We're piloting Agentforce for inbound case triage. Need to ground the model on knowledge articles + recent case history, but our compliance team flagged that the unfiltered case history can include unredacted PII.
Two-stage retrieval has been the cleanest pattern: pull the case history through a redaction layer first (the Einstein Trust Layer handles a lot of this by default but you can extend it), then ground the agent prompt on the redacted versions. Your compliance team should be able to approve the redaction policy once, not every prompt.
We added an Apex pre-flight on every grounding call that scrubs against a regex library + a NER model for names. Overkill for some teams but compliance loved it.