Is it safe to use AI with customer data? UK GDPR myths checked

You can use AI with customer data under UK GDPR, but whether it’s safe depends on which tool you use, what you put into it and what you let it decide. Your business stays responsible for that data even when a large AI provider processes it. Five common assumptions get this wrong, and two of them changed during 2026.

“The AI company is responsible, not me”

This is the assumption that causes the most trouble, because it decides whether anyone in your business thinks about the risk at all. When you choose to send customer emails or call notes to an AI tool, you decide why and how that data is used. In data protection terms that usually makes you the controller and the provider your processor. The ICO’s guidance on accountability in AI spends a whole section on working out these controller and processor relationships, because the obligations that follow are different for each side.

In practice, being the controller means three jobs land on you: having a lawful reason to use the data this way, telling customers about it in your privacy notice, and having a contract with the provider that covers how they handle it. A big provider’s terms don’t do the first two for you, and for the third you need to check that the plan you pay for comes with data processing terms.

Whether your data trains the model depends on the plan

Many owners assume everything typed into ChatGPT goes into training. That’s half right. OpenAI’s business data page says it does not use data from ChatGPT Business, Enterprise or its API, including inputs and outputs, to train its models by default. Its help article on model training says content from individual accounts may be used for training unless the user turns off “Improve the model for everyone” in the data controls.

The gap between those two pages is where customer data leaks. A member of staff pastes a complaint, with the customer’s name and order history, into a personal account on their phone, and that conversation is now covered by consumer terms you never agreed to. Start with a business-tier account for anyone who handles customer data, and a one-page rule that personal AI accounts are for public information only. That costs less than any policy document and closes the easiest route.

UK GDPR no longer bans most automated decisions

Until this year, the old Article 22 of UK GDPR restricted decisions made solely by a machine that had a legal or similarly significant effect on someone. The Data (Use and Access) Act 2025 replaced it. As Travers Smith’s summary of the reforms sets out, Articles 22A to 22D took effect on 5 February 2026 and allow solely automated decisions on most personal data, as long as safeguards are in place. Decisions that use special category data, such as health information, remain restricted.

The safeguards are the part people skip. Customers have to be told about the decision, be able to challenge it and be able to get a person to look at it again. If you let an AI system decide who gets credit terms or a refund, build the route for a human review before you switch it on, not after the first complaint.

Small businesses still need a DPIA for most AI uses

A data protection impact assessment (DPIA) is a written check of what could go wrong with a piece of processing and how you’ll reduce the risk. It sounds like paperwork for banks. The ICO’s AI guidance takes the opposite view: it says that in the vast majority of cases, using AI involves processing that is likely to result in a high risk to individuals, which is the legal trigger for a DPIA.

For a small business the assessment doesn’t need to be long. It should answer these questions in plain language:

1. Which customer data goes into the tool, and does it need to? Removing names and account numbers before the AI sees a document cuts most of the risk.
2. Where is it stored, and for how long?
3. What does the AI decide or produce, and which person in your business checks that output before it reaches a customer or changes their account?
4. Which less risky option did you consider, and why didn’t you choose it?

The ICO’s guidance expects that last question to be answered, and it’s the question most likely to change your design. A tool that reads invoices to pull out totals and supplier names may need no personal data at all if you strip the contact details first.

AI agents are already on the regulator’s desk

An AI agent is a system that takes actions on its own, such as replying to an email or placing an order, rather than producing text for a person to use. It’s tempting to treat agents as next year’s problem. The ICO doesn’t. In October 2026 it announced that ten foundation model developers had made or committed to changes in how they handle personal data, and said its scrutiny now extends to AI agents. Its call for evidence on agentic AI is open to businesses that deploy these tools, not only developers, and A&O Shearman’s note on the call says the evidence will feed the ICO’s statutory code of practice on AI and automated decision-making.

The ICO’s position, as reported in that coverage, is that an agent acting on its own doesn’t excuse poor compliance. If an agent emails the wrong customer’s details to someone else, that’s your breach. Our advice for now: let agents draft replies and look things up, but keep a person approving anything that goes out to a customer or changes their account. You lose some of the time saving and keep control of the part that would hurt.

Assumption	What the source says	What to do
The AI provider is responsible	Controller and processor roles carry different obligations (ICO)	Treat your business as the controller; get the provider’s data processing terms
Everything you type trains the model	Business plans and the API are not used for training by default; individual accounts may be (OpenAI)	Business accounts for staff who handle customer data
Automated decisions are banned	Allowed for most data with safeguards since 5 February 2026 (Data (Use and Access) Act 2025)	Build a human review route before going live
DPIAs are for big firms	Most AI processing is likely to be high risk (ICO)	Write a short DPIA before you start
Agents can wait	Regulator scrutiny of agents began in October 2026 (ICO)	Keep a person approving customer-facing actions

The practical upshot is to pick your first AI project carefully. If you’re still deciding where to start, the guide on which processes to automate first favours internal, low-risk work for the same reason, and the AI automation work we do starts with what data a tool needs before what it can do.

Frequently asked questions

Do I need to tell customers I use AI?

Your privacy notice has to explain how you use their personal data, so if customer data goes into an AI tool, the notice should say so in plain words. You don’t need a banner on every email, but a customer who asks should get a straight answer about what the tool sees.

Can I use AI on data stored outside the UK?

If your provider processes data outside the UK, the UK’s international transfer rules apply, so check where the provider processes data and what transfer mechanism its terms rely on. The ICO’s guidance on international transfers is the place to start.

Is an on-premise or open-source model safer?

It keeps the data on systems you control, which removes the provider question. It also makes you responsible for security and updates, which a small business may find harder to get right than choosing a well-run business plan.

If you’re weighing up an AI project that touches customer data, book a free call and we’ll tell you honestly whether it’s worth doing and what it would cost.