An AI agent is profiling under GDPR whenever it uses personal data, by automated means, to evaluate something about a person: their reliability, financial situation, health, behaviour or likely next move. A risk score, a churn flag or a vulnerability marker all qualify. Profiling is lawful, but it brings transparency duties, a right to object and, sometimes, Article 22.
What does GDPR say about profiling?
GDPR defines profiling in substance as any automated processing of personal data that evaluates personal aspects of someone, in particular to analyse or predict their performance at work, economic situation, health, preferences, interests, reliability, behaviour, location or movements.
Three things follow. First, profiling needs a lawful basis under Article 6 like any other processing, and must respect the Article 5 principles of fairness, accuracy and data minimisation. Second, people have a right to object to profiling based on legitimate interests or a public task, and an absolute right to object to profiling for direct marketing. Third, where profiling feeds a decision based solely on automated processing with legal or similarly significant effects, Article 22 applies.
Profiling also counts heavily in the test for when a data protection impact assessment is needed under Article 35. Systematic and extensive evaluation of people that leads to decisions with significant effects is a standard trigger.
Who does it apply to?
Every controller whose agents evaluate individuals, whether customers, applicants, employees or claimants. It applies in the same way under UK GDPR, though the UK has amended its automated decision-making rules; our comparison of UK GDPR and EU GDPR for AI agents covers where the two now part.
It does not apply to evaluations of companies, unless the company is effectively a person, such as a sole trader.
When does an agent's work amount to profiling?
Agents profile more often than their owners notice, because evaluation is built into how they work. Ask what the agent concludes about the person, not what the agent is called.
| Agent task | Profiling? | Why |
|---|---|---|
| Answering a balance question | No | Retrieves a fact; evaluates nothing |
| Sorting complaints by topic | Usually no | Classifies the complaint, not the person |
| Flagging a customer as potentially vulnerable | Yes | Evaluates health, circumstances or behaviour |
| Scoring a loan applicant's affordability | Yes | Evaluates economic situation |
| Predicting which customers will leave | Yes | Predicts behaviour |
| Ranking collection cases by likelihood to pay | Yes | Evaluates reliability and economic situation |
| Summarising a claim file for a handler | Borderline | Profiling if the summary rates the claimant |
Two agent-specific risks deserve attention. The first is inference in passing. An agent asked to draft a reply may note, in its working memory, that the customer "seems financially stretched". That is an evaluation of economic situation, stored and reusable. Whether that note is personal data the agent keeps is a question our piece on whether GDPR applies to what an AI agent remembers answers: it is.
The second is chained use. A profile created for one purpose, such as vulnerability support, can be picked up by another agent doing collections or marketing. Each new use needs its own lawful basis and must be compatible with the original purpose.
How do you meet the requirements?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Know where profiling happens | Register of agent capabilities that evaluate people, with what they infer | Capability register marked for profiling |
| Lawful basis for each profile | Basis recorded per capability; balancing test where relying on legitimate interests | Records of processing; balancing assessments |
| Transparency | Privacy notice says which agents evaluate people, on what data, with what effect | Published notice; version history |
| Right to object | Objection stops the profiling capability for that person, not just marketing emails | Objection log; suppression flags the agent reads |
| Accuracy | Inferences labelled as inferences, never as facts; customers can correct the inputs | Field-level labelling; correction log |
| Minimisation | Agent can only see the data fields the evaluation needs | Source permissions per capability |
| No reuse without check | Profiles scoped to the capability that created them | Access rules; compatibility assessments |
| Article 22 where relevant | Decisions with significant effects held for a named person, or an exception documented | Hold records; exception assessments |
| DPIA | Assessment before profiling goes live | Signed DPIA |
If legitimate interests is your basis, the balancing test must weigh how intrusive the inference is. Our guide to GDPR legitimate interests and AI agents works through that test for agents.
What will a supervisor or auditor ask?
- Which of your agents evaluate people, and what do they conclude?
- Where are those conclusions stored, and which other agents or teams can read them?
- What lawful basis supports each profile, and where is the balancing test?
- How does a person object, and what stops when they do?
- Do any profiles feed decisions with significant effects, and does a person with real authority take part in those decisions?
- How do you check the inferences are accurate and not unfairly skewed against any group?
What does the profiling rule not cover?
It does not ban profiling. It does not require consent in most cases. It does not, on its own, require a human to review every assessment; that duty comes from Article 22 and only for solely automated decisions with significant effects. Our explainer on automated decision-making under GDPR sets out where that line falls. Profiling rules also say nothing about whether an agent's assessment is any good. That is a question of testing and certification, not data protection.
For the wider map of obligations, 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.
Questions people ask
Is a vulnerability flag set by an agent special category data?
It can be. If the flag reveals or implies health information, such as a mental health condition or a disability, it is health data and needs a condition for special category processing as well as a lawful basis. Label the flag carefully and restrict who can read it.
Does a customer have to be told each time an agent profiles them?
No. The duty is to tell people, in your privacy information, that profiling happens, what it involves and its consequences. Individual notice is needed when a profile drives a significant automated decision about them.
Can an agent keep a profile after the customer leaves?
Only as long as you need it for the purpose it was created for, and that purpose usually ends with the relationship. Set a retention period per profile type and make the agent's memory obey it.