Intent detection is working out what a customer wants from the words they used, so the right workflow can run.
The job
A customer writes “hi is the 4412 one coming today or what”. Behind that sentence is an
order-status request with an order number in it. Intent detection is the step that turns
the sentence into “check order 4412 and report status”, which is something software can
act on.
Older tools did this by matching phrasings to a list of intents you had to define and
feed with examples. Setting up fifty intents, each with twenty sample sentences, was the
bulk of a chatbot project. Anything a customer said that was not close to a sample fell
through to a fallback message.
Why it stopped being the hard part
Language models handle paraphrase, typos, mixed languages and indirect requests without
being given samples. In Singapore and Malaysia that last point matters more than it
sounds. Real customer messages mix English with Malay, Mandarin or Hokkien inside one
sentence, and older intent classifiers broke on exactly that.
So the problem that used to consume setup time is now largely solved. Which has moved
the difficulty, not removed it.
Where the difficulty went
Understanding the request is now cheap. Doing something about it is not. Consider four
messages that are all understood correctly and all go nowhere useful without a connection
to a real system.
| Message | Intent | What it needs |
|---|---|---|
| “is 4412 coming today” | Order status | Read the live order |
| “can I move Friday to Monday” | Reschedule | Write to the calendar |
| “got the blue one or not” | Stock check | Read inventory |
| “how do I pay for the set lunch” | Purchase | Send a payment link |
If you are evaluating tools, this is the thing to probe. Accuracy on understanding is
table stakes now. Ask what happens in the second column.
Two intents worth handling deliberately
“I want a person.” Treat it as its own intent with the highest
priority. It should work first time. See live chat
handover.
Several intents in one message. “Where is my order and can I change the
address” is two requests. A system that answers the first and drops the second produces a
second message and a slightly more annoyed customer.
How to check it on your own data
Take fifty real messages from last week. Do not clean them up. Keep the typos, the
voice-note transcriptions, the mixed languages, the ones with no punctuation. Run them
through whatever you are evaluating. You are looking for two things: did it understand,
and could it finish. Count those separately, because the fixes are different.
Related terms
Read next
Start for free
AeroChat runs customer service and sales workflows you describe in plain English. There is a free plan, so there is nothing to decide today.
Start for free on the Shopify App Store
Not on Shopify? Create an AeroChat account here.