Sompo Japan Insurance Inc.

Sompo Japan Insurance faced challenges in handling large volumes of internal inquiries efficiently. Staff needed to manually search through extensive manuals and Q&A documents to formulate responses, resulting in time-consuming inquiry handling processes.

In October 2024, Sompo Japan launched a trial of 'Oshison LLM' (provisional name), an AI system built on a large language model trained on the company's internal manuals and Q&A data. The system automatically generates optimal answer candidates with source references, allowing staff to efficiently craft responses. Proprietary text structuring techniques were used to improve AI accuracy.

The trial was rolled out to approximately 1,000 employees nationwide. Initially limited to auto insurance and liability insurance inquiries, with plans to expand coverage. The system is expected to reduce time spent on inquiry responses and improve overall operational productivity.

• Training LLMs on domain-specific internal documentation (manuals, Q&A) can meaningfully accelerate knowledge worker response workflows.
• A human-in-the-loop design—where staff verify AI-generated answers and sources before sending—balances efficiency gains with quality control.
• Phased rollout by product line (starting with auto and liability) allows controlled validation before broader deployment.