An anti-fraud system routes flagged transactions to human reviewers. Audit shows reviewers approved 98% of cases with a median handling time of 3 seconds. The control is documented as human oversight. What is the most accurate finding?

The trigger

Not a bad decision. A queue sized so that the only way to clear it is to approve. Three seconds is enough to see that a case exists and click; it is not enough to open the transaction history, consider the merchant, or form a view that could differ from the model's.

Why it propagated

Three mechanisms, all structural rather than personal:

• Deference to the automated recommendation. When a system presents a suggested outcome, people accept it far more often than they would decide independently, and the effect is strongest under time pressure. Presenting the recommendation is itself a design choice and it is rarely reviewed as one.
• Incentives measured on throughput. If reviewers are assessed on cases cleared, the review converges on the fastest defensible action, which is approval.
• Asymmetric feedback. A wrongly rejected transaction generates an immediate complaint; a wrongly approved one surfaces weeks later as a chargeback attributed to the model. Reviewers learn from one of these and not the other.

Why the other options fail

• "Reviewers are performing well; the model is precise." Possibly true and not evidenced by this data. A 98% approval rate is consistent with a precise model and with an unread queue, and three seconds cannot distinguish them. When two hypotheses fit and one implies the control is absent, the control is not evidenced.
• "Every case received a documented human decision." This is the substitution the finding exists to name: a record of a decision is not a decision. It is also the reason the control passed every previous review.
• "It is a training issue." Training does not create time. A reviewer who understands fraud perfectly still cannot assess a case in three seconds, and framing it as training moves the remediation to the people least able to fix it.

The structural fix

1. Size the queue to the review, not the review to the queue. Decide the time a real review takes, multiply by volume, and either staff it or send fewer cases — which means raising the model's flag threshold and accepting the fraud that passes.
2. Instrument handling time and agreement rate as control metrics, with thresholds. A control whose only metric is completion is a control that cannot fail, which is why this ran for so long.
3. Break the deference deliberately: withhold the model's recommendation until the reviewer records an initial judgement, or route a small sample of unflagged cases so agreement rate has a baseline.
4. Measure the outcome, not the activity. Sample approved cases against eventual chargebacks, so reviewers receive the feedback the asymmetry denies them.

The general lesson, and when not to apply it

Human oversight is a capacity commitment, not a checkbox. Wherever a design says a person will review, the questions that decide whether it is real are: how many items per hour, what does one take, what happens when the queue grows, and what does the person see before they decide. If those four are unanswered, the oversight is documentary — which matters increasingly, because human review is the mitigation most often claimed when an automated decision is challenged.