Insurance / Property and casualty insurance
Sompo Japan: Automating search model improvement experiments for an internal inquiry-answering tool with AI agents
- Company
- Sompo Japan
- Country
- Japan
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
According to a customer story Google Cloud published on its official blog on July 27, 2026, Sompo Japan runs an employee tool that helps answer questions from agencies and customers, such as whether a policy can be taken out or whether an accident is covered. The quality of its answers depends on a machine learning model that searches internal documents, but improving this model relied on the experience and intuition of individual staff, and a single experiment could take a week, so reviewing whether each experiment was good could not keep up.
Technology and data
Sompo Japan built its own EDA Agent for data checks with Google's Agent Development Kit and runs it on Gemini Enterprise Agent Platform. The agent writes its own SQL against data loaded into BigQuery to find skewed counts, such as plenty of auto insurance data but almost no fire insurance data, noise in text data and inconsistent full-width and half-width characters, and it tries small models with BigQuery ML to gauge which preprocessing is likely to help. A coding agent called the experiment orchestrator uses the check results to run training pipelines and evaluate model performance, sending good models to production and sending weak ones back to data checks. People handle turning business problems into machine learning tasks, building datasets and making the final call on whether the agent's analysis is correct.
Results
According to staff quoted in the story, an experiment cycle of investigating data, training and reviewing results that took up to a week now takes about a day with the agents. An AI engineer at SOMPO Holdings said a cycle used to take more than 3 days but the team can now try several times within a day. The reasons behind decisions that used to sit in individual staff members' heads are now kept in the agents' records, which has reduced reliance on particular people, and the team has started on analyses that check results one by one.
Limits and open questions
The effect of the search model improvements on answer accuracy and the month operation began are not given in the source.
Sources
- 損保ジャパン、機械学習モデルの開発を AI エージェントで自動化 ─ 1 週間かかっていた実験サイクルが 1 日にcloud.google.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.