What is intent classification?
Intent classification is the AI task of labeling a customer message with the underlying goal, separately from the surface text. “I want my money back”, “refund me”, and “I cannot believe you charged me” all map to the same intent: refund request.
Why classify
Intent labels drive three downstream actions: routing (send refunds to billing department, send technical questions to tier 2), self-service deflection (show the refund article instead of opening a chat), and analytics (count how many billing questions came in this week).
Approaches
Old systems used keyword matching, which failed on paraphrase. Modern systems use embeddings plus a small classifier or directly prompt a large language model. The LLM approach is more accurate but more expensive per inference. message.com runs a fast classifier on the hot path and escalates to a heavier model only for ambiguous cases, so latency stays low without sacrificing accuracy.
Output
An intent classifier returns a label plus a confidence score. Below a confidence threshold, the message falls through to a human rather than getting auto-routed. This is the same guardrail pattern we use for RAG generation.
See it working on message.com.
Unified inbox, grounded AI, flat pricing. 14 days free, no credit card.
Start free trial →