You are defining escalation criteria for ambiguous cases in an automated workflow.

Option D defines observable and repeatable escalation triggers. Confidence thresholds identify cases where the system lacks sufficient certainty. High-impact categories ensure that legally, financially, medically, or operationally consequential decisions receive human attention regardless of nominal confidence. Classifier-generated ambiguity flags add an independent control, while an explicit user request preserves the individual’s ability to obtain review. These triggers should be documented, tested, logged, and mapped to reviewer authority and response-time objectives. Options A and C make escalation discretionary and inconsistent. Option B is purely reactive because the potentially harmful automated action has already occurred. A sound workflow escalates before consequential execution and records the reason, evidence, reviewer, decision, and outcome.

Study Guide references/topics: Human-in-the-loop governance; escalation thresholds; ambiguity detection; consequential decisions; review auditability.