GDPR Compliant LLM: Transfers, DPAs and EU-Only Inference

8 min read

A GDPR compliant LLM is not a product you install. The GDPR regulates the processing of personal data, and the controller, the organization deciding why and how data is processed, carries the duty to comply. A model or a platform can make that easier or harder. It cannot be compliant on your behalf.

For LLMs, the hard GDPR questions cluster in two places. The first is the same as for any system: lawful basis, minimisation, rights, retention and a risk assessment. The second is specific to hosted AI: personal data in prompts goes to a third party, often a processor with servers or support staff outside the EU. Running inference inside your own EU infrastructure removes most of that second group for the inference step. The first group stays with you.

This post is general information, not legal advice. Work through the specifics with your data protection officer or counsel.

Where personal data shows up in an LLM system

Map the data before you map the law. Personal data reaches an LLM through:

  • Prompts, including text users paste in: emails, tickets, CVs, contracts
  • Documents retrieved into context by a search or RAG pipeline, and the index built from them
  • Logs, if they store request or response text
  • Outputs, which can contain or infer personal data
  • Evaluation and fine-tuning datasets
  • The model itself, if it was trained on personal data

That last point is not theoretical. In Opinion 28/2024, the European Data Protection Board (EDPB) said that AI models trained on personal data cannot in all cases be considered anonymous, and that anonymity has to be assessed case by case.

Lawful basis and data minimisation

Every processing purpose needs a lawful basis under Article 6. For internal LLM use, such as summarizing support tickets or drafting replies, many organizations rely on legitimate interests. That requires the three-step test the EDPB restates in Opinion 28/2024: a legitimate interest, processing that is necessary for it, and a balancing test showing that people's rights do not override it. The necessity step has teeth. If the purpose can be met without personal data, processing it is not necessary.

Special category data under Article 9, such as health data, also needs one of the Article 9 conditions on top of an Article 6 basis.

Minimisation (Article 5(1)(c)) turns into concrete engineering choices:

  • Strip or pseudonymise identifiers before text reaches the prompt where the task allows
  • Send the relevant passage, not the whole file
  • Scope retrieval so a query can only pull documents the user may already see
  • Do not log full prompt text by default

Processors and DPAs

When a vendor processes personal data on your behalf, it is a processor, and Article 28 requires a contract, usually called a data processing agreement (DPA). It must cover the subject matter, duration, nature and purpose of processing, the processor acting only on your documented instructions, confidentiality, security, sub-processors, help with data subject rights, and deletion or return of data at the end.

For a hosted model API, read three things in that DPA: where data is processed and stored, which sub-processors are involved, and whether the provider uses your data for its own purposes such as training. Processing for its own purposes falls outside a processor role.

Self-hosting changes the cast. If the model runs on your hardware, no outside party processes the prompts. If it runs in your own cloud account, the cloud provider is still your processor and the DPA you already have with it applies. A software vendor whose staff never access the data is not processing it, as long as support works that way in practice.

International transfers after Schrems II

Chapter V of the GDPR restricts transfers of personal data outside the European Economic Area. In 2020 the Court of Justice's Schrems II judgment (Case C-311/18) invalidated the EU-US Privacy Shield. Standard contractual clauses survived, but exporters must assess whether the destination country's law undermines them and add supplementary measures where it does. The EDPB's Schrems II FAQ and its Recommendations 01/2020 on supplementary measures set out how.

In July 2023 the Commission adopted the EU-US Data Privacy Framework adequacy decision, which allows transfers to US organizations certified under the framework. The EU General Court dismissed a challenge to it in September 2025, and an appeal is pending before the Court of Justice. The framework is in force today, but it has been challenged before and the outcome is open. A design that does not depend on it is easier to defend.

One detail catches teams out. The EDPB's FAQ notes that providing access to data from a third country, for instance for administration purposes, also counts as a transfer. A provider with an EU data center can still create a transfer through remote support staff or sub-processors elsewhere. Residency and jurisdiction are different questions; the sovereign AI post explains why an EU region of a US provider still falls under the CLOUD Act.

Building a GDPR compliant LLM setup inside EU infrastructure

Here is how the main GDPR questions change when inference runs on infrastructure you control in the EU:

Question Hosted model API Self-hosted in your EU infrastructure
Transfer outside the EEA for inference Depends on provider regions, support access and sub-processors None, if hardware, storage and admin access stay in the EU
Processor for inference The API provider, plus its sub-processors None on your own hardware; your cloud provider if in the cloud
Provider use of data for its own purposes Check the terms Not applicable
Prompt and log retention Provider's policy plus yours Yours alone
Lawful basis, minimisation, rights, DPIA Your responsibility Your responsibility

The last row is the point. Keeping inference in the EU settles transfer and processor questions for that step. It does nothing for lawful basis, minimisation or data subject rights.

Data subject rights

People can ask to access, correct or erase their personal data (Articles 15 to 17), object to processing (Article 21), and not be subject to decisions based solely on automated processing that significantly affect them (Article 22). You generally have one month to respond.

For an LLM system that means three things. Logs and stored prompts must be searchable by person, or you cannot answer an access request. Stored text must be deletable. And any workflow where a model's output decides something about a person, such as screening applicants, needs real human review, not a rubber stamp.

Fine-tuning on personal data deserves extra caution, because removing one person's data from a trained model is far harder than deleting a row from a database.

DPIAs

Article 35 requires a data protection impact assessment (DPIA) when processing is likely to result in a high risk to people. The EDPB-endorsed WP248 guidelines list nine criteria, among them innovative use of new technology, large-scale processing and sensitive data. Processing that meets two criteria will usually need a DPIA. An LLM over customer or employee data often meets two.

A useful starting point is the EDPB support pool report on AI privacy risks and mitigations for LLMs. It was written by an external expert and does not represent the EDPB's official position, but it walks through LLM-specific risks and controls.

Retention of prompts and logs

Storage limitation (Article 5(1)(e)) means you keep personal data no longer than the purpose needs. Logs are where this slips. Suppose an internal assistant handles 2,000 requests a day and you keep full prompt text for 30 days:

2,000 requests x 30 days = 60,000 stored prompts, each holding whatever a user pasted in.

Split the log in two. Metadata (timestamp, user or token, model, token counts, latency, status) is low risk and can be kept longer for security and audit. Full text should be off by default, or kept separately with a short, documented retention period and tight access.

The EU AI Act timeline, briefly

The AI Act sits alongside the GDPR rather than replacing it. Per the European Commission's AI Act page:

  • 1 August 2024: entered into force
  • 2 February 2025: prohibited practices and AI literacy duties apply
  • 2 August 2025: obligations for general-purpose AI models apply
  • 2 August 2026: most remaining provisions, including transparency rules, apply
  • 2 December 2027: high-risk obligations for sensitive use cases such as employment and education, as extended by the AI Omnibus
  • 2 August 2028: high-risk obligations for AI embedded in regulated products

Check the page before planning around a date. The timeline has already moved once. Our LLM governance guide covers deployer duties in more detail.

GDPR LLM checklist

  • Data map covering prompts, documents, index, logs, outputs and training data
  • Lawful basis recorded per purpose, with a legitimate interests assessment where used
  • Article 9 condition identified for any special category data
  • Minimisation built into prompts and retrieval
  • Article 28 DPA with every processor, including your cloud provider
  • No transfers outside the EEA for inference, or a documented transfer tool and assessment
  • Vendor support access reviewed for hidden transfers
  • Access, erasure and objection requests can be met against logs and stored text
  • DPIA screened and, where needed, completed
  • Retention periods set separately for metadata and full text

Where AI Refinery fits

AI Refinery, a self-hosted LLM platform, installs inside your own infrastructure: an EU cloud region, your VPC, on-premises servers or an air-gapped network. Nothing phones home, so prompts and responses stay in the environment you choose. That supports a GDPR program by keeping the inference step inside infrastructure you control and taking a model vendor out of the processor chain. It does not make you GDPR compliant, and it carries no GDPR certification. Compliance remains your organization's responsibility.

It provides RBAC, SSO, approval workflows and a full audit log, and every model exposes an OpenAI-compatible endpoint, so applications already using a hosted API can move by changing a base URL. For document search you pair it with a vector database of your choice, as described in RAG with a local LLM.

It is not the right fit if you need the most capable frontier models and are comfortable with a hosted provider's DPA and transfer terms, or if you have no team to run GPU servers. The private LLM vs OpenAI API comparison weighs that choice, and the self-hosted LLM guide covers what running your own model takes.