Intent

People say the same thing in very different ways. “How warm will it be tomorrow?”, “Do I need a jacket tomorrow?”, and “Weather tomorrow” are three different sentences. Each time the same thing is meant: the user wants the weather forecast. This exact meaning is what technology calls the intent, i.e., the purpose behind the utterance. A voice assistant must first determine the intent before it can respond sensibly. Only afterward is it clear which information it should fetch or which action it should trigger.

Why assistants fail without intent recognition

A computer program cannot be prepared individually for every possible sentence. There are infinitely many phrasings, but only a manageable number of things a user actually wants. For a banking app, that might be thirty concerns: check account balance, block card, execute transfer, change address. The intent is the bridge between the open-ended language of humans and the limited list of a program’s functions.

Economically, this is a major factor. Companies automatically route customer inquiries to the right department because a human on the phone is expensive. If the intent is misidentified, the customer ends up in the wrong queue and calls back frustrated. That’s why companies measure very precisely how often the recognition is correct and how often it misses.

It’s important to distinguish this from a related concept: entities. The intent states WHAT is wanted, while entities provide the details for it. In “Book me a flight to Rome on Friday”, the intent is “book flight”. “Rome” and “Friday” are the entities, i.e., the concrete details that fill out the request.

From keyword lists to trained classification

Early systems worked with fixed rules. If the word “weather” appeared, the weather intent was meant. This worked as long as users used exactly the expected words. Sentences like “Do I need to bring an umbrella?” fell through the cracks because no keyword matched.

Today’s systems learn the mapping from examples. Developers collect many real phrasings for each intent, often several dozen per request. A model evaluates these examples and then classifies even unfamiliar sentences. It doesn’t compare words but the meaning of entire sentences. That’s why it recognizes “bring an umbrella” as a weather question, even though the word weather is missing.

The result is usually not a simple yes or no. The system provides a confidence value for each possible intent. If the highest value falls below a predefined threshold, the system asks back: “Did you mean your account balance or a transfer?” Large language models no longer strictly need this predefined list. They derive the intent directly from the text and then select the appropriate tool themselves.

Intents in assistants, hotlines, and search engines

The principle is most clearly seen in voice assistants like Alexa or Siri. Anyone programming their own extension there first defines a list of intents along with matching example sentences. The chat windows on websites of mail-order retailers, insurers, and phone providers work the same way.

Search engines also think in terms of intents. They distinguish whether someone wants to buy something, look something up, or navigate to a specific page. In online marketing, this is called search intent, and entire texts are written with it in mind. It’s the same basic idea as with chatbots, just in a different domain.

In company announcements, the term usually appears alongside figures on recognition rates. A typical misconception here: a high value doesn’t mean the system understands language. It sorts sentences into pigeonholes that humans have set up beforehand. If the matching pigeonhole is missing, it still picks one – and confidently answers off the mark.