The same basis the underlying purpose needs. GDPR Article 6 attaches a lawful basis to each purpose of processing, not to the tool doing it. An agent handling a customer's account query relies on the basis that covers servicing the account, usually contract. The harder question is whether everything the agent does stays inside that purpose.
Which bases are available, and which fit agents?
Article 6 offers six lawful bases: consent, contract, legal obligation, vital interests, public task and legitimate interests. For agents in regulated businesses, three do most of the work:
- Contract: processing necessary to perform a contract with the person or take steps at their request before entering one. Servicing an account, processing a claim under a policy or assessing an application the person submitted.
- Legal obligation: processing necessary to comply with the law. Customer due diligence, sanctions screening, regulatory reporting.
- Legitimate interests: processing necessary for the organisation's legitimate interests, unless overridden by the person's interests and rights. Fraud prevention and some internal analytics often rest here, after a documented balancing test.
Consent fits optional features but is fragile as a basis for core work, because it must be freely given and can be withdrawn. When it does fit is examined in GDPR consent as a lawful basis for AI agents. Special category data, such as health information, needs a further condition on top of an Article 6 basis.
The article-level detail is in GDPR Article 6 lawful basis for AI agents processing customer data.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Why does an agent make the basis harder to keep?
The basis is chosen for a purpose. An agent can drift from that purpose in the course of a task.
Necessity is tested per action. Contract and legal obligation both require the processing to be necessary. An agent that reads a customer's full transaction history to answer a question about one fee is processing more than necessary for that purpose.
Purposes blend. An agent servicing an account may notice the customer could qualify for a better product and suggest it. That is marketing, a different purpose with a different basis, and possibly a need for consent under e-privacy rules.
Decisions shift the analysis. Where an agent makes solely automated decisions with significant effects, Article 22 adds its own conditions, which narrow the usable bases to contract, law or explicit consent.
What may the agent do, and what stays with a person?
| Agent activity | Likely basis | Stays with a person |
|---|---|---|
| Answer the customer's question about their account | Contract | Disputes over the account |
| Screen a new customer against sanctions lists | Legal obligation | Confirming a true match |
| Flag possible fraud on a transaction | Legitimate interests, documented | Blocking the account |
| Suggest a new product | Separate purpose; check basis and consent | Approving the campaign rules |
| Use health information in a claim | Contract plus a special category condition | Decisions on the claim |
An illustrative example
Take a bank whose agent processes personal loan applications. Assessment of the application rests on steps taken at the applicant's request before a contract, and affordability checks partly on legal obligation under lending rules. The bank's privacy team reviews what the agent actually reads.
It finds the agent pulls the applicant's full transaction history across all accounts, including joint accounts with another person, and uses it to infer spending patterns. The joint account holder never applied for anything. The bank restricts the agent to the applicant's own accounts and to the statement period its lending policy requires, and records why each source is necessary. The full process is set out in how to process personal loan applications with an AI agent.
Which controls follow from the basis?
- A purpose and basis per capability, recorded in the records of processing.
- Source permissions tied to necessity, so the agent reaches only what the purpose needs.
- Refusals for out-of-purpose work, such as marketing from a servicing agent.
- Legitimate interests assessments where that basis is used.
- A data protection impact assessment where the agent's processing is high risk; see whether AI agents need a data protection impact assessment.
- Processor terms and transfer checks for the model provider, including where it is abroad; see whether personal data can be sent to a model provider in another country.
The full map is in AI agent regulation, mapped to controls.
Questions people ask
Do we need a separate lawful basis for using an AI agent?
No. The basis attaches to the purpose. Using an agent instead of a person does not need its own basis, but it may change whether the processing is necessary and proportionate.
Can a model provider rely on our lawful basis?
When it acts as your processor on your instructions, yes. If it uses the data for its own purposes, it needs its own basis, which your contract should rule out.
Does UK GDPR offer the same bases?
UK GDPR, with the Data Protection Act 2018, has the same six bases. Check UK-specific guidance on legitimate interests and automated decisions.