Abstract
Abstract
Purpose.
Many operational AI systems act automatically when a model’s confidence exceeds a threshold and defer to a human otherwise. This paper argues that the operating threshold is an AI risk-acceptance control the level of residual automated error an organisation tolerates without human review and asks how it should be set and governed rather than left as a tuning parameter.
Methods.
We formulate accept-or-defer automation as selective classification under asymmetric error costs and derive the operating threshold from an explicit expected-cost objective, situating the choice within AI risk-management standards and accountability duties. We demonstrate the method on two real deployments in a regulated professional-services practice: a document-classification workflow and a confidence-gated entity-resolution system.
Results.
The cost-derived operating point is empirically robust: across a plausible range of cost ratios the threshold stays within a narrow band, and the entity-resolution scores are sharply bimodal, so only a small fraction of cases require human judgement. Per-category thresholds and calibration further stabilise the operating point.
Conclusion.
Treating the threshold as a documented, justified and periodically reviewed risk-acceptance control makes delegated machine authority auditable. The approach generalises to any high-stakes setting in which AI acts above a confidence threshold.