A threshold is a confidence cutoff that determines whether an automated decision proceeds or escalates to a human. It turns uncertain model judgment into an explicit architectural boundary. section 1.2.2 section 4.6
Consider claims triage: a pilot classified medical claims with 94% accuracy and reduced triage time from 40 minutes to 4. However, the remaining 6% outside the approved confidence boundary had unknown consequences. A high-acuity claim misclassified as routine could wait 48–72 hours, creating downstream costs of $8,000–$50,000 per case. section 3.1
Use thresholds as part of a bounded-autonomy design:
- Define the confidence band in which an agent may act and the boundary beyond which it must escalate. section 4.6
- Estimate the consequences of decisions outside the boundary, rather than evaluating accuracy alone. section 3.1
- Enforce the threshold architecturally so escalation still occurs when the agent’s reasoning is wrong. section 1.2.2
- Assign escalation ownership and keep accountability with the named business owner when execution shifts between agents and humans. section 4.6
Claims triage
At 50,000 claims annually, a 2% harmful error rate could produce $8M–$50M in expected downstream failure costs, outweighing the pilot’s projected $1.8M labor savings. The threshold therefore governs both autonomy and the business case. section 3.1



















