Many teams set a fixed confidence cutoff, such as ESCALATION_THRESHOLD = 0.90, to decide when an AI agent should act or hand a task to a human. That approach assumes a single probability threshold applies to every decision, ignoring that the cost of an error varies widely across contexts. The author argues the threshold should be a price, derived from the ratio of the cost of a mistake to the cost of escalation.

The rule follows (1 – p) × cost_of_error < cost_of_escalation, where p is the calibrated probability that the agent is correct. In a refund‑triage example, a routine refund costs £15 if the agent errs, while escalating costs £4. The resulting threshold is 1 – 4/15 ≈ 0.73, so a 90 % confident agent can act. If the same confidence is applied to a potential account takeover, the error cost jumps to £2000; the new threshold is 1 – 4/2000 ≈ 0.998, meaning the agent should always defer.

Accuracy calibration matters because the confidence output is not the same as the agent’s true correctness. Overmeyed agents can mislead the threshold, turning what looks like a 90 % chance into only a 78 % reality. The author recommends logging decisions, bucket‑wise calibration, and mapping with isotonic regression or Platt scaling before applying the cost rule.

Implementing the approach requires defining cost classes—refunds, security incidents, etc.—and gathering real‑world cost figures. Teams should involve finance and support ops to estimate cleanup expenses, then compute thresholds per class. Ongoing recalibration guards against drift-loaded errors and keeps the autonomy decision grounded in actual risk.