Skip to content

Glossary · I

Intent recognition

A woman with headphones sits thoughtfully at a laptop

Definition

Intent recognition is the automatic identification of the request (intent) behind a spoken or written utterance, usually together with the details it needs, such as numbers, dates or amounts.

"I just wanted to ask why my bill is higher this time." A person knows at once what this is about: a billing query. Intent recognition teaches software to understand the same.

#Intent and entity

An intent is the request, the thing a person wants to achieve. Entities are the details needed to resolve it. In the example above, the intent is "billing query"; the entities would be customer number, billing month and amount.

Modern methods use language models and recognise requests even in unfamiliar wording, with typos or in colloquial speech. If a detail is missing, the system asks for it specifically instead of restarting the whole dialogue.

Good recognition also notices several requests in one message. "I've moved and would like to change my tariff while I'm at it" contains two. Both are captured and handled in turn.

#Where it is used

  • In self-service. IVR systems and voicebots recognise the request from free speech.
  • In routing. The ACD routes a contact to the right skill based on the request.
  • In email and messaging. Incoming items are sorted and prioritised by request.
  • In analytics. Speech analytics shows which requests are on the rise.
An utterance becomes a case: request recognised, matching queue chosen, context on the desktop. Sample.

#How a good intent catalogue comes about

Quality depends less on the method than on the catalogue of requests. It should be derived from real conversations, not from the org chart. Customers think in requests, not departments.

Checkpoints for the catalogue

Derive requests from real conversations and messages.
Start with a manageable number of frequent requests.
Draw clear lines between similar requests.
Accept "unclear" as a valid result, with a follow-up question or handover.
Add new wording from operations on a regular basis.

#Intent recognition in Univerx

In Univerx, AI agents, voicebots and call distribution share the same requests. You maintain and test the catalogue in Agent Studio, without programming. Whatever cannot be assigned goes to a person marked "unclear" and feeds into the next revision of the catalogue.

See Univerx on your own example.

In a demo we show the platform along your channels and requests, from AI agents to video, workforce scheduling and analytics.

Online, in German or English, with no preparation on your side.