How does NLU work?
In a traditional conversational system, designers list the intents to recognise, such as “reset password” or “change plan”, with example training phrases for each. A model learns to classify new messages into those intents and to extract entities from them. Google’s Dialogflow documentation describes an intent as categorising an end user’s intention for one conversation turn, and calls the matching step intent classification.
In a voice system, NLU sits after speech recognition: speech-to-text produces the words, NLU interprets them and the dialogue logic decides what happens next. Google’s machine learning glossary defines NLU as the subset of natural language processing that determines the intentions of something said or typed.
Large language models changed the approach. An LLM can interpret phrasing it was never shown, follow context across turns and reply directly. Vendors have moved with it: Microsoft’s documentation, checked on 8 October 2026, gives 31 March 2026 as the full retirement date of its LUIS service, and says LUIS’s successor, conversational language understanding, retires from Azure Language on 31 March 2029, with new projects directed to Microsoft Foundry.
Take “I can’t get into my account since I changed my email.” An intent model might label it a login problem, with the email change as a detail. An LLM-based agent can respond to the whole sentence without that intermediate label.
Why does NLU matter for a voice agent on a website?
Spoken questions are looser than typed ones. People restart sentences, add filler words and describe a problem before naming it. A voice agent has to work out what they mean from imperfect transcription and incomplete phrasing, and ask a clarifying question when it can’t.
With intent-based NLU, anything outside the intent list falls through to a fallback reply. With an LLM, the agent can respond to unexpected phrasing, but it can also misread a request with confidence, so test with vague and oddly worded questions as well as tidy ones.
How does Kawl Now understand callers?
Kawl Now has no intent model to build or train. The voice model interprets what callers say directly, and you shape its behaviour with instructions and approved knowledge rather than intents, entities and training phrases.
Before the call, the caller types a short topic, which the agent receives as context, labelled as untrusted. Agents greet callers in English, and there is no language setting.
Live captions and saved transcripts show how the caller’s words were transcribed, which helps you tell a misunderstanding from a wrong answer when you test.
Common questions
What is the difference between NLU and NLP?
Natural language processing (NLP) is the whole field of computers working with human language, including translation, summarisation and speech. NLU is the part concerned with meaning: working out what a person intends from what they said or typed.
Do large language models replace NLU?
In many conversational products, the LLM now does the understanding as part of writing its reply. Intent classification is still useful where each message has to trigger one of a fixed set of actions, such as a menu or a booking flow.