New Policy Optimizes Automation in Human-AI Service Systems

Hybrid human-AI service systems, where tasks are first handled by automation (e.g., a chatbot) and then escalated to human agents if unresolved, face a critical challenge: determining the optimal level of automation. Over-reliance on automation can lead to high chatbot costs, while insufficient automation can overwhelm human staff. This paper addresses this trade-off by studying a system where tasks of various types arrive sequentially, with unknown chatbot success probabilities and human service rates. The researchers introduce the UCB-DPP policy, which combines Upper Confidence Bounds (UCB) for learning unknown system parameters with Drift-Plus-Penalty (DPP) control for making queue-aware decisions. This policy dynamically learns when to automate tasks, allocating resources to the chatbot while considering the impact on human service queues. UCB-DPP is proven to achieve a regret of $\widetilde{\mathcal{O}}(K\sqrt{T})$ and ensures the mean-rate stability of human-service queues, meaning it effectively manages workload and prevents bottlenecks. Simulations confirm that this proposed policy significantly outperforms natural baselines, offering a robust solution for optimizing resource allocation in complex human-AI collaboration environments.