How AI intent detection works
Intent detection answers one question about a message: what does this person want? A customer who writes "I was charged twice" wants a refund, "where is my parcel" is asking about an order, and "how much is the pro plan" is asking about pricing. Detecting that automatically is what lets a chatbot, helpdesk or inbox send each message to the right place.
From text to an intent label
You paste a message and the tool sends it to a language model along with a list of possible intents. On this page the list is fixed: question, complaint, refund request, order status, cancellation, pricing inquiry, technical support, purchase, feedback and greeting. You can also type your own list of 2 to 10 short labels to see how the model classifies messages for your business.
The model compares the meaning of the message with the labels, not just keywords. "I want my money back" contains no word "refund" but still maps to a refund request.
What the confidence score means
Alongside the intent you get a confidence score. A high score means the message clearly fits one label. A low score means the message is vague, very short or fits more than one intent, and a person or a follow-up question may be needed. In your own automation, a common pattern is to act automatically above a threshold and hand everything else to a human.
One intent or several
Real messages are often mixed: "My order is late and I want a refund." Turn on the multiple intents option to get every intent found, with its own score, instead of only the main one.
Where intent detection is used
Support teams use it to route tickets, chatbots use it to decide which flow to start, sales teams use it to spot purchase and pricing questions, and product teams use it to count what customers complain about most. The same engine is available as the intent detection API, with a free monthly quota on every plan.
Limits
Intent models can be wrong on sarcasm, slang and very short messages like "ok" or "help". Check the confidence score and test your own labels with real messages before you rely on the result.