget
Deloitte’s inaugural GenAI Workforce Survey was conducted online by Ipsos UK between 7 May and 10 June 2026 among a representative quota sample of employees and self-employed people aged 18 to 70. Sixty-three per cent said they knowingly use generative AI for work. Among those users, 46% reported using free tools, 34% used external tools paid for by an employer, 17% used in-house tools and 17% said they personally paid for at least one work-related GenAI tool.
The £958 million figure is Deloitte’s estimate of annual personal spending across UK workers, not a total taken from bank or card transactions. That distinction matters. The survey measures what people report doing, then extrapolates spending across the workforce. It still points to a substantial transfer of software costs from employers to individuals, but it should not be read as an audited national bill.
A personal subscription can look trivial in isolation. A worker may decide that a monthly chatbot subscription is worthwhile because it helps with writing, research or repetitive administration. Once that behaviour is repeated across a large workforce, it becomes a material economic signal. Employees are deciding that the tool is useful enough to buy even when their organisation has not supplied it.
Shadow AI is now part of the workplace AI story
Nearly a third of GenAI users in the survey, 31%, said they use these tools without their employer’s knowledge. Deloitte describes this as shadow AI. The term covers a familiar technology-management problem in a new form: people adopt a service because it solves an immediate problem before security, procurement and governance processes have caught up.
The risk is not simply that somebody has installed an unapproved app. A general-purpose AI service can receive draft contracts, customer information, internal strategy, source code or other material that an employee would never deliberately publish. Whether that creates a problem depends on the service, its settings, the data entered and the organisation’s rules. The important point is that employers cannot manage a use case they do not know exists.
Britain’s National Cyber Security Centre has already warned that more autonomous AI agents need practical controls rather than instructions alone. The Deloitte findings describe an earlier stage of the same governance problem. Before businesses worry about an agent taking actions across connected systems, many first need to establish which AI tools staff are already using and what information is being sent to them.
That sits alongside broader official evidence that workplace AI is already spreading beyond specialist teams. The UK government’s AI Adoption Research found that among businesses already using AI, an average 30% of staff currently use it. A separate UK Business Data Survey found that only 17% of AI-using businesses reported having formal or informal AI policies or guidelines. The studies use different samples and measures, but together they show why governance can lag behind day-to-day use.
Workers are using AI every day, but mostly for ordinary tasks
Almost a quarter of UK workers, 24%, told Deloitte they now use GenAI every day for work. The most common reported activities were searching for information and drafting emails, each cited by 43% of users, followed by summarising documents at 31%. Those are not exotic applications. They are ordinary pieces of knowledge work that can be repeated throughout a working week.
That helps explain why adoption can outrun formal company programmes. A worker does not need an enterprise transformation plan to decide that summarising a long document or producing a first draft is worth a subscription. The benefit is immediate and personal, while the questions about procurement, record keeping and information security sit elsewhere in the organisation.
The productivity picture still needs careful interpretation. Users reported saving an average of around 70 minutes a week, but self-reported time saved is not the same as measured productivity. A recent controlled trial covered by LiveAIWire found that experienced developers took 19% longer when AI coding tools were allowed, despite believing afterwards that the tools had made them faster. Different tasks can produce different outcomes, and perception is not a substitute for measurement.
The hidden cost is training as well as software
Deloitte’s wider survey material says around half of GenAI users have received no formal training. That is a problem because using a chatbot is easy while using one safely and critically is not. Workers need to know what information can be entered, when a generated answer needs independent checking, how to preserve records where required and which tasks should stay outside the tool altogether.
Training also matters for human capability. LiveAIWire recently reported experimental evidence that AI assistance can improve visible performance while hiding weaker unaided skills. A company that measures only how quickly work gets finished may therefore miss a second question: what can the employee still do when the assistant is unavailable, wrong or unsuitable for the task?
The answer is unlikely to be banning every personal AI subscription. Staff often adopt unofficial tools because official alternatives are missing, awkward or slow to arrive. A more useful response is to make the legitimate route easier than the shadow route, with approved tools, clear data rules, practical training and a way for workers to raise new use cases without waiting months.
The £958 million estimate reveals who is driving adoption
The most revealing part of Deloitte’s research is not that British workers like generative AI. It is that a meaningful share are prepared to pay for it themselves. That reverses the usual enterprise technology story in which management buys a system and then persuades employees to use it.
Here, some workers are behaving more like consumers. They choose a tool, learn it, absorb the monthly cost and only later create a governance question for the organisation. The pattern may be good news for AI vendors, but it leaves employers with a harder task than simply selecting software. They need to understand a market that is already operating inside their workforce.
For employees, the practical issue is simpler. A personally purchased tool can still be subject to workplace rules, confidentiality duties and professional obligations. Paying for the account does not make company information personal property. The survey suggests that this boundary is becoming important for ordinary workplace interactions at scale, not just specialist AI projects.
The £958 million estimate should therefore be treated as a signal rather than a precise national invoice. It shows that AI adoption has become valuable enough to many workers that they are reaching into their own pockets. The next stage for employers is to decide whether those private purchases remain invisible, or become part of a deliberate and measurable way of working.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.