How should you handle a mistake that came from AI?

The instinct to explain the origin is understandable and it makes the situation worse. To the person affected, the relevant facts are what was wrong, what is being done, and whether it will recur. Adding that a system produced it introduces a party they have no relationship with and implies a distribution of responsibility they did not agree to. It also raises an obvious question — who checked it — that the explanation was meant to avoid.

So the external handling is unchanged. Correct it, say what happened in terms of what was wrong rather than how it was produced, and say what will prevent a recurrence. That last part is where the origin becomes relevant, because the prevention is different from the prevention for a human error, but it can be stated without attribution: a check has been added, this now goes through review.

Internally the origin matters a great deal and it should be recorded specifically. What produced it, what the input was, why the check that existed did not catch it, and whether the same failure could occur elsewhere. That last question is the one that distinguishes this from an ordinary mistake: a systematic error affects every item processed the same way, so a single visible instance may indicate a set nobody has looked at.

That scope question should be answered before the correction is complete. If a workflow mislabelled one record, it likely mislabelled others. If a generated document contained a wrong figure, the same figure may be in the previous four. Checking the population rather than the instance is the specific discipline this kind of error requires, and it is the part most often skipped because the immediate problem has been solved.

The control that follows should be the one that would have caught this class rather than this instance. Adding a review step for one type of document is narrower and more durable than an instruction to be careful, and a deterministic check on the specific property that was wrong is better than either. The test is whether the same class of error could pass again, and the answer should be no rather than less likely.

One thing to resist: removing the automation entirely as a response to a single error. That may be right and it should be a decision made on the measured rate and consequence rather than on the salience of one incident. A process that produced one visible error and a hundred correct outputs may still be better than the manual alternative, and reversing it on the strength of the error you noticed is how a business ends up with neither the automation nor the control.