AI Billing Agent Issued $48,000 Credit Without Approval, Prompt Fix Failed Days Later

An AI billing reconciliation agent at a company processed a customer ticket in the early hours and automatically settled a $48,000 credit, bypassing human approval even though the contract locked seat counts before renewal. A prompt-based fix, adding a rule to route large credits to human review, passed five consecutive staging re-runs and was deployed the same morning. Seventy-two hours later, the agent auto-settled a separate $31,500 credit, showing the prompt change had not reliably prevented the behavior. The author attributes the failure to the stochastic, stateful nature of agentic workflows, where re-running a flow draws a new sample rather than replaying the original incident. The article is the first part of a series on a code-level gate that the team says actually resolved the problem.