No AI agent has a lawful basis of its own; each purpose it serves needs one. Under GDPR Article 6, pick a basis per purpose: contract for servicing an account, legal obligation for regulatory checks, legitimate interests for fraud prevention or internal efficiency, and consent only where it is freely given. Record the choice before launch.
What does Article 6 require?
Processing of personal data is lawful only if at least one of six bases applies. In substance:
- Consent. The person has consented to processing for one or more specific purposes.
- Contract. Processing is necessary for a contract with the person, or for steps they asked for before entering one.
- Legal obligation. Processing is necessary to comply with a legal obligation the controller is subject to.
- Vital interests. Processing is necessary to protect someone's life.
- Public task. Processing is necessary for a task in the public interest or in the exercise of official authority.
- Legitimate interests. Processing is necessary for the legitimate interests of the controller or a third party, unless overridden by the person's interests or rights. Public authorities cannot use this basis for their tasks.
Necessary is the important word in four of them. It means the processing is a targeted and proportionate way to achieve the purpose, not just helpful.
Who must decide the basis?
The controller, before processing starts. The basis must be recorded, given in privacy information, and is hard to change later: you cannot switch to a different basis because the first one fails. If the controller uses a vendor's agent, the vendor does not choose the basis; you do.
Why does an agent need a basis per purpose?
An agent is a tool that can serve many purposes, sometimes in the same conversation. A servicing agent might answer a balance query (contract), run a sanctions check before a payment (legal obligation), flag possible fraud (legitimate interests or legal obligation), and suggest a product (legitimate interests or consent, depending on the marketing rules). Each is a different purpose with a different basis.
Three agent-specific traps:
Basis drift. An agent set up under contract to service accounts starts summarising conversations for product development. That is a new purpose, and contract does not cover it.
Necessity stretched by convenience. An agent reading a customer's full transaction history to answer a question about one payment may be doing more than the contract basis supports. The basis covers what is necessary for the purpose, not everything the agent finds useful.
Inference as new processing. When an agent infers something new about a person, such as financial difficulty from spending patterns, it creates new personal data. That inference needs a basis too, and may be special category data, which needs more. An inference about health drawn from pharmacy payments is the classic example, and few teams notice they have created it.
How do you pick and document a basis for each purpose?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Identify purposes | Each agent capability mapped to a named purpose | Capability-to-purpose map |
| Choose a basis per purpose | Basis selected with reasoning; necessity tested | Lawful basis assessment per purpose |
| Contract basis | Processing limited to what the contract needs | Data map tied to contract terms |
| Legal obligation basis | Law identified for each check the agent runs | Obligation register entry |
| Legitimate interests basis | Purpose, necessity and balancing test documented | Legitimate interests assessment |
| Consent basis | Specific, informed, freely given, withdrawable consent | Consent records; withdrawal handling |
| Inferences | New data the agent creates assessed for basis and category | Inference assessment |
| Tell people | Privacy information names purposes, bases and the agent's use | Privacy notice versions |
| Stop drift | New purpose for an existing agent triggers basis assessment | Change records |
Record the result alongside the purpose in your GDPR Article 30 records of processing, so the two stay aligned.
Consent deserves caution for agents; it is often the wrong basis for core servicing, as GDPR consent as a lawful basis for AI agents explains. The short version of the whole question is in what lawful basis an AI agent needs to process personal data.
How does this compare outside the EU?
The US has no single lawful basis regime. Financial privacy for consumer data is shaped by GLBA and by state laws such as California's CCPA as amended by the CPRA, which work through notice, opt-out and use restrictions rather than six bases. A firm running agents in both should map each purpose to both regimes; our note on whether GLBA applies to AI agents handling customer data covers the US side.
What will a supervisor or auditor ask?
- What purposes does this agent serve, and what is the basis for each?
- Why is this processing necessary for the contract?
- Show us the legitimate interests assessment.
- What new data does the agent infer about people, and on what basis?
- How did you tell customers?
What does Article 6 not settle?
It is not enough on its own for special category data, which needs an additional condition. It does not decide whether a solely automated decision is allowed; Article 22 does. It does not make processing fair or minimal; other principles apply. And it does not fix the lawful basis for the model provider's own uses, which should be excluded by contract. For the full map, 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
Can one lawful basis cover everything an agent does?
Rarely. An agent serving several purposes needs a basis for each. Using one broad basis for all usually means some processing is not actually covered.
Is legitimate interests a safe default for agents?
It is flexible but not a default. It needs a documented balancing test, gives people a right to object, and is unavailable to public authorities for their tasks.
Do we need a new basis when we switch to an agent from a manual process?
Not necessarily. If the purpose is the same, the basis may be the same. Check whether the agent processes more data, or makes new inferences, than the manual process did.