A Market-Conduct-Safe, Evidence-Anchored Decision Packet for AI-Assisted P&C Underwriting Design principles for auditability, adverse-action defensibility, and controlled discretion

Abstract

AI-assisted underwriting in regulated P&C insurance is constrained not by model capability, but by the inability to produce durable decision evidence that withstands market conduct review. Regulators audit files, not models. This paper proposes a Decision Packet: a standardized, auditable, replayable record binding underwriting outcomes to contemporaneous evidence, governing constraints, decision authority, and alternatives considered. We define a protocol comprising a message envelope, evidence bundle contract, decision graph, and compliance replay manifest. Two worked examples (Personal Lines Auto and Commercial Workers’ Compensation) demonstrate implementation. The proposed construct is buildable today and enables scalable, defensible AI-assisted underwriting in regulated environments.