Short answer
An intent classifier is the model that reads a message and returns a label for what the sender wants, such as order status or a refund, usually with a confidence score. Routing, automation and reporting all start from its label.- Also called
- Intent model, intent classification model, intent recognizer
- Input
- A customer message: chat, email or ticket text
- Output
- One or more intent labels with confidence scores, or none
- Built with
- Rules, a trained model, or a language model given examples
Intent classification is the task of sorting messages by what the sender wants. The intent classifier is the component that does it: the model sitting at the front of a support system, turning free text into a label the rest of the system can act on.
What a classifier is
A classifier is any model that assigns an input to one of a set of categories. A spam filter is a classifier with two categories. An intent classifier has as many categories as the team has request types, and its input is a customer's message.
For each message it returns a label and a score:
- "Where's my parcel? It said Tuesday." Order status, 0.91.
- "Tracking says delivered but there's nothing here." Missing parcel, 0.86.
- "Can I swap the medium for a large?" Exchange, 0.88.
Types of intent classifier models
There are four common ways to build one, and they differ most in what it takes to change them.
| Type | How it decides | When your intents change |
|---|---|---|
| Keyword rules | Matches words or patterns | Rewrite the rules |
| Trained classifier | Learned from many labelled messages | Relabel data and retrain |
| Embedding similarity | Compares the message with saved examples | Add or move examples |
| Language model | Reads intent names, descriptions and examples | Edit the descriptions and examples |
Trained classifiers are fast and cheap per message but need a large labelled set for every intent. Language-model classifiers cost more per message and can be changed in minutes, which matters when a store adds a product line or a new policy.
Intent classifier examples
One morning in a Shopify store's inbox shows what the model has to handle.
| Message | Label | Note |
|---|---|---|
| "I put the wrong address, can you fix it before it ships?" | Address change | Urgent until the order is fulfilled |
| "Cancel order 4831, and do you ship to Canada?" | Cancellation, Shipping policy | Two requests, two labels |
| "Is the linen shirt see-through?" | Product question | Answered from the product page |
| "Hi, following up on my email from Monday" | None | No request to classify yet |
The last row matters as much as the others. A classifier that must always pick a label will force that message into the nearest intent.
In Aide, a message that fits none of the team's topics gets a suggested topic right on the message, with a name, a group and a one-line description. The team builds its topics from the messages that actually arrive instead of writing the list up front.
The classifier in an AI agent
In an AI agent, the classifier is usually the first step. Its label decides which procedure the agent follows, which data it pulls, whether it may act or only draft, and which team gets the conversation if it hands off. Some systems call this a router or an intent classifier agent: a small model whose only job is to send each message down the right path before a larger model writes anything.
In Aide, the topic label decides what happens next: each topic runs at AI Agent, Drafts or None, and a procedure can require a minimum topic confidence before it runs. In Zendesk, Front or Gorgias, the topic is also written back to the ticket as a tag or a ticket field.
What makes a good intent classifier
A good classifier is judged on the messages it gets wrong, not the ones it gets right.
- It can return none when nothing fits, instead of a weak match.
- It returns more than one label when a message holds more than one request.
- It sorts into the team's own intents, not a generic list.
- Its scores are compared with real outcomes, and an intent confidence threshold decides when a label is trusted.
- Its labels are visible on each message, so the team can correct them.
The practice this page rejects is judging a classifier by how rarely it leaves a message unlabelled. A classifier that always answers looks complete, and is wrong on every message that fits nowhere.
Where Aide fits
Aide's intent classifier is a large language model that scores every customer message against the team's own topics, using each topic's name, its description and the real messages the team saves to it as examples. It picks up to three topics, best first, or none when nothing fits well. Aide does not train a separate model for each account; it learns from the examples the team adds, and moving an example to the right topic is how a team corrects it. See what your topics tell you.