Skip to content

Sovereignty and law · 6 min read · As of November 2026

AI and the GDPR in business: five criteria to check

Whether an AI application fits the GDPR follows from its architecture and contracts. What a provider promises can be checked on five points.

Companies that use AI process personal data sooner than they expect: in emails, contracts, support tickets or job applications. The same duties apply as for any other data processing. The question “are we allowed to do this?” has no blanket answer. It can be settled article by article under the GDPR. This piece sorts out the criteria on which a provider's data protection promise rests.

At a glance

  • Whether an AI application fits the GDPR follows from an architecture and contract decision. A model does not bring that along by itself.
  • Five criteria decide: legal basis, a data processing agreement, a sound answer on third-country transfers, data minimisation without silent retraining, and data subject rights that work technically.
  • A data centre in the EU is necessary, and it must be combined with control over the technology. What also counts is who controls access to the model and the data.

Why a data protection promise proves nothing without criteria

Any provider can claim that its AI is lawful. A buyer who cannot check the claim is buying an assertion. The five criteria below are the points where a data protection officer or legal counsel would start. They do not replace a legal review of your case. They show what matters.

No. Criterion How to recognise it
1 Legal basis (Art. 6) A documented balancing of interests (Art. 6(1)(f)) or a voluntary, informed, revocable consent. A blanket “AI clause” in the terms and conditions is not enough
2 Data processing (Art. 28) A data processing agreement with every AI provider in use, covering instructions, sub-processors, technical measures and deletion duties
3 Third-country transfer (Art. 44 et seq.) A sound answer on whose law applies to the provider: certification under the Data Privacy Framework, standard contractual clauses with additional safeguards, or avoidance through self-hosting in the EU
4 Data minimisation and purpose limitation (Art. 5) No silent reuse of inputs for model training without a legal basis of your own, plus traceable pseudonymisation and tenant separation
5 Data subject rights (Art. 15 to 21) A technically working answer to a deletion request, beyond what the contract text says

Legal basis and data processing

Every processing of personal data needs a legal basis under Art. 6 GDPR. In practice this is usually legitimate interest (Art. 6(1)(f)) or consent (point (a)). Legitimate interest requires a documented balancing against the interests of the data subjects. In its Opinion 28/2024 on AI models, the European Data Protection Board confirmed a three-step test for this. A blanket AI clause in the terms and conditions does not carry that burden.

When an external AI provider processes data on your behalf, it becomes a processor. You then need a data processing agreement under Art. 28. Many cloud services offer one by now. The real weak spot usually lies elsewhere.

The third-country transfer

If the provider is based in the United States or belongs to a US group, the third-country transfer under Art. 44 et seq. applies. Since 2023 the EU-US Data Privacy Framework provides an adequacy decision again, and transfers to certified companies are in principle permitted. The framework is legally contested. The General Court of the EU dismissed an action against it in 2025, and an appeal is pending before the Court of Justice.

It does not solve the core problem. After the Schrems II ruling, standard contractual clauses alone are not enough either when the recipient is subject to US law. The US CLOUD Act and FISA 702 oblige US-controlled providers to hand data to authorities even when the servers are in Europe. Neither the framework nor contract clauses remove that power of access. This is where many “EU region” offers lose their footing.

Data residency in the EU therefore says nothing yet about who controls the technology. A data centre in Frankfurt does not protect against extraterritorial access if the operator is subject to the law of a third country. A clean route is self-hosting: open language models with weights you control run on your own servers in the EU. With self-hosting on your own servers in the EU, the data does not leave your own environment, and the third-country transfer can be avoided through the structure. More in the article Data residency and sovereignty.

Data minimisation, purpose limitation and no silent retraining

The principles of Art. 5 apply to AI as well. Only data needed for the specific purpose may be processed (data minimisation), and only for that purpose (purpose limitation). A common violation is silently retraining models on customer or employee data, which is not allowed without a legal basis of its own. With self-hosted models the company decides whether, with what and when to train or fine-tune. With many public cloud services you are left with contractual promises.

Data subject rights and deletion

Data subjects have the right of access (Art. 15), rectification (Art. 16), erasure (Art. 17) and objection (Art. 21). Once personal data has flowed into the weights of a model, deleting a single item cleanly is hardly possible in practice. This is a central risk of generative systems.

An evidence-based architecture reduces that risk. Knowledge lives in a controlled, searchable knowledge base. The model retrieves it at run time and attaches its source to every statement. A deletion request is met in one place: the record leaves the knowledge base and the AI can no longer cite it. How the check of the evidence works technically is described in the article RAG with sources.

Technical measures, impact assessment and the EU AI Act

Under Art. 32, technical and organisational measures are mandatory: encryption, access control, tenant separation, logging and a roles and permissions concept. For processing likely to result in a high risk, Art. 35 requires a data protection impact assessment before going live, and AI is often a candidate. In parallel the EU AI Act applies, which regulates AI systems by risk class and becomes applicable in stages. GDPR and AI Act apply side by side. The details are in the article The EU AI Act for mid-sized companies.

The framework keeps getting denser. Anyone who sets up an architecture today that builds in data control, traceability and controlled deletion from the start will have to retrofit less later. The overall picture is on the page about sovereign infrastructure.

Frequently asked questions

Can a US cloud service be used in line with data protection law?

Only under conditions and with a residual risk. You need a legal basis under Art. 6, a data processing agreement and a sound solution for the third-country transfer. Transfers to providers certified under the EU-US Data Privacy Framework count as formally permitted. The framework is legally contested and does not stop access by authorities, because US providers are subject to the CLOUD Act and FISA 702. For sensitive personal data, self-hosting open models is the cleaner route. A binding assessment for your case comes from your data protection officer.

Is a server in the EU enough for data-protection-compliant AI?

The location shows where data sits. If the operator is subject to the law of a third country, authorities can compel access even to data stored in Europe. What counts is who has technical and legal access.

How do you meet the right to erasure with AI systems?

If a model was trained on personal data, deleting single items is hardly possible in a clean way. An evidence-based architecture solves this because the knowledge sits in a controlled knowledge base, outside the model weights. A deletion request is met in that one place.

Which five criteria should you ask a provider about?

The legal basis for each use case, a signed data processing agreement, a reliable answer on the third-country transfer, the commitment that your data is not secretly used for training, and a technically traceable way to carry out a deletion request. If one of these has no answer, the data protection promise stays unchecked.

Questions about your GDPR check?

We go through the five criteria with you for your own use case.