Alibaba Cloud is using AI to help it use less AI

Alibaba Cloud, a major player in the tech industry, has devised an innovative approach to optimize its AI usage. The company, known for developing large language models, has created a system called DualLane to analyze customer tech support queries. This system aims to avoid using large language models when possible, resulting in faster, cheaper, and more accurate solutions.

At the recent SIGKDD 2026 conference, Alibaba Cloud unveiled DualLane in a paper titled "DualLane: Fast and Reliable LLM Agents for Interactive AIOps via Dual-Path Planning." The paper details how the system works by classifying incoming tickets into high-frequency routine scenarios and low-frequency long-tail scenarios. For routine issues, DualLane swiftly resolves them using pre-defined templates, while for more complex cases, it employs AI agents.

However, the paper also highlights the challenges Alibaba Cloud faced when initially relying solely on AI agents. Incorrect responses, parameter errors, and dependency issues plagued the system. To address these problems, DualLane takes a dual-path approach, utilizing fast and slow paths. The fast path, requiring only a few tokens, quickly assesses high-frequency scenarios and refers customers to templates for self-resolution. The slow path, capable of handling up to 3,000 tokens, is reserved for complex cases.

Alibaba Cloud's DualLane has shown promising results, with offline benchmarks indicating a 96.5% accuracy rate and superior latency performance compared to other scheduling tools like LLMCompiler and React. The company has already implemented DualLane in production, demonstrating its commitment to maximizing efficiency and addressing issues such as memory oversubscription, network failover, and poor observability of virtual networks.