Your AI Agent Passed Every Eval. Finance Still Killed It.

AI agents can pass every technical quality evaluation yet still be shut down for economic reasons. A customer operations agent was terminated because its fully-loaded cost per resolved ticket exceeded that of the human team it was meant to replace. The core problem is that standard evaluation harnesses measure if an agent is correct but not if it pays for itself. The crucial missing metric is "cost per successful outcome," which accounts for the cost of failed attempts and escalations to humans. This economic reality, not just technical performance, ultimately determines whether an AI agent survives in a production environment.