What is intent detection?
Intent detection is the determination of what a caller wants, as distinct from the words they chose to ask for it. The same intent arrives as a direct request from one caller and as a long story from another. It is the input to every routing decision the call then makes.
Also called: Intent classification, Intent recognition
Why callers describe a situation instead of naming a request
A caller rarely opens with the intent. They open with context, often with the thing that annoyed them, and the request appears in the third sentence or never appears at all. Menu systems addressed this by refusing to listen, which is precisely what callers dislike about them. An agent detecting intent properly has to hold several candidates open while the caller talks, and has to ask one clarifying question rather than guessing, because a wrong intent commits the call to a path the caller cannot reverse.
- The intent is frequently absent from a caller's opening sentences.
- A wrong intent commits the call to a path the caller cannot easily leave.
- Confidence matters as much as accuracy, since a doubtful intent should prompt a question.
- New intents appear with every campaign, so the set is never settled.
How intent detection works
- The intent set is defined from real recordings rather than from a product catalogue.
- The transcript is classified against that set with a confidence value returned.
- Entities such as amounts and dates are extracted alongside the intent.
- Low confidence produces a clarifying question instead of a routing decision.
- Unrecognised utterances are logged as candidates for a new intent.
- Intent accuracy is monitored per campaign, since the inbound mix keeps changing.
How intent detection is evaluated
Intent detection is assessed on the utterances that are genuinely ambiguous rather than on the clear majority. Measure the following on sampled live calls:
- intent accuracy against a sample labelled by a person
- accuracy on utterances containing no explicit request
- confusion between the closest pair of intents
- share of calls falling outside the intent set
- clarifying question rate and its effect on completion
Limitations, risks and common failure modes
- Intent sets drawn from product structure fail to match how callers describe things.
- A confident wrong intent is more damaging than an admitted uncertainty.
- Recognition errors on a single word reverse the detected intent.
- Rare intents carry the highest consequence and the least training data.