Learning When to Automate: Queue Control in Human-AI Service Systems

What the math gives to ML

The paper combines confidence-bound learning for unknown automation quality with Lyapunov drift-plus-penalty control for preventing expensive fallback queues from becoming unstable. The transferable asset is a queue-aware routing rule that trades cheap, uncertain computation against costly expert computation while explicitly controlling backlog. A practical adaptation is to route requests between a small model and a large model, using empirical success estimates for the small model and virtual queues to enforce latency, escalation-rate, or compute-budget constraints.