What is intent detection?
Intent detection is the AI task of identifying what a customer wants from their message, such as tracking an order, starting a return or canceling. It assigns each message to one of a set of predefined intents, usually with a confidence score, so the request can be routed to the right team, workflow or AI agent.
Also called: intent classification, intent recognition, intent matching
Why intent detection matters
Intent detection is the first step in automated customer service. Routing, the choice of workflow and reporting on contact reasons all depend on it, so an error here carries through to everything that follows. It is also called intent classification or intent recognition.
Most systems set a confidence threshold, and messages below it go to a fallback. Hard cases include messages with more than one request, vague or badly worded messages, and messages in several languages.
Intent detection can use a classifier trained on labeled examples, a large language model, or both. Combining a trained classifier with a language model can be more accurate than either on its own. Common ecommerce intents include order status (WISMO), returns, cancellations, order changes and product questions.
How intent detection works in practice
In DigitalGenius, intent detection follows these steps:
The message is cleaned of email chains, HTML and links, and its language is detected.
The AI predicts a broad category, such as Order Status or Return Questions, then a specific intent within it, such as Where Is My Order or How To Return.
Each prediction has a confidence score. If the AI is confident about the category but not the specific intent, it returns only the category. If it is confident about neither, the message is classed as "other" and handled by a fallback.
For messages with more than one request, a second likely intent can be returned.
The detected intent starts the matching workflow and can be added to the helpdesk ticket as a tag for reporting.
Topics that the trained model does not cover can be detected with keyword rules.
Example
A customer writes: "Where's my package?" The AI detects the category Order Status and the intent Where Is My Order, and the order tracking workflow starts. Another customer writes: "My parcel still hasn't arrived, and I'd like to return the blue one when it does." The AI returns order status as the main intent and a return as the second, so both requests can be handled.
Intent detection and DigitalGenius
DigitalGenius uses AI trained on ecommerce conversations, in up to nine languages, to understand what customers mean in their own words, then routes each request to the right agent or workflow.
Read more
FAQ
Is intent detection the same as intent classification?
Yes. The terms are used interchangeably. Both describe matching a customer's message to one of a set of predefined intents, such as order status or returns.
What happens when the AI isn't sure of the intent?
If the confidence score is below a set threshold, the system uses a fallback. Depending on the setup, it may return only the broad category, ask the customer a clarifying question or pass the message to a human agent.
Can one message have more than one intent?
Yes. Customers often ask two things at once, such as where an order is and how to return part of it. Good systems detect each intent and handle them in turn.
· By Iain Moss