Yes, if the conditions are met. You need a lawful basis for the processing, a contract that binds the provider as your processor, and a valid route for the international transfer, such as an adequacy decision or approved contractual safeguards with an assessment of the destination. With an agent, you must also control what it sends.
What do the rules require, in plain words?
Under GDPR, every processing of personal data needs a lawful basis under Article 6 and must respect the Article 5 principles, including data minimisation and purpose limitation. A model provider that processes data on your instructions is normally your processor, which requires a contract setting out what it may do with the data. Our guide to the lawful basis an AI agent needs to process personal data covers the first step.
Sending personal data outside the European Economic Area is restricted. In substance, a transfer is allowed where the European Commission has found that the destination offers adequate protection, or where appropriate safeguards are in place, such as standard contractual clauses approved by the Commission, backed by an assessment of whether the destination's laws undermine them. Narrow exceptions exist but are not suitable for routine agent traffic. The UK runs an equivalent regime under UK GDPR and the Data Protection Act 2018.
Other markets have their own cross-border rules. The UAE's federal Personal Data Protection Law and the DIFC Data Protection Law each address transfers, covered in the UAE Personal Data Protection Law and AI agents. Nigeria, Kenya, South Africa, Brazil, India and others have data protection laws with their own approaches.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Who does it apply to?
The organisation deploying the agent is usually the controller and carries these duties. The model provider is usually a processor, with its own duties under the contract and the law. If you use an agent platform that calls a model provider on your behalf, the platform is often a processor and the model provider a sub-processor, and the transfer analysis follows the data through each step.
What changes when the system is an agent?
A person filling in a form decides what data to include. An agent decides for itself, step by step. It reads a customer record, decides what is relevant, and sends it to the model as part of a request. Without controls, it may send a whole history when it needed a postcode.
So the transfer question becomes partly a design question: what may this agent send, for which task, to which provider, in which region? The guide on GDPR and prompt data goes into the detail. Three points matter most:
- Minimisation per task. Each capability should send only the fields it needs.
- Provider terms. Whether the provider keeps prompts, for how long, and whether it may use them to improve its models.
- Region. Many providers offer processing in specific regions, which may remove or reduce the transfer.
How do you meet it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Lawful basis and purpose | Purpose stated per agent capability | Record of processing entry |
| Processor contract | Contract with platform and model provider, listing sub-processors | Executed contracts |
| Transfer route | Adequacy, or contractual safeguards with a transfer assessment | Transfer assessment on file |
| Minimisation | Field-level limits on what each capability may send | Configuration, and logs of fields sent |
| Retention by provider | Contract terms on retention and model training use | Contract clause; provider confirmation |
| Region | Region pinned where available | Configuration record |
| Accountability | Each call logged with provider and region | Filed verdicts |
What does that look like in practice?
Say a German insurer wants its claims agent to summarise medical reports using a model hosted outside the EEA. Health data is special-category data, which raises the bar further. The data protection officer asks for a transfer assessment and finds the provider offers EU-region processing. The insurer pins the EU region, removes names and policy numbers from what the agent sends, and confirms by contract that prompts are not kept for model training. The transfer question largely disappears, and the record shows the region and fields for every call.
The wider rules picture is in our guide to AI agent regulation.
Questions people ask
Does removing names make data anonymous?
Usually not. Data that can still be linked to a person, through a policy number or a rare combination of details, remains personal data. Removing direct identifiers reduces risk but rarely takes the data outside the rules.
Is a transfer assessment needed for every provider?
For each transfer route that relies on contractual safeguards, yes. Many organisations keep one assessment per provider and destination, reviewed when terms or laws change.
What if the agent sends data the brief did not allow?
Treat it as a data incident and assess whether it is reportable. Then fix it with field-level limits outside the model, since an instruction in the brief alone will not hold.