Working out what the customer wants before anyone has read the message.
Intent detection. Intent detection is the automatic classification of what a customer is trying to accomplish, based on the words in a message — a refund request, a bug report, a cancellation. Intent detection lets support systems route, prioritize and answer conversations before a person has read them.
The team defines a taxonomy of intents — the finite list of things customers come to do — and a classifier assigns each incoming message to one, with a confidence score attached. Older systems did this with keyword rules or small trained models, which needed hundreds of labelled examples per intent and broke on paraphrase; modern systems use large language models that classify from the intent descriptions alone and cope with “my money hasn’t come back yet” meaning refund. Downstream, the label drives everything: routing to the right team, priority, which automation is allowed to answer.
The distinction practitioners learn the hard way: intent is not topic. “Refund” the topic contains “where is my refund?” (a status question), “how do refunds work?” (a documentation question) and “I want a refund” (a churn risk) — three different intents needing three different responses. A taxonomy built on topics routes all three identically, and identically wrong.
Every downstream decision inherits the classification. A misrouted conversation bounces between teams, and each bounce adds a full queue-wait to the customer’s clock — routing error is the quietest way to double handle time. Automation raises the stakes further: an auto-answer fired on the wrong intent is a confident reply to a question the customer did not ask, which is worse than silence.
And taxonomies drift. Products change, so the list of things customers come to do changes — but the taxonomy only changes if someone tends it. An untended taxonomy does not fail loudly; it just misroutes a growing sliver of traffic while every dashboard built on its labels stays green.
Build the taxonomy from evidence: pull a few hundred real conversations and name what customers were actually trying to do. Eight to fifteen intents is the workable range to start — fewer and the labels mean nothing, more and no two people apply them consistently. Resist mirroring your org chart; customers do not know your departments, and an intent named after one routes by your structure rather than their need.
Then treat the confidence score as the boundary between machine and human: high-confidence classifications route or answer automatically, low-confidence ones go to a person, and where you set that threshold is a policy decision about error tolerance, not a technical detail. Measure precision per intent — a 95% average can hide one intent that is wrong half the time — and re-sample quarterly, because the drift never announces itself.
Intent detection is the automated half of an old discipline — these entries cover the manual version and the systems that act on the label.
Give your customers faster answers and your team their evenings back.
Desktop and mobile. Free for solo.