When you need this
- Customer support bot sees ticket PII
- GDPR review flagged LLM data flows
- Logs contain names and IBANs from prompts
- Need DE-specific entity recognition
Prerequisites
- Gateway or app middleware hook before provider call
- Entity types list for your domain (incl. DE IBANs, Steuer-ID patterns)
- Legal stance on tokenization vs blocking
Tools
Run at inference boundary, not only in batch pipelines.
Deploy as single ingress before adding a second LLM vendor.
Steps
- 1
Map data flows
Document every field that enters prompts and logs. Classify PII, financial, and health-adjacent fields with the DPO.
- 2
Deploy Presidio analyzer + anonymizer
Run as sidecar or pre-gateway service. Configure DE/EU entity recognizers; tune confidence thresholds on real tickets.
- 3
Block or redact before provider call
Replace detected entities with tokens; keep rehydration mapping inside your trust boundary only if operators need it. Prefer block for high-risk fields.
- 4
Cover logs and traces
Apply the same redaction to Langfuse/OTel attributes — otherwise PII bypasses the gateway via observability.
- 5
Audit and test quarterly
Red-team with synthetic PII payloads. Log redaction events for compliance evidence; track false-negative rate.
Adoption pitfalls
- Redacting prompts but not retrieval chunks
- Storing raw prompts in eval datasets
- Over-redaction that breaks task quality with no tuning loop
Adoption checklist
- No raw PII reaches third-party LLM logs
- Redaction failures alert on-call
- Data flow diagram updated for audits
- Legal sign-off on tokenization approach
SEER REAL assessment / sprint
Assessment traces PII into LLM providers and logs. Sprint typically puts Presidio in front of LiteLLM for one high-risk app and adds a quarterly red-team.