How we handle PII masking at the prompt layer, not just at the response layer

Most PII masking ships at the response layer. The compliance gap is real.

Why prompt-layer masking matters:

- Customer support agents often receive raw PII in inbound messages

- That PII flows into the model context, even if the response doesn't echo it

- Model training feedback, if enabled, can absorb raw PII

- Some compliance regimes require masking before any processing, not after

How we built it:

- Regex + ML-based PII detection at the prompt input

- Token replacement with reversible references stored in an isolated vault

- Model sees masked prompt, response is unmasked using the same vault

- Vault entries auto-delete on configurable retention policy

The result, prompts logged for audit don't contain raw PII. Responses to customers remain accurate.