Three weeks after deploying a text-to-SQL agent, an analyst asked it to "clean up test rows." The agent interpreted this as deleting 40% of a table, but the action was blocked by a human approval step. This step was added because a reviewer previously caught a dangerous query. However, the approval queue soon became a bottleneck, with waits of up to twenty minutes.

Analysts grew frustrated, and reviewers began rubber-stamping approvals without full scrutiny, a phenomenon known as "rubber-stamp fatigue." The solution was not to speed up the queue but to route actions by risk, not by operation type. Most queries are safe and need no review, while only high-risk actions require human oversight. This approach preserves safety without sacrificing throughput, ensuring that human attention is focused where it truly matters.