Alibaba Cloud DualLane routes routine support tickets away from full LLM planning

Alibaba Cloud DualLane, a two-path AI-agent system for Elastic Compute Service support tickets, has been deployed in the company’s production environment for more than a year, according to an ACM paper by 11 Alibaba Cloud Computing researchers. The system sends frequent, repeatable requests to pre-validated execution templates instead of generating a full large-language-model plan, while retaining LLM-based planning for lower-frequency cases.

The paper was published in the proceedings of KDD ’26, held in Jeju Island, South Korea, from August 9 to 13. It describes a narrow operations use case, not a reported reduction in Alibaba Cloud’s overall AI training, inference, GPU use, energy consumption or spending.

Alibaba Cloud sells infrastructure and services for foundation-model development and deployment, which it markets as optimized for performance and efficiency. DualLane points to a more specific operational problem: deciding when an agent needs to reason through a new support request and when it can reuse a human-verified workflow.

How does Alibaba Cloud DualLane handle support tickets?

In the workflow described by the researchers, ECS support engineers turn customer requests into technical questions, consult specialized tools and escalate cases that remain unresolved. Alibaba Cloud built an agent intended to plan those tool calls, combine their outputs and produce a response.

DualLane runs a fast and slow route in parallel. The fast route identifies high-frequency routine scenarios using a small number of tokens and uses pre-designed, human-verified execution plans. This bypasses full LLM plan generation, then directs the request to a template for the relevant scenario.

Low-frequency, long-tail or complex requests use the slow route. That route uses an LLM-driven, two-stage process that separates task decomposition from parameter generation. The split is designed around the skewed distribution of support work: repeated problems can be standardized, while uncommon requests need a more flexible planning process.

The paper also says the fast path includes LLM-driven scenario matching. The design therefore does not remove AI from routine cases. It avoids the more expensive full-plan generation step when a validated workflow is available.

What results did Alibaba Cloud report?

Alibaba Cloud’s researchers reported 96.5% accuracy in offline benchmarks, a 4.2-second median plan-execution time in production, and a 7.1% rate of queries incorrectly addressed because of agent-induced errors. They also reported up to 44.6% faster plan-and-execution performance than the ReAct and LLMCompiler baselines evaluated under real-world workloads.

Those are company-authored measurements, rather than independent validation. The supplied evidence does not include ticket volumes, the share routed through either path, absolute latency reductions versus an earlier production system, or quantified token, dollar, GPU or energy savings. Operators assessing such results should distinguish benchmark and workflow-specific measures from broader deployment claims, as outlined in AI model evaluation.

The practical significance is the routing pattern, not a claim that all support automation should use fewer models. For a cloud provider handling recurring operational issues, verified plans can constrain a class of errors involving tool selection, parameters and dependencies. Alibaba Cloud’s published results indicate that it has put that pattern into ECS support; they do not establish performance across its broader cloud portfolio.

This story draws on original reporting from The Register.