Running LLM automations under the GDPR means treating every prompt, retrieved document and output that contains personal data as regulated processing. Minimise the input so the model sees only what the task needs, then confirm a processor agreement, EU hosting options and sub-processor terms before production use.
Key takeaways
- Data minimisation is a legal principle, so an LLM should receive only the personal data the task genuinely needs.
- Pseudonymised data remains personal data under the GDPR when a person can still be identified.
- EU hosting is one factor and does not replace a processor contract, a lawful basis or sub-processor checks.
- Logs, retention settings and automated decision rules must be reviewed before the automation goes live.
What does the GDPR require from an LLM automation?
It requires that the personal data you send and store is limited to what the purpose needs, and that any company processing it for you is bound by a written contract. These are two separate obligations, and an LLM automation has to meet both.
Article 5(1)(c) of the GDPR states that personal data shall be adequate, relevant and limited to what is necessary in relation to the purposes for which it is processed. Article 28 of the same regulation says a controller may use only processors providing sufficient guarantees, and that the processing must be governed by a contract or other legal act with specified terms.
An LLM provider that handles your prompts on your behalf is usually a processor, although the exact role depends on the service and on what the provider does with the data for its own purposes. Your team therefore needs to read the provider's terms rather than assume them.
This article is general information and not legal or tax advice. Rui Codex offers technical implementation support for AI automation projects; it does not provide legal review, compliance audits or certification, so involve qualified counsel for your own interpretation.
How do you minimise personal data in prompts and outputs?
Remove every personal data field that the task does not need before the model call, and limit which documents the model can retrieve. Minimisation is a design decision made before the first prompt is written, not a clean-up step afterwards.
Start from the purpose. If an automation classifies supplier invoices, the model needs the amount, currency, due date and supplier identifier, not the full mailbox history. Write that purpose in one sentence and use it to justify every field you include.
- Write the purpose of the automation in one sentence.
- List every field the prompt template would include, and mark which ones the task truly requires.
- Strip or replace identifiers such as names and contact details where the task allows, using stable placeholder tokens.
- Restrict retrieval so the search step can reach only the documents the requesting user role is allowed to see.
- Set retention periods for prompts, responses and logs, and delete data on schedule.
Pseudonymisation reduces risk but does not take data outside the GDPR. Recital 26 of the GDPR explains that pseudonymised data which could be attributed to a person through additional information remains personal data. Store the token mapping separately and restrict who can reverse it.
Minimisation has a cost. Removing context can lower output quality, especially for free-text tasks such as summarising customer messages. Test the reduced input against a sample of real cases before deciding what to remove, and record the result.
Does EU hosting make an LLM automation GDPR compliant?
No. EU hosting can reduce transfer questions, but it does not establish a lawful basis, a processor contract or control over sub-processors and support access.
Where data is stored and where it is processed are separate questions. A provider may store data in the EU while support staff, monitoring tools or sub-processors access it from other countries. Each of those access points can be a transfer, which Chapter V of the GDPR governs when personal data goes to a third country.
When you assess hosting, ask five concrete questions: where inference runs, where logs and stored files are kept, which sub-processors touch the data, from which countries support staff can access it, and whether the region can be fixed in your contract. Get the answers in writing rather than from marketing pages.
EU hosting fails as a control when the answers are vague. A common failure is a region setting for storage paired with default settings that send diagnostic data elsewhere. Check the default configuration, not only the option you selected.
For the technical side of mapping data flows between your ERP, CRM and AI tools, Rui Codex offers implementation support through its AI automation services and system integration work.
What should a processor agreement cover for an LLM provider?
It should set out the subject matter, duration, nature and purpose of the processing, the types of personal data, the categories of people affected, and your documented instructions. Article 28(3) lists these elements.
Article 28(3) of the GDPR lists the mandatory content of the contract between controller and processor. Use that list as a checklist when you read a provider's data processing terms.
- Processing only on documented instructions from you.
- Confidentiality commitments for staff authorised to process the data.
- Security measures appropriate to the risk.
- Rules for engaging sub-processors, including notice and your right to object.
- Assistance with data subject requests and with personal data breach notification.
- Deletion or return of the data at the end of the service.
- Information needed to demonstrate compliance and to allow audits.
Two gaps appear often. Standard terms may permit use of submitted content for product improvement, and retention periods may differ from your own policy. Read the default settings and the contract together, and record which setting you chose and why.
A contract is only as good as the controls behind it. If staff paste customer records into a chat tool outside the approved integration, the paperwork does not protect that data.
Hypothetical examples of minimisation and hosting choices
Here are two hypothetical scenarios:
- A hypothetical invoice-triage automation first passes entire email threads to the model, including signatures, private phone numbers and quoted histories. The revised design extracts only the supplier identifier, amount, currency and due date, and discards the rest before the model call. The second version is easier to justify under the minimisation principle and easier to explain to a data protection officer.
- A hypothetical support-ticket summariser sends full ticket text to an external model whose support access terms are unclear. The team first confirms in writing where inference runs and which sub-processors touch the data, then limits ticket fields to subject, product and an anonymised description. If those answers cannot be obtained, the team picks another provider or keeps the summariser inside its own infrastructure.
Neither scenario reports a result. Their purpose is to show the order of decisions: purpose, fields, contract, then hosting.
Where do LLM automations fail under the GDPR?
They most often fail through unbounded inputs, logs that keep more than the prompt needs, and outputs acted on without review. Each failure is visible in the design if the team looks for it.
Logs and prompt history
Observability tools often record full prompts and responses by default. Those logs become a second copy of the personal data, with their own retention and access rules. Mask personal fields at the logging layer and give logs the same retention schedule as the source system.
Outputs and automated decisions
Article 22 of the GDPR restricts decisions based solely on automated processing that produce legal or similarly significant effects for a person, subject to listed exceptions and safeguards. An LLM that drafts a rejection letter without a human decision point is a design risk. Where outputs only support someone's work, document that a reviewer checks them.
Inferred data is personal data too. A model that infers a health condition or political opinion from text raises questions about special categories under Article 9 of the GDPR. Avoid prompts that invite such inference, and filter outputs that contain it.
How should a team start an LLM automation project safely?
Start with a data inventory for one automation, settle the minimisation and contract questions, and only then build the integration. A narrow first project keeps the review manageable and gives the team a template for later automations.
Engineering choices follow from that inventory. Restrict credentials to the minimum scope, separate development data from production data, and keep a record of model versions and prompt templates. Those records are useful when a data subject request or an internal audit arrives.
Rui Codex builds AI automation and system integrations for businesses in Belgium, the Netherlands and Luxembourg. For background before a technical discussion, read the security and GDPR engineering articles on the Rui Codex blog.
Conclusion
Running LLM automations under the GDPR is an engineering task as much as a legal one. Minimise what goes into prompts, verify where data is stored and accessed, and make sure the processor terms match how the system actually works. Each decision is cheaper to make before code reaches production. To discuss the technical design of an automation, request a project consultation at https://ruicodex.com/Contact or email info@ruicodex.com to discuss your project. You can also see delivered projects at https://ruicodex.com/CaseStudies.
Frequently Asked Questions
Is EU hosting enough to make an LLM automation GDPR compliant?
No, EU hosting is not enough on its own. You still need a lawful basis, a processor contract that meets Article 28, control over sub-processors, and an assessment of any access from outside the EU. Hosting location answers only one part of the question.
Does pseudonymising personal data take it outside the GDPR?
No, pseudonymised data remains personal data when it can be attributed to a person with additional information. The mapping table must be protected and access to it restricted. Anonymised data is different, because the person can no longer be identified by means reasonably likely to be used, and that standard is demanding.
Do I need a processor agreement with an LLM provider?
Usually yes, when the provider processes personal data on your behalf. The contract should cover instructions, confidentiality, security, sub-processors, data subject assistance, deletion and audit information. Check the provider's default settings for training use and retention, because standard terms may differ from your own policy.
Does the GDPR ban automated decisions made by an LLM?
No, the GDPR does not ban them, but Article 22 restricts decisions based solely on automated processing that have legal or similarly significant effects for a person. Exceptions and safeguards apply. Where an output affects someone's rights, keep a human decision point and document it.
Sources cited in this article
- GDPR (eur-lex.europa.eu)