Multi-agent RAG for automated underwriting

Artificial intelligence is increasingly making its way into actuarial practice—a field that requires the analysis of unstructured documents and heterogeneous data, as well as compliance with strict regulatory procedures. Actuaries now have a wide range of design possibilities, from traditional rule-based automation to large language models (LLM), RAG (Retrieval-Augmented Generation) systems, and multi-agent architectures capable of planning, retrieving data, invoking tools, and reflecting.

An Underwriting Experiment

To test the practical applicability of these new architectures, the authors developed an agent-based framework for automated underwriting (straight-through underwriting) of small commercial policies known as Business Owner Policies (BOPs). In a synthetic but realistic environment, the researchers compared three pipelines:

• a baseline single LLM;
• a naive RAG system;
• a multi-agent Agentic RAG, combining targeted search, third-party source verification, and explicit multi-step rule evaluation.

Results

The agent-based system demonstrated the best overall results. The greatest improvement in quality was observed in multi-step scenarios and situations with missing information. According to the authors, structured search and a reflection mechanism help the model avoid unfounded decisions in automated processing.

Source: cs.AI updates on arXiv.org