Home / Insights

Can you train AI on your own customer data? What UK law allows

The data is yours, the customers are yours, and the model would run inside your own walls, so training on it feels like a private matter. Under UK data protection law it is a new purpose for personal data that was collected for something else, and that is the test the whole question turns on.

The short answer

  • Training a model on personal data is processing, so it needs a lawful basis of its own. Holding the data lawfully for its original purpose does not carry over.
  • The first hurdle is purpose limitation: is training compatible with the reason the data was collected? If not, you need consent or a fresh legitimate-interests assessment.
  • Legitimate interests is the usual basis, and it requires a documented three-part test. The "recognised legitimate interests" added by the Data (Use and Access) Act 2025 do not cover model training, so the balancing test still applies.
  • Transparency and individual rights do not switch off. People must be able to understand and object to the use, which is why first-party data you can explain beats scraped or bought data you cannot.
  • A data protection impact assessment is effectively mandatory for training on personal data at any scale.

Why "we already hold it" is not the answer

Every use of personal data needs a lawful basis under Article 6 of the UK GDPR. A basis attaches to a purpose, not to the data itself. You collected customer records to deliver a service, take payment and meet a legal duty, and those are the purposes your original basis covers. Training a model is a different purpose, so it needs its own basis and its own justification.

Before you even reach the choice of basis, purpose limitation under Article 5(1)(b) asks whether the new purpose is compatible with the original one. Compatibility turns on the link between the two purposes, the context in which the data was collected, the nature of the data, the possible consequences for the individual, and the safeguards in place. Training an internal fraud model on transaction data a customer gave you to process payments is often defensible on that test. Training a general-purpose product feature on the content of private support messages is much harder, because the customer would not reasonably expect it. This is the same distinction that decides whether the data you hold can be put to a second use at all.

Choosing a lawful basis for the training purpose

In practice three bases are relevant, and most enterprise training runs on the third.

BasisWhen it fits trainingWhat it costs you
Consent Where you can ask cleanly and the person can genuinely refuse without losing the service. It must be specific, freely given and as easy to withdraw as to give. Withdrawal has to be handled, which is hard once data is inside a trained model.
Contract Rarely. Only where the training is genuinely necessary to deliver the service the person signed up for. Necessity is a high bar. "It improves the product" is not the same as "the contract cannot be performed without it".
Legitimate interests The common route for internal models built on first-party data. A documented three-part test: a real purpose, necessity, and a balancing exercise against the individual's rights and expectations.

The Data (Use and Access) Act 2025, which received Royal Assent in June 2025, added a set of "recognised legitimate interests" that can be relied on without the balancing test. AI model training is not among them; the list covers specific public-interest purposes such as safeguarding and crime prevention. For training, the full legitimate-interests assessment still applies, and it has to be written down before the processing starts rather than reconstructed afterwards.

What the ICO has said about scraped data, and why it matters even if you do not scrape

The ICO ran a five-part consultation on generative AI through 2024 and published its outcomes at the end of that year. Its position on web-scraped training data is instructive: legitimate interests is realistically the only available basis, and it is genuinely difficult to satisfy, because people whose data is scraped without their knowledge cannot exercise their rights, which weakens the balancing test and strains the transparency duty.

That reasoning is the argument for first-party data, put in legal terms. When you train on data your own customers gave you, you can tell them, you can honour an objection, and you can pass the balancing test, precisely because the relationship and the transparency exist. The weakness the ICO identifies in scraped data is a weakness your first-party data does not have, provided you actually use the transparency you are capable of.

The duties that do not go away

A lawful basis is necessary but not sufficient. Four further duties apply to any training on personal data.

Transparency

Articles 13 and 14 require you to tell people how their data is used. A privacy notice that lists your ordinary operations but is silent on model training does not cover it. Naming the use plainly is also the cheapest way to strengthen a legitimate-interests balance.

Special category data

Health, biometric, ethnicity and similar data need an additional condition under Article 9. If your training set contains any, or a model could infer it, the bar rises and consent or a substantial-public-interest condition is usually required.

Automated decisions

If the trained model then makes or heavily informs a decision about a person, the rules on automated decision-making apply. The Data (Use and Access) Act adjusted these, permitting more automated decisions where safeguards, including a route to human review, are in place. The safeguards are the condition on which those decisions are allowed.

A data protection impact assessment

Training on personal data at scale is high-risk processing, so an Article 35 impact assessment is effectively required. It is also the document that records the purpose-limitation reasoning and the legitimate-interests test, so it is the artefact a regulator would ask for first.

The constraint

The binding limit here is a documented lawful basis for the training purpose itself, tied to why the data was originally collected, rather than the size or quality of the dataset. Acquiring more data, or better tooling to process it, without settling purpose limitation and the legitimate-interests test increases the exposure rather than the readiness, because the deciding question is a legal and governance one about the purpose, and no amount of data engineering answers it.

The question to put before any training run

Ask, for the specific dataset and the specific model: what is our lawful basis for training as a purpose, is training compatible with why we collected this data, can we tell the people affected and honour an objection, and is the impact assessment written down. A training run that can answer those four is on solid ground under UK law. One that cannot is a data protection incident waiting to be discovered, regardless of how good the model turns out to be.

Sources

  1. Information Commissioner's Office, response to the consultation series on generative AI, December 2024.
  2. ICO, guidance on AI and data protection: how do we ensure lawfulness in AI.
  3. Data (Use and Access) Act 2025, Royal Assent 19 June 2025, including recognised legitimate interests and the automated decision-making provisions.
  4. UK GDPR, Articles 5(1)(b), 6, 9, 13, 14 and 35, as applied through the ICO guidance above.