Using ChatGPT, Claude and Copilot at work lawfully, in a UK business.

Whatever your policy says, your staff are already using these tools. Ofcom’s 2026 adults’ media survey, 7,533 UK adults interviewed between 29 September and 28 November 2025, found that 54 per cent now say they use AI tools such as ChatGPT, Copilot or Gemini, up from 31 per cent a year earlier; among 25 to 34 year olds it is 74 per cent and among 16 to 24 year olds 79 per cent. Among those who use AI, 47 per cent say they do so to support work or study.

Businesses have not caught up with their employees. DSIT’s UK Business Data Survey 2026, 4,450 businesses interviewed between October 2025 and January 2026, found that 41 per cent of businesses handling digitised data use AI, but only 17 per cent of those had any policy on its use, formal or informal. Just 5 per cent had a formal, written policy; 12 per cent had informal guidance. Among micro businesses 8 per cent had a formal policy, among sole traders 3 per cent. Only 53 per cent had even heard of AI regulatory guidance, and of those, 19 per cent found it clear.

MIT’s 2025 study of large organisations describes the same gap from the other side: while only 40 per cent of companies said they had purchased an official LLM subscription, workers from over 90 per cent reported regular use of personal AI tools for work. The report calls it a shadow AI economy. The practical consequence for a UK firm is that the question is not whether to allow these tools but which accounts, which data and which rules.

There is no separate law for generative AI in the UK. UK GDPR and the Data Protection Act 2018 apply whenever personal data goes into a tool, and the ICO’s guidance on AI and data protection, last updated on 15 March 2023, sets out how the principles apply: accountability and governance, transparency, lawfulness, accuracy, fairness, security and data minimisation, and individual rights. The ICO also publishes an AI and data protection risk toolkit and separate guidance on explaining decisions made with AI.

The sentence a small business should read first is in the accountability chapter: in the vast majority of cases, the use of AI will involve a type of processing likely to result in a high risk to individuals’ rights and freedoms, and will therefore trigger the legal requirement for a DPIA. Where you assess a particular use as not high risk, you still need to document how you made that assessment. A DPIA for pasting a client email into a chatbot is a page; a DPIA for a tool that screens applicants is not.

The Data (Use and Access) Act 2025 changed the rules on automated decisions. Per the ICO, all of the Act’s data protection provisions are now in force, and it opens up the full range of lawful bases, including legitimate interests, for significant automated decisions, so long as safeguards remain and special category data is not involved. If a tool in your business makes or materially shapes a decision about a person, the safeguards, telling them, letting them contest it, getting a human to look, are the design requirement.

The same name, ChatGPT, Claude or Gemini, covers two products with opposite defaults on training. The consumer version may use what you type to improve the model unless you opt out; the business or API version does not by default. The two look identical on screen, which is why a policy that names the account rather than the tool is the only kind that works. Read on the providers’ pages in September 2026, the position is this.

Vendor statements on the vendors’ own pages, read 14 September 2026. The rule that follows is simple: personal or confidential data goes only into an account whose terms are in the "No" rows.

In UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 00081 (IAC), promulgated on 17 November 2025, a three-judge panel of the Upper Tribunal dealt with legal representatives who had used generative AI in case preparation. At paragraph 60 the Tribunal said: "Uploading confidential documents into an open-source AI tool, such as ChatGPT, is to place this information on the internet in the public domain, and thus to breach client confidentiality and waive legal privilege", adding that such conduct might warrant referral to the regulator and should in any event be referred to the Information Commissioner’s Office.

The Tribunal drew a distinction that matters for every business, not only law firms: closed-source AI tools which do not place information in the public domain, such as Microsoft Copilot, are available for tasks such as summarising without these risks. The line it drew is between a public tool on personal terms and an enterprise tool inside a controlled environment, which is the same line the training table above draws.

Two cautions. The Tribunal was describing an immigration case and its observation is guidance rather than a full judgment on privilege, and it did not address a public tool used with its privacy or training settings changed. Until an appellate court does, the safe assumption for any business that holds legal advice, contracts under negotiation or client confidences is the one the Tribunal made: a consumer AI account is the public domain.

A DPIA is a written assessment of the risks of a processing activity and how you will reduce them. The ICO’s AI guidance says most AI use of personal data will need one. For generative AI tools in a small business the triggers below are the ones that arise in practice; if none applies, write down why, because the ICO expects the assessment to be documented either way.

A written policy is what 95 per cent of AI-using UK businesses do not have. It does not need to be long. Each line below is a rule, and a policy that fits on one page will be read; one that runs to twelve will not. Adapt the account names to the tools you actually buy, and keep the rules about data and accounts even if you change every tool, because they are the part the law cares about.

A policy nobody understands is a document, not a control. The briefing that makes it work is short and concrete, and it answers the questions people actually have rather than the ones the regulator asks. In our experience it takes about an hour, and the most useful part is the demonstration, because most people have never seen the difference between a work account and a personal one on the same screen.

A twenty-person professional services firm currently has staff using personal ChatGPT and Claude accounts, some paid for on expenses. Client emails, draft contracts and the occasional CV are going into them, which after the Upper Tribunal’s observation is a risk the firm’s insurer would want to know about. The change is a set of accounts, a page of policy and a morning of training, and it can be priced from published figures.

Accounts: a business tier at each provider staff already use. Anthropic lists a Claude Team standard seat at $25 a month billed monthly or $20 on an annual plan, so twenty seats are $500 a month, or $4,800 a year on the annual plan, in dollars plus VAT; OpenAI’s business tier and Microsoft’s Copilot licences are priced on their own pages and should be read there on the day. Against this, the personal Pro subscriptions being expensed at $20 a month each are already $400 a month for twenty people, on terms that may train on the content. The business tier is not much more expensive; it is a different contract.

Policy and training: the one-page policy above takes an afternoon to adapt. A half-day training session with us is £1,200 fixed, and it ends with the policy, the approved-tool list and a working prompt library the team keeps. A DPIA for the common uses is a short document if drafted alongside. Total for the year, at the assumptions above: about $4,800 to $6,000 in seats and £1,200 in training, for a business that is now inside the "No" rows of the training table rather than outside them.

Seat prices from claude.com/pricing on 14 September 2026, in US dollars excluding VAT. The consumer row is the cost most firms are already paying without the contract.

The £1,200 half day is built for a whole team at once, and it is practical rather than theoretical: the accounts to use and why, live demonstrations of the tools on your own documents with the personal data removed, the checking habits that catch confident errors, and the policy, finished in the room and left with you. Leadership sessions add the DPIA and the automated decision-making safeguards; developer sessions add the API terms, regions and cost controls. Every session ends with an artefact the team owns, because a workshop with nothing left behind decays in a fortnight.

A full day is £2,000 and a multi-day programme up to £6,000, fixed, delivered within three weeks of booking. Sessions are aimed at one audience at a time, because a mixed room produces material too shallow for the builders and too technical for the decision makers. If the honest recommendation is that a policy and a business subscription are all you need this year, that is what we will say.

Every figure and legal statement above was read on its source page on 14 September 2026. Adoption figures are Ofcom’s Adults’ Media Use and Attitudes Report 2026 and DSIT’s UK Business Data Survey 2026, with fieldwork dates given in the text; the shadow AI figures are from MIT NANDA’s 2025 report of large organisations. The ICO material is its guidance on AI and data protection, the accountability chapter on DPIAs, the AI guidance hub, and its Data (Use and Access) Act page as updated on 19 June 2026. The Upper Tribunal decision was read on the National Archives case law service; the quotation is paragraph 60 verbatim.

Provider positions on training and retention are from the providers’ own legal and help pages, with effective dates where stated; they change without notice. The ICO’s AI guidance hub, as read on 14 September 2026, lists no separate generative AI guidance, so this guide relies on the general AI guidance and the DPIA chapter. Prices are in the currency each vendor bills in, excluding VAT. Nothing here is legal advice.