Human-in-the-Loop

Explanation

Human-in-the-Loop checks that a qualified person reviews high-stakes AI decisions rather than letting the model act unsupervised, and it watches how often that reviewer disagrees with the model. This matters because accountability rules require a human to own consequential outcomes such as a declined loan, and a sustained pattern of reversals is an early warning that the model is drifting or miscalibrated. It is measured by the Human Override Ratio, computed as Override Ratio = Count(Human_Overrides) / Count(AI_Decisions_Reviewed), specifically for credit decisions with human review of declines. To implement it, route the relevant decisions (for example, all credit declines) into a review queue, capture each reviewer's accept-or-override action with reason codes and reviewer identity, and log timestamps and the model's original recommendation as evidence; the control sits at the decision-egress point of the lending pipeline and feeds a governance dashboard. The threshold is <10%, where a higher override rate indicates model issues; a breach should trigger investigation of the model and the affected decisions, with escalation to model risk owners and roles defined under the AI policy. Because this control is rated Immediate priority, treat any sustained breach as a signal to pause reliance on automated declines pending review.