Short answer
Intent detection is the task of working out what a person wants from what they wrote or said, such as tracking an order or cancelling it. In customer service it is the first step before any reply, routing or action.- Also called
- Intent recognition, intent identification, intention detection
- Answers
- What does this person want?
- Output
- One or more intents, or none
- Used in
- Customer service, chatbots, voice assistants, search
Customers rarely state their intent in the words a system expects. "Where's my stuff", "it's been ten days" and "did you lose my order" are one request. Intent detection is the job of seeing that, and it is usually done by an intent classifier: the model that assigns the label.
How intent detection works
Every intent detection step runs through the same sequence.
- Read the whole message, including the subject line, earlier turns and any order the customer mentions.
- Compare it with the intents the team has defined, such as order status, cancellation or returns.
- Score each intent by how well it fits.
- Keep every intent that clears the bar, since one message can hold several requests. Multi-intent detection is the name for that case.
- Return none when nothing fits well, rather than the closest guess.
The output is a label, or a few, that the rest of the system uses to decide what happens next.
Intent detection examples
Different words, one intent, and similar words with different intents:
| Message | Detected intent |
|---|---|
| "Where is order 4831?" | Order status |
| "It's been ten days, nothing yet" | Order status |
| "Tracking says delivered but there's nothing here" | Missing parcel |
| "I'm moving on Friday, is it out yet?" | Address change, possibly |
| "Love the jacket, do you have it in olive?" | Product question |
The third row looks like an order status question and is not. A missing parcel needs a claim with the carrier, not a tracking link, which is why detection has to go beyond keywords.
In conversations and calls
Intent can change while a conversation runs. A shopper who opens with "where is my order" may, three messages later, want to cancel it. Conversational intent detection reads each new message in the context of the thread, so the system follows the change instead of finishing the first request.
Call intent detection applies the same task to phone calls: the speech is transcribed, then the transcript is read for intent. The method is the same; the input is noisier.
In Aide, a procedure can be set to switch when the customer moves on to something another procedure fits better, so "where is my order" followed by "actually, cancel it" is picked up as a cancellation request. Each message in the thread names the procedure Aide followed and the ones it considered, with the reason.
Intent tagging and intent gates
Detection is only useful once something acts on it. Two common uses:
- Intent tagging. The detected intent is written onto the ticket as a tag or field, so the queue can be filtered, routed and reported by request type.
- Intent gates. An automation runs only when a specific intent is detected with enough confidence. A cancellation workflow, for example, should not start on a message that merely mentions the word cancel.
With Aide, a cancellation procedure can run only when the cancellation topic is detected, require a minimum topic confidence, and carry a plain-English When to use that says what it is not for, such as "Not when the customer only asks whether they can cancel later".
What intent detection misses
Most detection failures fall into four groups.
- The long tail. Rare requests that match no defined intent.
- New intents. A product launch or a policy change creates requests the list has never seen.
- Hidden second requests. "Thanks for the refund, and can I change my other order?"
- Forced matches. A message filed under the nearest intent because "none" was not an option.
The practice this page rejects is treating detection as permission. Knowing what a customer wants tells a system which path applies; whether AI may take that path is a separate decision for the team.
Where Aide fits
Aide detects intent on every customer message in chat, email and helpdesk tickets, including conversations a teammate is handling, against the team's own topics and in any language. It picks up to three topics or none, and when nothing fits it suggests a new topic on the message, so the list grows from real conversations. Topics are written back to Zendesk, Gorgias or Front as tags or a ticket field, and each topic's AI level, AI Agent, Drafts or None, decides what Aide may do with it. See AI ticket tagging.