The problem
- Raw PII in every AI prompt.GDPR Art. 5 minimization principle violated.
- Vendor retains prompt data.Personal data stored by a third party.
- No audit trail for personal data.DSAR requests impossible to fulfill.
The solution
Personal data never reaches the model.
PII stripped and replaced with tokens at the edge before the prompt reaches any LLM. The response is de-anonymized on the return path. GDPR Art. 5 data minimization by design — the model never sees real identities.
GDPR Art. 5(1)(c)
"Personal data shall be adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed — the data minimisation principle."
Every identifier replaced with a consistent, reversible token. Model reasons, never exposes.
Frequently asked questions
Is this anonymization or pseudonymization?
Technically pseudonymization (GDPR Art. 4(5)): tokens are reversible via a map held in your infrastructure. True anonymization (Recital 26) is irreversible. The model only ever sees pseudonymized tokens, never real identities.
Does the LLM still produce useful output without real names?
Yes. PII is replaced with consistent tokens ([PERSON_1], [EMAIL_2]), preserving grammar and relationships, and the response is de-anonymized on the return path so your users see real values.
How does this satisfy data minimization (Art. 5)?
The provider receives only what the task needs — never direct identifiers — which is data minimization enforced at the architecture layer rather than by policy.
Where is the token-to-PII mapping stored?
In your own infrastructure. It is never transmitted to the model, so the re-identification key stays under your sole control.