Alibaba Cloud Builds AI System to Route Support Tickets Away From AI

Alibaba Cloud, one of the world's largest distributors of large language models, published a paper at SIGKDD 2026 describing a production system whose main job is to keep support tickets away from LLMs. The system, called DualLane, classifies incoming tickets on two concurrent paths and, when the fast path recognizes a routine query, kills the slow path before the LLM ever fires. The fast path costs a couple of tokens; the slow path costs up to 3,000, or about $0.001 each.

The motivation was straightforward: human agents are slow and expensive, but LLM agents kept making things worse. The paper catalogs four failure modes — wrong tool selection, bad parameters, dependency errors, and misinterpreted outputs — that caused agents to respond incorrectly. DualLane sidesteps most of these by routing high-frequency tickets to hand-maintained templates instead. Offline benchmarks put accuracy at 96.5%, which Alibaba says beats LLMCompiler and ReAct. The company has deployed it to production.