Alibaba Cloud has developed a system designed to handle technical support inquiries more quickly by deploying large language models only when absolutely necessary. For common issues, the company uses pre-validated solutions.
The approach is called DualLane and is described in the paper “DualLane: Fast and reliable LLM agents for interactive AIOps via dual-path planning.” Alibaba Cloud presented the research at SIGKDD 2026. The research was prompted by issues with AI agents that handle support tickets autonomously. Such agents can select the wrong tools, use incorrect parameters, or, for more complex tasks, misprocess information from previous steps. Errors can also occur when formulating the final response, reports The Register.
Fast and slow paths
DualLane therefore processes every support query simultaneously via two paths. The fast path attempts to determine whether the issue is a known, common problem. If so, the system halts the slower processing, and the customer receives a solution based on a pre-made template.
The fast classification costs only a few tokens. For less common issues, the second route remains active, allowing an LLM agent to use up to approximately 3,000 tokens to arrive at a solution. The cost to Alibaba is negligible: the company estimates it at about $0.001 per ticket. The main benefit lies in speed and reliability.
The system can also detect patterns in questions that are initially processed via the slow route. When certain issues recur frequently, employees can create a new template after verification and have it handled via the fast route going forward.
Now in production
Alibaba Cloud is now using DualLane in its production environment. According to the researchers, in offline tests the system achieved an accuracy of 96.5 percent and lower latency than alternative approaches such as LLMCompiler and ReAct.
This means Alibaba is using AI not only to automate support tasks but also to determine when it is more efficient to keep a large language model out of the process altogether.