Glossary
Intent recognition
Intent recognition assigns a freely worded utterance to a request type, for example appointment, complaint or callback.
The term comes from older voicebots. There, a list of intents with example sentences was created in advance, and a classifier assigned every utterance to one of them. If nothing matched, a follow-up question or a transfer came next. Assistants based on large language models no longer need a fixed list in order to reply.
The assignment remains useful all the same: for transferring to the right team, for rules (“always offer a callback for complaints”) and for analysis. Only when requests are named consistently can you see why people call and which questions the assistant could not answer. A short, stable list of categories makes more sense than a fine-grained structure that nobody maintains.
In practice
Define five to ten intents that really occur in your business and give each a clear name. Describe for each, in a sentence or two, how to recognise it and what follows from it: answer directly, take details or transfer. Include a category “other”, so that nothing is forced into a box that does not fit. After a few weeks, look at what has ended up under “other”. If similar calls pile up there, a category is missing. If another one stays almost empty, it can go.
An example
An online shop for bicycle parts gets calls about delivery status, returns, product advice and invoices. A customer says: “I ordered a saddle last week and it doesn’t fit my seat post.” A keyword system would guess delivery status because of “ordered”. The assistant recognises a return, asks for the order number and explains the procedure. In the monthly statistics the retailer sees that many returns are explained by measurements that did not fit, and adds the dimensions to the shop.
A common misunderstanding
Intent recognition is often dismissed as yesterday’s technology because modern assistants answer freely. But understanding and classifying are two different things. Without named intents there is no reliable routing and no statistics. The opposite mistake is too fine a breakdown with forty categories. The more alike two categories are, the more often calls are assigned wrongly, and the less the statistics tell you.
At Call0
At Call0, calls produce summaries that include the request that was recognised.
Read on:
On this topic:
As of: 12 October 2026