Behind every message there is a goal. "Are you open on Saturday?", "what are your hours this weekend?" and "can I drop by on Saturday morning?" are three different phrasings of the same intent: finding out Saturday's opening hours. The intent is what the person wants to achieve, whatever words they use.
Intent in classic chatbots
The notion of intent comes from the previous generation of chatbots. They relied on a list of intents defined in advance by the team: "track my order", "change my address", "find the opening hours", "talk to someone". For each one, dozens of example sentences had to be provided, then the matching answer or path written.
A natural language processing tool learned to recognize these intents from the examples. When a visitor wrote, the program looked for the most likely intent and displayed the planned answer. Platforms such as Google's Dialogflow or Rasa popularized this approach.
It has strengths: every answer is controlled and the behaviour is predictable. It also has two limits. A question that matches no planned intent falls into an "I didn't understand" category. And the list keeps growing, with similar intents getting confused.
Intent in an AI assistant
An AI assistant built on a language model does not need a list of intents to answer. It understands the request directly and looks for the answer in the documents. The notion of intent does not disappear, though: it takes on a different role.
- To decide on an action: if the intent is "book an appointment" or "talk to someone", the assistant triggers a handover or a form rather than writing a plain answer
- To hand over: some intents, such as "dispute an invoice" or "file a complaint", are reserved for the team, however they are worded
- To measure: classifying conversations by intent shows what visitors are really looking for, and where documents are missing
Intent, topic and sentiment
Three related notions complement each other when analysing a conversation:
| Notion | Question it answers | Example |
|---|---|---|
| Intent | What does the visitor want to achieve? | Get a quote for a boiler |
| Topic | What is the conversation about? | Heating, servicing, energy grant |
| Sentiment | What frame of mind are they in? | In a hurry, but friendly |
A single intent can involve several topics, and a negative sentiment can signal the need to hand over even when the intent seems simple.
The intents that matter for an SME
For most small organizations, around ten intents cover most messages:
- Getting information: opening hours, address, services, terms
- Getting a price: public rate, quote, estimate
- Booking an appointment or requesting a call-out
- Following up on an existing request
- Reporting a problem or filing a complaint
- Talking to a person
Identifying these intents before going live helps write the assistant's prompt and prepare lead qualification for sales requests. You then know which intents the assistant handles alone and which it passes on.
The pitfalls
- Messages with several intents. "Do you also do window frames, and how much is a quote?" contains a request for information and a request for a price. A good assistant handles both
- The hidden intent. "My system is making a strange noise" could be a request for information, a call-out request or an emergency. A clarifying question is better than an assumption
- The intent that shifts. A visitor who is gathering information may end up asking for a quote. The analysis must look at the whole conversation, not just the first message
Intent in KernChat
KernChat does not require you to define a list of intents: the assistant understands questions as they are asked. When each conversation closes, an automatic summary extracts its main intent, the sentiment and the topics discussed. This information feeds the workspace dashboards and helps the team see what visitors ask most often.