What is Intent classification?

Pre-LLM conversational AI systems (Dialogflow, Rasa, Microsoft LUIS) center on the intent classifier. Developers define a list of intents, write 10-50 example utterances per intent, train a classifier, and at runtime the user’s message is routed to the matching intent and a corresponding handler runs. The pattern is reliable, deterministic, and limited — the system can only do what its intents allow, and any utterance that doesn’t map to a known intent is a failure.

Modern LLM agents don’t require an explicit intent classifier. The language model reads the conversation, considers the available tools, and decides what to do — which is functionally intent classification, but produced through the LLM’s reasoning rather than a dedicated classifier. This is more flexible (the model can handle utterances that don’t map cleanly to predefined intents) but less predictable (the model occasionally chooses the wrong action).

In practice, production AI voice agents often use a hybrid approach: explicit "always handle this way" rules for the highest-stakes intents (DV disclosures, custody transfers, in-custody arrests), and LLM-driven routing for everything else.

For a law-firm voice agent, the intent space is well-defined: new client matter, existing client follow-up, billing question, attorney callback request, urgent (time-sensitive) matter. The first turn typically classifies the intent; the rest of the call is qualification within the intent. A pure LLM agent without explicit intent handling can drift between intents in confusing ways. Adding an explicit intent-classification step at turn 1, then letting the LLM handle qualification within the chosen intent, produces more predictable outcomes.