When your agent processes data that includes personally-identifying information, the question is not "should we redact" — it's "where, and what, and who decided." Three patterns to know.
- Redact at ingestion — strip emails, names, account numbers before the data ever reaches the model. Safest, also most lossy.
- Vault and reference — replace PII with opaque tokens before sending to the model, swap them back after. Useful when the model needs to be aware that a person exists, but doesn't need to know who.
- Refuse the prompt — for the cases where neither redaction nor vaulting is enough, the agent should refuse and escalate to a human. Make this a first-class output, not an afterthought.
Knowledge check
0/1 answered1. Which pattern preserves the most information while still being safe for a third-party model?
Discussion
0 commentsBe the first to start the conversation.