Member-only story
Most of What You’re Calling an Agent Is a Workflow
Autonomous Agents are a cost. Most enterprise use cases are paying it for nothing.
You are not a medium member? This is the friend link to read the full blog …
An automotive manufacturer, service desk side of the house, somewhere north of 40,000 tickets a month coming in from the dealer network. A small team built an agent for it, and that agent did two things. It read an incoming ticket, worked out what kind of problem it was, and routed it to the right queue. Then, for one defined class of ticket, it resolved and closed the thing on its own.
The first half is a classifier and a routing table.
Read the text, assign one of maybe forty categories, look up the queue, hand it over. The step count is fixed. The branches are enumerable. There’s a correct answer for any given ticket that a human can confirm in four seconds. You could build it as a fine-tuned classifier, or as a single constrained model call returning one field from a closed list, and either way you get a deterministic assertion, a confusion matrix, a stable regression suite, and a system whose failure mode is “wrong queue” rather than “closed 600 tickets over a weekend.”
They built it as an agent with tools and autonomy, because the whole thing was one initiative and the second half needed to be…
