This looks to be a big deal for my use case.
I've already got rule-based PII redaction in place, but there are still some leaky or ambiguous edge cases that crop up from time to time. Putting another layer like this in the way (before we send things to the big model) is likely to dramatically improve my ability to sell these tools to management. The block diagram does look much more reassuring once we introduce this step.