New Agentic RAG system beats LLMs and naive RAG in commercial underwriting

A new paper from Robert Richardson and colleagues explores how agentic AI architectures can improve straight-through underwriting in insurance. The researchers compare three pipelines on a synthetic small commercial Business Owner Policy (BOP) dataset: a single-LLM baseline, a naive retrieval-augmented generation (RAG) system, and a multi-agent 'Agentic RAG' pipeline that plans, retrieves, calls external tools, and reflects on its own outputs. The agentic system explicitly enforces multi-step rule evaluation and checks third-party data, addressing key actuarial priorities like auditability and transparency.

The results show the agentic RAG system performed best overall, with the largest gains in multi-step and missing-information scenarios. Structured retrieval and reflection helped the model avoid unsupported straight-through decisions, a critical risk in automated underwriting. The study provides a concrete prototype for how regulated industries can deploy LLMs while maintaining human-in-the-loop governance.