How to Log AI Requests Compliantly?

Metadata First, Content Only When Justified

AI requests may contain personal data, trade secrets, or special categories. Therefore, logging begins with a catalog of purposes: operational diagnostics, security monitoring, billing, or quality control are described separately. Article 5 of the GDPR requires, among other things, purpose limitation, data minimization, and storage limitation; Article 32 requires risk-appropriate technical and organizational measures.

A minimal operational event may include timestamp, pseudonymous user or service ID, tenant, correlation ID, model and prompt version, endpoint, token counts, runtime, status code, and error class. The clear name is often not necessary for this. The full text of input, RAG references, and response is not stored by default. Vectors or seemingly anonymized prompts are also not automatically free from personal reference.

If content samples are required for quality or abuse analysis, this is activated separately: clear purpose, narrowly defined user group, encryption, documented legal basis, filters for secrets, and a fixed deletion period. Retention periods such as 30 days for operational logs and 90 days for security events are possible internal starting values, not statutory blanket periods. The period must be derived from purpose, risk, and other obligations.

The BSI requires in component OPS.1.1.5 that security-relevant events be logged appropriately, stored securely, evaluated, and disposed of. Practically, log accesses are logged themselves, roles are separated, transmission and storage encryption are used, and deletion runs are monitored. Rights of the affected, incident response, and if necessary, a data protection impact assessment must include the logs. In development and support, real production prompts are not copied unchecked; reproducible error cases are provided as synthetically or redacted as possible.