GDPR Checklist for LLM Applications (2026)

Every LLM feature you ship handles personal data — even if you did not intend it to. Users paste emails, contract clauses, internal IDs and customer names into prompts. Under the GDPR that is processing, and it triggers the same obligations as any other data flow in your product.

This checklist condenses what a DPO will actually ask you before signing off on an LLM feature. Use it as an internal review template.

1. Establish a lawful basis for prompt data

Article 6 requires a lawful basis for every processing activity. For most B2B LLM features that means legitimate interest (documented via a balancing test) or contract when the feature is part of the service the customer paid for. Consent is rarely the right basis for an internal productivity tool.

2. Apply data minimization at the proxy layer

The single most effective control is to strip PII before the prompt reaches the model provider. This is where a PII detection API or a governance proxy like Privaro fits — every prompt gets scanned, entities get tokenized or masked, and only sanitized text hits OpenAI, Anthropic or Gemini.

3. Run a DPIA for high-risk features

Article 35 requires a Data Protection Impact Assessment when processing is likely to result in high risk. Any LLM feature that touches health, finance, legal or biometric data falls in scope. The DPIA should cover:

• Categories of personal data flowing into prompts
• Model provider and sub-processors (with SCCs on file)
• Retention of prompts, completions and embeddings
• Mitigations (tokenization, human review, opt-out)

4. Lock down retention

Model providers cache prompts. Vector databases retain embeddings. Log stores keep completions. Define a retention policy per surface and enforce it with automated deletion — do not rely on manual cleanup.

5. Handle third-country transfers explicitly

Most LLM providers process data in the US. You need Standard Contractual Clauses on file, and for sensitive categories a Transfer Impact Assessment. EU-region endpoints from Azure OpenAI or Anthropic on Bedrock EU can eliminate the transfer altogether — worth the migration for regulated verticals.

6. Give data subjects a working DSAR path

If a user requests deletion, you need to prove that their prompts, completions and derived embeddings are gone. That requires request-level audit logs from day one, not something you bolt on when the first DSAR arrives.

Next step

If you want to see how your current LLM usage stacks up against this checklist, run our free AI risk assessment or read the AI compliance software overview for the control mapping to GDPR and the EU AI Act.