Alibaba is one of the world’s leading sources of large language models, yet its cloudy arm is proudly showing off its own technology that finds ways to avoid using them.

Alibaba Cloud has built a system to analyze its customers’ tech support queries that often avoids using large language models to solve the problem, an approach it says has proven faster, cheaper, and more accurate.

China’s top cloud provider shared its work at this week’s SIGKDD 2026, the annual conference staged by the Association for Computing Machinery's Special Interest Group (SIG) on Knowledge Discovery and Data Mining, in a paper titled “DualLane: Fast and Reliable LLM Agents for Interactive AIOps via Dual-Path Planning.”

The paper explains that Alibaba tries not to have its human staff handle all support tickets, because people are slow and expensive. But when the Chinese cloud champ tried using AI agents to handle many tickets, it found they can respond inappropriately.

“Incorrect agent responses arise from several sources,” the paper states. “First, the agent may make errors in tool selection, failing to select certain critical tools; second, the parameters generated by the agent for tool calls may contain issues such as missing parameters, incorrect parameter names, or wrong parameter types.”

It gets worse, as the third class of error sees agents stumble when asked to handle complex tasks with dependencies. “The agent may fail to accurately extract required parameters from prior-step results, leading to dependency errors,” the paper reveals.

“Finally, when synthesizing tool outputs into the final response, the agent may misinterpret the results or omit certain key information, resulting in an incorrect final answer,” wrote the paper’s eleven authors, all of whom work for Alibaba Cloud.

The outfit’s answer was to use AI to classify incoming tickets into high-frequency routine scenarios and low-frequency long-tail scenarios.

A tool called “DualLane” processes all incoming tickets concurrently on a fast path and a slow path.

If the fast path detects a high-frequency routine scenario, it kills the slow path. The fast path needs only a couple of tokens to make its assessment, but the slow path needs up to 3,000 tokens.

Alibaba Cloud isn’t concerned about token consumption, because 3,000 tokens cost it about $0.001 to process. It does care about resolving tickets and says DualLane makes that happen by referring fast path inquiries to “templates” that explain how customers can resolve each scenario.

DualLane also identifies issues that consistently appear in the slow lane, and when it identifies a “cluster” of similar issues suggests it might be worth elevating that problem to the fast lane. After manual review, Alibaba may choose to develop a new fast lane template that its systems recommend to customers.

The paper presents findings that DualLane outperforms other tools that schedule work for processing by LLMs, including LLMCompiler and React, and reveals that the Alibaba business unit has put DualLane into production.

“Offline benchmarks indicate a high accuracy rate of 96.5%, accompanied by superior latency performance.”

In all, the move is actually a very Alibaba Cloud thing to do: The Register has often observed the outfit sharing info about how it maximizes efficiency by addressing issues such as memory oversubscription, network failover, low GPU utilization rates, and poor observability of virtual networks. ®