Intent Detection

Intent detection is the natural-language-understanding step that reads a customer's message — "where's my refund?" — and maps it to one of a predefined set of labels, like refund_status, so a bot, router, or workflow knows what to do next. It sits at the front door of almost every support automation: triage, auto-answers, and hand-offs all depend on getting the intent right first.

Classically it is a classifier trained on labeled example phrases for each intent; newer systems use embedding similarity or an LLM prompted with the intent list. You evaluate it with intent-classification accuracy, plus macro-F1 across intents (so a few busy intents don't mask a dozen broken ones) and out-of-scope recall — how reliably it says "I don't know" instead of guessing.

That last part is where marketing and reality diverge. Vendors quote 90%+ on tidy, curated datasets, but on real chatbot traffic the number falls sharply, and every system trades in-scope accuracy against its willingness to fall back to a human.

What it hides: most engines assume one intent per message, when real messages carry several or none; the long tail of rare intents is where accuracy quietly collapses; and taxonomies drift as the product changes, so a model trained last quarter mislabels this quarter's questions. A confident but wrong intent doesn't look like an error — it looks like a customer being routed somewhere useless.