Design effective human oversight into AI systems with confidence-based routing and escalation thresholds.