A public-health intake agent stores full conversation transcripts for quality review, and those transcripts must have direct patient identifiers removed before they are retained long-term. Which capability fits this requirement?

Sensitive Data Protection is designed to discover and redact sensitive data patterns, including identifiers, in stored content, which is precisely the data-at-rest task described: transcripts that already exist and need identifiers stripped before long-term retention. Model Armor operates on the live conversational path, screening prompts and responses in real time for manipulation or leakage as an interaction happens, which is a different moment in the data's lifecycle than scrubbing an already-captured transcript before storage. Agent Gateway monitors and can rate-limit traffic, including the volume of storage requests, but it has no role in inspecting or transforming the content of what's being stored, so it can't remove an identifier from a transcript. Agent Identity restricts which principal can read a stored transcript, which is an access-control measure that limits who can see the identifiers, but it doesn't remove them, so anyone with legitimate read access would still see raw patient identifiers the requirement says should be gone. A scope caveat: redaction accuracy depends on the identifier patterns configured for detection, so an unusual identifier format could slip through and needs periodic review. As an operational check, a reviewer can pull a retained transcript sample and confirm no direct identifier pattern remains before it is used for quality review.