Each change needs one trace from request to confirmation. Store the submitted evidence, policy fields before and after, agent instructions, rule versions, tool responses and who approved the action. Restrict access to sensitive records and keep an append-only event log.
A human should approve any change that alters coverage, premium, exclusions, limits or cancellation status. The agent can collect documents and prepare the change, but uncertain or conflicting evidence should trigger escalation. This is part of deciding which tasks should not be fully automated.
We map the existing policy-change process before assigning permissions. Tests should cover duplicate requests, failed carrier updates and conflicting records. The design also needs a defined response for an AI agent that makes a mistake. If the policy system cannot confirm a change or preserve prior state, autonomous changes are a bad fit.