LLMs, GenAI, RAG, agents, evaluation · insight-1612
Redact personal data before the prompt leaves your boundary and keep the mapping local
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- ML engineer, Applied AI, Data scientist
- Industries
- aviation, healthcare
Summary: situation, action, reason, boundary
Situation
Support tickets and call transcripts with names, emails and booking references go to a hosted model for summarising or classifying.
Action
Detect entities and replace them with typed placeholders such as PASSENGER_1, keep the placeholder map in your own store per request, and restore values only for entitled readers. Measure detector recall per entity type on a labelled sample and set a release threshold.
Reason
The data-minimisation principle in GDPR Article 5(1)(c) asks for data limited to what is necessary, which this supports (a principle, not legal advice). Automated detectors miss entities, so recall must be measured and layered with other protections.
Boundary
Summarisation, classification and extraction where identity is not part of the answer.
Symptom or risk
The provider and your trace store both receive identifiers the task never needed, widening breach, retention and access scope.
Exceptions
Tasks that need the identity itself, where the provider must be approved for that data class.
Trade-offs
Placeholders can hurt quality when context depends on names.
Pitfalls
Redacting the message body but not metadata fields, and measuring recall on clean synthetic text.
Example
Hypothetical 400-ticket labelled sample: recall 99% for emails, 97% for booking references, 88% for names, so names need rules or review before release.
Lessons and concepts
Claims and their sources
GDPR Article 5(1)(c) states data minimisation, meaning personal data adequate, relevant and limited to what is necessary.
Presidio documents that its automated detection cannot guarantee finding all sensitive information.