Submissions referred by automated underwriting land here for a human decision. Each carries the climate cat-model score and the rules it tripped. Approve a risk and you can authorize the AI agent — backed by the platform's RAG knowledge system — to auto-clear comparable risks in the same geography.
Behind the "train AI agent" action is the workbench's retrieval-augmented (RAG) knowledge system — the carrier's own proprietary decision memory, built into the platform. Every underwriter decision is written to it as a vectorized record (embedded — stored as numbers the AI can search by similarity):
When a new submission arrives, the agent retrieves the most similar past decisions from this knowledge base (same geography + characteristic + score band) and proposes a decision grounded in precedent — not a black-box guess. The underwriter sets the guardrails; the agent acts within them and queues edge cases back.
Compliance rules are never delegated to the agent. Every retrieval and decision is logged with the model + knowledge-base version for audit.