Skip to content
AI

Intent recognition

Intent recognition is a method that works out what a customer message is asking for, such as an order query, a return or a product question, and routes it to the right answer or person.

Intent recognition means working out the request behind a sentence a customer wrote. “Where is my order?”, “My parcel has not arrived” and “When will #1042 arrive?” are three different sentences that carry the same request: finding out the order status. If a system can put them in the same box, you do not need a separate rule for every sentence. The customer writes in their own words, and the system finds which job the message belongs to.

What happens after the intent is found is a separate decision. An order query means looking at data, a return request means passing it to a team member, a product question means finding an answer in a catalogue. Intent recognition is the first step of that choice. It does not answer by itself; it says which route the answer takes. When it gets it wrong, the customer goes to the wrong door, so leaving an unsure message to a person is good design.

Why It Matters

  • The customer writes in their own words, so you do not need a rule for every variation.
  • When messages reach the right topic and the right person, the customer does not explain the same thing twice.
  • Counting which intents come up often shows what your store gets the most questions about.

How It Works

  1. A message arrivesThe customer types a plain sentence: “The shirt I bought came out too big.”
  2. The system classifies itIt decides which request the message belongs to: a return or exchange request, an order query, a product question.
  3. A route is chosen by intentAn order query goes to data, a product question to information, a return request to a team member.
  4. An unsure message goes to a personIf the system cannot find the intent clearly enough, it does not guess; the conversation is left to a person.

Example

In one store the incoming messages are sorted into three intents: order status, return or exchange, and product information. Mert K. writes “When will #1042 arrive?” and the message goes to order status. Elif Arslan says “The Linen shirt (₺1,490) came out small”; that is classed as a return or exchange and reaches Selin Kaya. Another customer writes only “hello”, so the intent is not clear, and the message is first met with a question. At the end of the week the store owner looks at which intent came up more often.

Intent recognition and keyword matching
Keyword matchingIntent recognition
What it looks atWhether a certain word is in the messageWhat the message is asking for
“My parcel has not arrived”Misses it if the word “order” is absentClasses it as an order query
A new phrasingA new rule has to be writtenCan put a similar phrasing in the same intent
When it is wrongGoes the wrong way on the wrong wordShould leave it to a person if unsure

Not to Be Confused With

Intent recognition is not the same as image understanding; interpreting a photo a customer sends is a different job. The “exit intent” on Şans Çarkı (the spin wheel) is also unrelated: it is a condition that opens the wheel when a visitor is about to leave the page, and it does not work out what a message is for.

Frequently Asked Questions

Is intent recognition the same as a chatbot?

No. Intent recognition is a method a chatbot can use; it is not the bot itself. A chatbot takes a message and replies. Intent recognition only says which request the message belongs to. Two tools using the same method can give very different results depending on where they find the answer.

What happens if intent recognition gets it wrong?

The customer goes down the wrong route: someone asking for a return sees an order status. The way to reduce this is to leave unsure messages to a person and to read misclassified examples regularly. No method is right every time; the chance of error is not zero. That is why a handover rule matters.

How much data does intent recognition need?

It depends on the method. A simple classifier needs sample messages for each intent; a method that uses a language model can work with fewer. What matters is that the examples come from real customer sentences and the list of intents stays short. Start with the requests your store gets most.

Related Terms