A strong answer. “Eighty-five percent tells me almost nothing on its own, because the number that matters is not the accuracy — it is the cost of the errors, and that depends entirely on what the feature does.

So my questions. What is the cost of a wrong answer? If it drafts an email a human reviews before sending, 85% is genuinely useful — it saves time and the human is the control. If it auto-approves refunds, 85% is fifteen percent wrong refunds and that is not a product, it is a liability.

Where do the errors fall? Uniformly distributed errors are survivable. Concentrated errors — always wrong for one customer segment, or wrong specifically on the high-value cases — are not, and the aggregate hides it.

Does the user find out? A wrong answer the user can immediately recognise is very different from one they cannot. Confidently-phrased plausible errors are the dangerous class, which is why I care about how the output is presented as much as how accurate it is.

What is the comparison? If the human process it replaces is right 80% of the time, 85% with a review step is an improvement, and nobody has measured the human baseline in most organisations.

So: I would ship at 85% behind a human review step, instrument which errors occur, and use that to decide whether the fully-automated version is ever reachable. And I would define the launch gate as an eval threshold on a golden set with a recorded baseline, not as a vibe — because otherwise the bar moves under commercial pressure.”