GDPR purpose limitation when an AI agent reuses data

Purpose limitation means personal data collected for one purpose cannot be reused for an incompatible one. An AI agent breaches it when a new capability draws on data gathered for something else, such as using complaint transcripts to target sales offers, without checking compatibility first. The check is a short, recorded test you run before the capability goes live.

What does purpose limitation require?

Article 5 requires personal data to be collected for specified, explicit and legitimate purposes, and not further processed in a way incompatible with them. The purpose has to be fixed at collection and stated, not reconstructed later.

Further processing is allowed if it is compatible. Article 6 sets out the factors for judging that, in substance:

• the link between the original purpose and the new one;
• the context of collection, especially what the person would reasonably expect;
• the nature of the data, with more caution for special category data;
• the possible consequences of the new use for the person;
• the safeguards in place, such as pseudonymisation.

If the new use is not compatible, you need a fresh lawful basis, often consent, or you cannot proceed. Archiving in the public interest, research and statistics are treated as compatible when appropriate safeguards are in place.

Who does it apply to?

Every controller under GDPR and UK GDPR. It is a controller's duty, so it sits with the firm deploying the agent, not the platform. A processor that reuses your data for its own purposes is not making a compatibility judgement; it is breaching its instructions.

Why do agents put purpose limitation under strain?

Agents are built to find what is useful. Give an agent broad read access and it will use whatever helps it do the task in front of it. That is the property that makes agents productive and the property that breaks purpose limitation.

Three patterns cause most of the trouble.

Shared sources across capabilities. A firm connects one document store to several agents. The complaints agent reads complaint transcripts to resolve complaints, a compatible purpose. The sales agent is pointed at the same store to "understand customer needs". Same data, new purpose, no check.

Memory carried across tasks. An agent that remembers what it learned in one conversation can apply it in another. If an agent learns during a hardship call that a customer has lost their job, and later uses that in a cross-sell decision, the data has moved purposes inside the agent.

Repurposing by prompt. Someone changes an agent's brief from "answer billing questions" to "answer billing questions and identify upgrade opportunities". One sentence changes the purpose of processing for every customer the agent touches.

There is also the question of the model. Using general-purpose models does not change your purpose analysis; the model is a means, not a purpose. Our explainer on what a general-purpose AI model is covers the distinction, and the separate duties of model providers are covered in whether general-purpose AI model rules apply to agent builders.

How do you meet the requirement?

The practical answer is to tie data access to purposes, and purposes to capabilities, so a change in one forces a review of the others.

The memory row overlaps with retention. How long scoped memory should last is covered in GDPR storage limitation and AI agent memory.

What does a compatibility check look like in practice?

Take an illustrative example. A lender's collections agent reads payment history to agree repayment plans. The business wants a second capability: the same agent proposing a top-up loan to customers who complete a plan.

Link between purposes: weak. One supports someone in difficulty; the other sells credit. Reasonable expectations: low; a customer in arrears does not expect their hardship file to drive a sales pitch. Nature of data: may include health or employment detail. Consequences: possible harm from more debt. Safeguards: none would cure the mismatch.

Result: incompatible. The top-up capability needs its own agent, its own data sources and its own lawful basis, if it should exist at all. Writing this down takes an hour. Discovering it after a complaint takes months.

What will a supervisor or auditor ask?

• For each agent capability, what is its purpose, and which data sources can it read?
• When did a capability last gain access to a new source, and where is the compatibility assessment?
• Can an agent use something it learned in one task in an unrelated task?
• Who approves changes to an agent's brief, and is privacy part of that approval?
• How did you tell customers about new uses of their data?

What does purpose limitation not cover?

It does not set retention periods; storage limitation does. It does not decide the lawful basis for the original collection. It does not cover the logging an agent does to evidence its own behaviour, which has its own justification; that tension is set out in EU AI Act record-keeping vs GDPR storage limitation. And it says nothing about the quality of the agent's output. For the full set of rules, see AI agent regulation, mapped to controls.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.