Alibaba's DualLane: classify first, reason on the tail

Most AI support automation is built backwards. Every ticket goes to a language model, the model reasons its way to an answer, and you pay in latency and wrong answers for the 80% of tickets that were the same ten questions. Alibaba Cloud built the inverse and put it in production: a support ticket system whose first move is to decide whether it needs an LLM at all.

What actually happened

The system is called DualLane, described in a SIGKDD 2026 paper from eleven Alibaba Cloud researchers, "DualLane: Fast and Reliable LLM Agents for Interactive AIOps via Dual-Path Planning." The Register covered it yesterday.

Every incoming query starts down two paths at once. The fast path spends roughly two tokens classifying whether this is a known, high-frequency problem. If it is, the system answers from a pre-validated template and kills the slow path before it finishes thinking. The slow path is a conventional LLM agent with up to about 3,000 tokens of reasoning, and it only gets to finish on the genuinely unusual tickets.

Reported results: 96.5% accuracy in offline tests, with lower latency than agent-planning baselines like LLMCompiler and ReAct. Alibaba has it running in production. The paper also puts the cost of a 3,000-token pass at roughly $0.001 — which is the number people will quote and the least interesting one here. A tenth of a cent was never the problem. The problem is a customer waiting eleven seconds for a paragraph that gets their refund policy wrong.

Why it matters for your business

The result generalizes past AIOps because the shape of the queue is the same everywhere. Pull your last 500 tickets and sort by intent — where's my order, how do I return this, reset my password, is this in stock. That head of the distribution doesn't need reasoning. It needs correct identification and a good canned answer, which is a classification problem you can measure, version, and fix.

Three things this pattern buys you that "just wrap it in an LLM" doesn't:

Auditability. When a templated answer is wrong, you edit the template and it's wrong zero times after that. When a generated answer is wrong, you edit a prompt and hope. We've written before about deterministic tools beating bigger models; this is the same lesson at the routing layer.

Latency where customers notice it. The common questions are the ones with impatient askers. Answering them in milliseconds while the hard ones take seconds is the right allocation.

A real escalation boundary. A confidence threshold on the classifier gives you one honest place to hand off to a human, instead of an agent that improvises when it should stop.

Running both paths concurrently is the clever bit — you don't pay a routing penalty on the hard tickets, because the slow path started at the same moment. Copy that. Don't gate the LLM behind the classifier; race them and cancel.

Sources: The Register, Techzine.