Near-Zero AI Override Rates Are a Warning Sign, Not a Success
Banks celebrating near-zero AI override rates may be measuring friction cost rather than model accuracy.
A European bank's celebrated 0.03% AI credit-decision override rate masked a critical blind spot: the audit never tracked override requests abandoned before completion. Near-zero override rates often indicate systems so burdensome that humans stop trying to intervene, not AI systems that get every decision right. Risk officers should demand vendors split override metrics into two categories—decisions overridden versus requests abandoned—to reveal whether human correction is actually possible.
At a large European bank in 2024, an internal audit flagged an AI credit-decision override rate of 0.03% over 18 months. Leadership treated that number as a compliance success. The audit report did not measure the abandonment rate of override attempts—how many requests were started, found too difficult, and never completed. Lower override rates are routinely celebrated as a sign of accurate, trustworthy AI. But when the rate drops to near zero, the metric is often measuring friction cost, not model quality. An override path that is so expensive, slow, or administratively heavy that people stop trying is not a system that gets every decision right. It is a system that has closed the door on human correction. Near-zero override rates are a warning signal, not a marker of AI accuracy. The number that matters is not just “decisions overridden.” It is “override requests abandoned before completion.” Separate override-rate dashboards into two metrics: 'decisions overridden' and 'override requests abandoned before completion.' Risk officers should demand that split from every vendor. Otherwise, audit-readiness becomes a paper exercise, and the real data path—the one that tells you whether a human can actually intervene—stays unexamined.