Task-based Qwen routing cuts ticketing costs and latency
掘金·人工智能 reports that one company routed its ticketing-system requests across qwen-turbo, qwen-plus, and qwen-max according to task complexity. The author says monthly token spending fell from 4,180 yuan to 1,590 yuan, while classification accuracy rose from 91.2% to 93.7%. After concurrency was reduced from 32 to 12 and requests were queued by task, end-to-end P95 latency reportedly dropped from 6.8 seconds to 2.1 seconds.
This single-company account offers concrete evidence that task classification, model routing, summary compression, and workload-aware queues can materially reduce both inference cost and latency, though the figures are self-reported.
我们公司的工单系统跑了六年，去年年底加了一个 AI 助手：客服把用户的话粘进去，自动判断该派给哪个组、紧急程度几级、要不要拉人。
一开始很朴素——所有请求都丢给同一个模型。跑了两周，账单下来了，我盯着那个数字看了很久，然后决定重做。
不是模型不好。是我们把一个需要"想得很深"的活和一个只需要"归个类"的活，交给了同一个脑子。
一、先看一个很土但很有用的分法
工单系统的 AI 请求，拆开看其实只有三类：