Intent recognition classifies a caller utterance into a predefined intent — "book appointment", "reschedule appointment", "billing question", "complaint". In modern voice-AI stacks the LLM itself usually does this, often combined with a small hardened classifier layer for high-frequency standard cases.
Solid intent recognition needs a reality-based intent inventory: too many intents (>30) make the model uncertain, too few (<5) leave relevant cases unhandled. A two-stage model usually works best: 6–12 top-level intents plus optional slot fields ("for what date?", "which treatment type?").
Operationally critical: each intent class needs a fallback strategy for low-confidence cases, a measured confusion-matrix report from real calls, and a review cycle in which newly appearing concerns are either added as an intent or escalated to a human.