Trusting AI agents: A reinsurance case study

Sophia Barnes, a Computer Science Master's student at Stanford University, developed a multi-agent AI system to automate and streamline the process of converting unstructured Excel submission pack data into structured catastrophe records in the reinsurance industry. The system employs multiple specialized agents to handle tasks such as extracting catastrophe loss data, matching events to historical records, populating cedant loss data, and comparing it to existing data. The architecture leverages a "human-in-the-loop" approach to ensure accuracy by allowing user intervention and confirmation at crucial steps, which provides flexibility and adaptability to varying submission pack formats. This modular design not only prevents agents from being overwhelmed by excessive information but also facilitates reliable execution through Temporal's Durable Execution, Workflows, and Activities, ensuring that human oversight maintains control over the AI's decisions. The project demonstrates how integrating human input within multi-agent systems can enhance reliability and efficiency, with potential applications beyond the reinsurance sector.