A falling override rate can mean weaker human review, not a better agent

If people check an AI agent's work only where the agent raises a flag, the biggest risk is what it never flags. Michele Grant made this point on the AI Engineer podcast, in an episode about whether organizations can trust agents that work. She added that a falling override rate, meaning humans correct the agent less often, can be a warning sign rather than good news.

What was said

Grant used an illustrative scenario: an agent builds reports during an acquisition integration, work that analysts did before. She described four stages of trust. First, the agent runs beside the analysts and the team compares both reports. Second, the agent writes a draft and a senior analyst reviews every line item. Third, the analyst checks land initiatives plus a random sample of cleanups. Fourth, the analyst sees only high-risk anomalies, backed by periodic independent sampling.

She told listeners to pay close attention to that last stage. In her words, "once humans only look at exceptions, your biggest vulnerability is what the agent failed to flag." Her advice was to track missed material issues through independent sampling, monitor error severity and watch the override rate. She also said: "A falling override rate does not automatically mean the agent is getting smarter. It can also mean the depth of review has slipped." Independent sampling, she said, tells you which of the two you are seeing.

Why it matters

Our reading: an override rate is easy to put on a dashboard, and a falling number looks like progress. Grant's point is that the number alone cannot separate a better agent from a tired reviewer. Leaders who approve more autonomy for an agent need a second measure that does not depend on the reviewer's attention.

The other side

Grant's example is illustrative, not a measured case. The excerpts give no data on how often override rates fall because review has weakened. They also do not say how large the independent samples should be, how often to draw them or what they cost. That matters because exception-only review is attractive for saving analyst time.

She also did not argue against exception review. She presented it as the final stage, reached only after the earlier stages. And she said the agent's payoff in her scenario was verified evidence weeks ahead of schedule, which gives leaders time to act. The open question is how much sampling keeps that benefit without giving the savings back.