Agents that handle money need hard stops. Casefile runs a claim through an extractor, an investigator and a reviewer that can send work back, under a supervisor with a step limit, a send-back limit and a token and dollar ceiling checked before every call. Every payout waits for a person. Pick a claim below and step through what each agent did; problems were planted in some claims as an answer key.
The documents are what the extractor reads. Each step is one agent's turn; the whole claim state is saved after every step, so any run can resume from any point and end the same way.
Across 900 claims, with no false flags on claims without the problem.