Automated invoice approval system faced audit questions it could not answer

A developer who helped build an agent that auto-approved clean supplier invoices was later asked by finance why one approval occurred, and the logs recorded only the outcome, not the reasoning. The team realized they had logged what the system did but not why, which data it used, or which rules and model versions were in effect. Because financial entries must be traceable to rules and accountable parties long after code changes, the team began recording inputs, rule and model versions, confidence, and a plain-language reason for every decision. They also separated the agent from the ledger, so a rule-checking service screens proposals and routes risky cases to humans, and they restricted the agent's access to approved actions. Matching rules and prompts are now versioned and reviewed like code, since configuration changes can amount to financial policy changes.