GDPR and AI: why running locally is the simplest answer

Every time an employee pastes a customer email, a contract or a patient record into a cloud AI service, personal data is processed by a third party. That is not forbidden. But it drags along a chain of questions someone in your organisation must be able to answer. With a local language model, most of them disappear.

The questions you otherwise have to answer

Who is the processor? The cloud provider processes data on your behalf, so you need a processor agreement under Article 28. The big providers have standard agreements, but it is you who must be able to show that the terms are adequate, that sub-processors are known, and that data is deleted when you say so.

Where is the data processed? Many AI services run fully or partly outside the EU. Then the third-country transfer rules in Chapter V of the GDPR apply. Since the Schrems II ruling in 2020 and the subsequent EU-US Data Privacy Framework, the legal position has shifted several times. The Framework survived its first annulment challenge in September 2025, but the appeal is now before the Court of Justice. What is approved this year may be challenged the next.

Who can compel disclosure? The US CLOUD Act obliges providers under US jurisdiction to hand over data in their possession, custody or control, regardless of whether it sits inside or outside the United States. Picking the "EU" region in the cloud console does not by itself put the data out of reach. How far the Act extends to EU-established subsidiaries of US groups turns on who actually controls the data, and the EDPB and EDPS have called that question unresolved.

Is your data used for training? Terms differ between consumer and enterprise agreements, and they change. Someone has to read them and keep watching them.

Can you honour data subject rights? If a customer requests erasure or a copy of their data, can you show what was sent to the AI service and have it deleted there?

None of these questions is impossible. But each one costs time for legal, IT and the business, and the answers must be documented and kept current.

What happens with a local model

The model runs on your server, in your network. Then it looks like this:

• No third party processes the data. No processor agreement for the AI processing itself.
• No transfer to a third country. Chapter V does not apply to that part.
• No foreign legislation that can compel disclosure from a provider, because there is no provider in the chain.
• No risk that your data is used to train someone else's model.
• Data subject rights are handled in your own systems, which you already have routines for.

What remains is what you are already responsible for in every system: a legal basis for the processing, access control, logging, retention periods, and an impact assessment if the processing is extensive or sensitive. That is a much shorter list, and it sits entirely within your control.

What it does not solve

A local model does not make the processing automatically lawful. If you feed in personal data you have no right to process, it is just as wrong locally as in the cloud. And a server without access control is a leak waiting to happen. Local is simpler, not exempt.

The AI Act, in addition, sets requirements based on what you use AI for, not where it runs. An HR system that ranks candidates is a high-risk system under Annex III whether the model sits in Växjö or Virginia. The Digital Omnibus pushed the obligations that follow from that classification back to December 2027, but it did not change the classification itself, and the systems you build this year are the ones that will have to meet them.

Practical advice

If you already have a data protection function: bring them in early. A local model is often the proposal they are most relieved to receive. If you do not: start with a simple inventory of which types of data would pass through the model. That decides both how you should configure access and whether you need an impact assessment.

It is also one of the steps included in the assessment.