Intent

An intent is the thing a user is trying to accomplish with a message: book an appointment, check an order, complain, reach a human. In classic chatbot design you defined a list of intents, gave each one a set of example phrasings, and the bot classified every incoming message into one of them before doing anything else.

Why it matters when you are buying or building

Intent-based design is why the previous generation of chatbots was so frustrating. If a user said something outside the training examples, they fell into a fallback bucket and got "I did not understand that." Coverage was a treadmill: every new way a customer phrased something was another example you had to add.

Modern LLM agents mostly do not need this. The model reads the message and decides what to do from the instructions, no intent list required. That is a real improvement, and it is most of why current agents feel less brittle.

Intent has not disappeared, though. It has moved:

• Routing. In a multi-department setup, something still has to decide sales versus support versus billing. Whether you call that a classifier or a routing prompt, it is intent detection.
• Analytics. You cannot improve what you cannot count. Tagging conversations by what the caller wanted is how you find out that 40% of your calls are one question you could answer on the website.
• Deterministic flows. Some paths should not be improvised. Payment capture, identity verification, consent collection. You detect the intent, then hand off to a scripted flow.

What I have seen in production

On the multi-tenant voice agents I run for a telecom client, routing is the intent problem that actually matters. Callers do not say "sales." They say something specific about their situation and expect to land in the right place. Getting that wrong is worse than a bad answer, because the caller now has to explain themselves twice. When a route fires, the transfer itself has to be wired correctly too, and that is a separate failure mode: see call transfer flows.

The analytics use is the one clients underrate. Tagging what people asked for, over a few hundred real conversations, is usually the highest-value output of the first month. On lead-gen builds it is how you find out which qualifying question is losing people. See chatbot analytics.

Practical advice: do not build an intent taxonomy up front. Ship with a well-written prompt, collect real transcripts, and let the actual categories emerge. The list you would have guessed and the list the transcripts give you are rarely the same, and the transcripts are right.