Escalation Threshold
The rule determining when a model's output is acted on automatically and when it is routed to a person.
Human-in-the-loop is not a binary property of a system; it is a threshold set per decision, and setting it well is what makes automation both safe and worthwhile.
The inputs to the threshold: the consequence of being wrong (a mis-tagged support ticket versus a mis-approved refund), the reversibility of the action, the model's measured accuracy on that task, and the cost and latency of human review.
The design patterns, in ascending autonomy: human approves every action (safe, and it caps throughput at human capacity); human approves above a threshold — by value, by risk, or by model confidence, which is the common and usually correct arrangement; human reviews a sample after the fact, which detects drift without gating throughput; and fully automatic with an appeal route.
What makes the human step effective rather than decorative:
Show the reasoning and the sources, not just the conclusion — a reviewer who cannot see why cannot judge whether.
Design against rubber-stamping. A reviewer approving 500 items an hour is not reviewing; sampling fewer items properly is better than nominally checking all of them.
Feed decisions back into the evaluation set, so overrides become training and test data. This is what allows the threshold to be raised over time with evidence, which is the whole point of the arrangement.