Intent Formulation

People say the same thing in many different ways. “How warm will it be tomorrow?”, “Will I need a jacket tomorrow?”, and “Weather Berlin tomorrow” essentially mean the same thing at their core. A computer program has to recognize from these different sentences what the person actually wants to achieve. This goal is exactly what is called the intent. Intent formulation is the step in which this intent is given a fixed name and a fixed structure, for example “check weather” along with the additional details of location and time. Only after this does the program know which action it should carry out.

Why assistants fail without a clear intent

A system that determines the intent incorrectly then does everything else wrong as well. Someone who says “Cancel my order” and instead receives a shipment tracking update is rightfully annoyed. The intent is the switch point at the beginning of the track: if it’s set wrong, even the best rest of the system won’t help anymore.

For companies, this is directly tied to money. A customer service chatbot is supposed to answer simple inquiries itself and only forward difficult ones to humans. How well this succeeds is measured by how reliably intents are recognized. Error rates of just a few percent may sound harmless, but with a million inquiries per month, they mean tens of thousands of mishandled cases.

A second reason is overview. Anyone who has cleanly named all of their users' intents can also see what these users want most often. Such lists are often more valuable to product teams than any survey, because they reflect actual behavior.

From example collection to recognized intent

Classically, developers define a list of possible intents. For each one, they collect example sentences, often several dozen. A model learns from these to match new sentences to the appropriate intent. In addition, it marks the important individual details within the sentence, such as location, date, or order number. These details are called slots and fill the intent with concrete values.

A catch-all for anything unknown is important. If someone asks about something that isn’t provided for at all, the system should openly say so. The most common error in practice is that a request is assigned to a wrong intent with high confidence instead of being treated as unclear. Good systems ask follow-up questions in such cases.

With large language models, the approach has shifted. These models can classify intents even without their own collection of examples, if you simply describe the possible categories to them. However, intent formulation does not disappear entirely as a result. Because which intents exist at all and which action each one triggers still has to be determined by a human.

Voice assistants, hotlines, and search bars

The principle is most clearly encountered in voice assistants like Alexa or Siri. Developers who offer their own functions there define exactly these kinds of intents along with example sentences. The phone menu that asks “What is this about?” instead of requiring button presses also works according to this pattern.

Search engines distinguish intents in broader categories. Does someone just want to read up on something, navigate to a specific page, or buy something? This classification determines whether an answer box, a link, or a product list appears at the top. In marketing, this is therefore referred to as search intent.

In tech news, the term often comes up in connection with AI agents, i.e. programs that carry out multiple steps themselves. There too, the question at the very beginning is what goal the user actually has. This should not be confused with prompt engineering: that is about the skillful formulation of the input to a model, whereas this is about the clean description of the underlying goal.