Escalation policy

Confidence threshold for auto-handling: 70%

This case study models a patient-intake chatbot triaging symptom reports by urgency. When its own confidence in a triage decision falls below the threshold, it escalates to a human nurse instead of deciding alone. A lower threshold escalates more often, safer but costlier; a higher threshold automates more, cheaper but riskier for edge cases.

Outcome over 500 simulated intakes

Auto-handled safely—

Escalated to nurse—

Risky auto-handled (missed escalation)—

Intake confidence distribution (red = risky cases below true-safety line)