What is intent detection?
- Intent detection
- Intent detection is working out what a caller actually wants from the words they used, so the system can act on the request rather than only transcribing it.
What intent detection means in practice
People do not phrase requests the way a menu does. "My tooth is killing me" and "I need an emergency appointment" are the same intent in completely different words.
Older systems matched keywords and broke constantly, because the keyword you did not think of is the one the caller used.
Language models changed this. The request is understood from meaning rather than from a phrase list, which is why the caller no longer has to guess your vocabulary.
What people get wrong
Eight ways to ask a vet for the same thing
Say you manage a veterinary clinic and you've set up one intent called "book a sick visit". Here's how callers actually ask for it over a single morning. "Can I get Biscuit in today?" "My cat's been throwing up since last night." "Do you have anything this afternoon?" "She's not eating and I'm worried." "I need someone to look at my dog's paw." "Are you taking walk-ins?" "He got into the trash and now he's acting weird." "I think he needs to be seen."
Not one of them contains the word "appointment", and only one contains "visit" in any form. A keyword list would have caught perhaps two. They're the same request eight times, and the third one can't even be understood without knowing it was said to a vet.
The request and the details are separate jobs
People often fold two things into "intent". One is the request itself: book, cancel, ask a price, reach a person. The other is the set of details needed to act on it, which engineers call entities or slots. For a sick visit those are the pet's name, the species, the symptom, and how soon.
The split matters when something goes wrong. Say the assistant offers a grooming slot to the owner of a vomiting cat. There, the request was misread. If it books the sick visit under "Biscuit" for a dog when Biscuit is a cat, the request was right and a detail was wrong. Those need different fixes. A misread request usually means your description of that call type is thin, so add the phrasings you saw. With a wrong detail, the likelier cause is that the question collecting it is unclear, or that the answer never got confirmed back to the caller.
How GreetKeeper handles it
GreetKeeper routes on what the caller describes rather than on which key they pressed, so nobody has to translate their problem into your departments.
Where the intent is one you have flagged as urgent, the routing rule fires immediately rather than at the end of a script.
The transcript shows what was understood next to what was said, which is how you find the phrasings your rules are missing.
Intent questions
What happens with an intent nobody anticipated?
It should hand over rather than improvise. That is the whole value of a hard boundary, and the transcript then tells you which intent to add.
Does it work on rambling callers?
Better than keyword matching did, though a caller who covers four topics in one sentence is still hard. Reading those transcripts is the fastest way to tighten your rules.
Can I see which intents fire most?
Read the transcripts for a week and the pattern is obvious. Three or four requests account for most calls in nearly every small business.
Related terms
Hear it take one of your calls
Two minutes, your own scenario, no card.