TITLE:
Hierarchical Multi-Agent Systems for Automated Commercial Property Insurance Underwriting
AUTHORS:
Muhammad Imran Sajid
KEYWORDS:
Multi-Agent Systems, Large Language Models, Insurance Underwriting, Commercial Property Insurance, LangGraph, Document Classification, Risk Assessment, LangChain, Kingdom of Saudi Arabia
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.6,
June
22,
2026
ABSTRACT: Commercial property insurance underwriting requires integrating evidence from heterogeneous document sources—financial statements, property inspection reports, legal certificates, and claims histories—a process that is time-consuming, labour-intensive, and prone to inconsistency. This paper presents a hierarchical multi-agent system that automates end-to-end commercial property underwriting using Large Language Models (LLMs), designed and evaluated in the context of the Kingdom of Saudi Arabia (KSA) regulatory framework. The proposed framework orchestrates nine specialised components through a directed LangGraph StateGraph workflow, processing three categories of input artifact: a structured JSON proposal form, an unstructured PDF/DOCX document bundle (financial, inspection, legal, and claims documents plus reference PDFs), and a domain-knowledge Excel template encoding Risk-Score Rating (RS-RR) criteria. A DocumentClassifier powered by GPT-4o-mini routes each document page into one of eight semantic buckets before four specialist GPT-4o agents analyse their respective domains in sequence. A ReferenceContextBuilder distils KSA underwriting guidelines, policy wordings, and exclusion clauses from reference PDFs, while a separate RiskEngineeringExtractor derives per-category scores across 26 fire-protection and physical-risk dimensions from the Excel template. A SynthesisAgent consolidates all specialist analyses and a DecisionAgent applies the organisation’s risk methodology to produce a final risk rating and premium recommendation in Saudi Riyals (SAR). The system generates seven structured output artifacts including a KSA-aligned underwriting pack, a full LLM interaction log, and a scored risk-engineering assessment. Across five expert-designed commercial property proposals spanning the risk-class spectrum, the system completed each end-to-end underwriting in under three minutes, produced bit-identical outputs on repeated runs through a deterministic premium calculation, and assigned the risk class anticipated by the underwriters who designed each case. These results support the feasibility of bringing LLM-based multi-agent underwriting into regulated insurance markets; a larger expert-validated evaluation is identified as future work.