Insurance Underwriting

An underwriting recommendation affects who gets cover and at what price. Model assistance in that work raises two questions at once, about applicant data and about accountability for the answer. This starter addresses the data question well and leaves the accountability question to your configuration.

The review step is the part you must finish. It ships with no settings, so a recommendation is not routed to anyone until you configure that stage. Treat it as a review point to enable, and not as active routing on day one.

An applicant national identifier number is masked before the prompt travels upstream. Three blocked phrases cover a secret, a client account export and wire instructions. Injection screening runs ahead of both, so a crafted submission cannot turn those checks off.

The market abuse screen looks for insider trading language and for material non-public information in the returned text. A recommendation that reads as guaranteed returns is blocked, and so is an instruction to buy this stock. Weak text fails too. The floor is 0.86 combined, with relevance at 0.78 and accuracy at 0.80, weighted alike.

No fairness signal is configured in this starter. Testing for disparate outcomes, and documenting that testing, remains separate work under state fairness rules and the NAIC model bulletin. Marking at the end of the chain blocks nothing and keeps nothing. Underwriting authority stays with a licensed underwriter, because this starter approves and declines nothing.