What practical steps can be taken to avoid exposing sensitive financial/PII data when using LLMs for tasks like loan modeling?

Two main strategies were described. First, the team is working on PII scrapers (with input from another member) to strip sensitive data before it ever goes into an LLM, though this is still in review. Second, in practice, the safest approach is to manually limit what's shared — for example, when modeling loans, only a screenshot of the specific paragraph describing the day-count convention and interest calculation assumptions was dropped into the LLM, deliberately excluding other sensitive details like the balloon payment amount in some contexts. This was acknowledged as somewhat of a 'cop-out' answer since it currently relies on the user deciding what's appropriate to share. Longer-term, the plan is to build tools that automatically extract/redact sensitive data before it's sent to LLMs, and eventually to consider hosting their own models on their own servers, though that hasn't been tackled yet. The concern raised is that anything dropped into an LLM could resurface during a lease audit or an adversarial situation with a national tenant.