How do you protect sensitive data (PII) when using third-party LLM APIs?

30-second answerSay your answer out loud first, then reveal.

Minimise what you send. Detect and redact or pseudonymise PII before the call (replace with reversible tokens like <PERSON_1>, and re-insert after the response if needed). Classify data sensitivity and route the most sensitive workloads to self-hosted or approved private deployments. Use enterprise agreements (no training on your data, zero or limited retention, regional processing), encrypt in transit, control and redact logs, and audit everything.

PII protection flow: input goes through PII detection and a sensitivity check; restricted data goes to a self-hosted model, confidential data is pseudonymised using an encrypted short-TTL token map, sent to a zero-retention regional external LLM API, and re-identified in the response.

Considerations

Detection quality: Indian identifiers (Aadhaar, PAN, GSTIN, IFSC), names in multiple scripts, free-form addresses. Combine regex, NER models and allow/deny lists, and evaluate recall.
Pseudonymisation vs masking: consistent placeholders preserve meaning ("PERSON_1 called PERSON_2"), so the model can still reason. Pure masking ([REDACTED]) loses relationships.
Context leakage: retrieved documents and tool outputs also contain PII, so apply the same pipeline to them.
Logs and traces: often the biggest leak. Redact before logging; restrict access; apply retention limits.
Contracts and settings: data processing agreements, no-training clauses, retention settings, data residency (e.g. India region for regulated sectors).
Output checks: ensure responses don't expose other users' data.

PreviousDesign an AI writing assistant embedded in a document editor (like a "copilot" in Google Docs or Notion).

Little by little, you're building something great.

How do you protect sensitive data (PII) when using third-party LLM APIs?

Considerations

Related