Intent classification is the layer that turns a free-form utterance into a structured label, 'book_appointment,' 'request_refund,' 'check_status.' In the pre-LLM era, intent classification was its own machine learning pipeline with labeled training data and a confidence threshold. In 2026 it is usually a single LLM prompt with a constrained output schema.
The simplification is real but the quality bar has not moved. Misclassifying intent is the most common reason an agent goes off-rails: a 'reschedule' intent treated as 'cancel,' a 'pricing question' treated as 'booking request.' The downstream tool calls then go to the wrong place and the call falls apart.
Production intent classification has two non-obvious requirements. It needs to handle multi-intent utterances ('I need to cancel my Tuesday and book a new one Thursday'), and it needs to refresh its decision as the call evolves. A caller's intent at minute one is often not their intent at minute three.
Vorel intent classification runs every turn, not only on the first utterance. If the caller shifts mid-call from booking to a complaint, the agent updates its action plan accordingly.