← Back to concept explainers

Concept Fable

Intent recognition

Conclusion first:Intent Classification is the ability to determine which category of intent a user request belongs to, in order to decide the subsequent process, tools, Agent, or response strategy. It can be done by rules, traditional classifiers, or LLMs. In enterprise Agents, intent recognition is often used for customer service triage, workflow selection, permission judgment, and routing, but it needs to handle multiple intents, ambiguous requests, low-confidence follow-up questions, and continuous evaluation.

Intent Classification · Orchestration

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The counter first determines what the customer actually came to do

City hall had only one large counter. Customers started talking as soon as they arrived, and the clerk listened while guessing whether they needed tax, household registration, roads, or a shop permit. Regulars could explain clearly; newcomers often wandered in circles for ages.

The earliest method was to have all customers line up in the same queue. When it was someone’s turn, the teller asked more detailed questions. The line grew longer and longer, and people who really only wanted to fill out one form still had to wait for complex disputes to be handled.

The supervisor posted a keyword list: if the word tax appears, go to taxation; if the word housing appears, go to household registration. A few days later, errors occurred. Someone asked about shop rent tax, and both windows said it was not their responsibility.

Another teller suggested letting customers choose their own window. As a result, many chose wrong, waited half an hour, and were sent back, angrier than before.

Later, a triage clerk was added at the hall entrance. He did not solve problems, only asked a few short questions: Are you here to inquire, modify, complain, apply, or supplement materials? Which object is involved? Is there a deadline?

The triage clerk tags each request with an intent label, then routes it to the corresponding window. Vague requests get follow-up questions first, rather than being forced into a category.

The supervisor also regularly reviews misclassified cases. If many customers use the same kind of vague language, the entry point needs to change; if a new business type appears often, add a new category.

The hall quickly quieted down. Not because every window had become smarter, but because the entrance first judged the request type clearly, so the later process could match. The triage clerk said: before acting, first determine what kind of thing the other party actually wants to accomplish. Sort into the wrong door, and no matter how hard you work afterward, you are taking a detour. Later the supervisor found that entrance triage also exposed new needs. If many people cannot clearly state the same thing, the hall changes how it asks; if new matters appear frequently, it opens a new door for them. It also reduced misunderstandings. When customers hear their purpose repeated back, they often correct it on the spot: I am not complaining, I need to supplement evidence. Once the direction is accurate, there are fewer detours afterward. The clerks also argued less, because the matters each window received were more like what it should handle. The clerks also argued less, because the matters each window received were more like what it should handle, and processing truly became faster.

Reveal

This story is about: intent recognition

Intent Classification is the ability to determine which category of intent a user request belongs to, in order to decide the subsequent process, tools, Agent, or response strategy. It can be done by rules, traditional classifiers, or LLMs. In enterprise Agents, intent recognition is often used for customer service triage, workflow selection, permission judgment, and routing, but it needs to handle multiple intents, ambiguous requests, low-confidence follow-up questions, and continuous evaluation.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • City hall main counter: AI applications with a unified entry point
  • Tax, household registration, road service windows: different workflows, tools or Agents
  • Keyword table: fragile rule classification
  • Guests choose their own window: shifting the classification burden to users
  • Entry triage officer: Intent Classifier
  • Query, modification, complaint, application: intent categories
  • Ask follow-up questions for ambiguous requests first: handling low confidence
  • Final insight: Intent Classification identifies the request type before execution, providing the basis for subsequent routing

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

routerNLUworkflow
京ICP备2026014272号