What is intent classification?
Intent classification is the task of assigning a customer's message or spoken request to one of a predefined set of intent categories, such as check balance or report fraud, usually with a confidence score, so that a system can route the request, answer it or start the right process. It is a core step in natural language understanding.
The output is simple: a label and a number saying how sure the system is. What happens next depends on that number. Above a set threshold, the system acts on the label. In a middle band, it confirms with the customer before acting. Below it, the request goes to a fallback, which is usually a clarifying question or a person.
How intent classification works
There are three common ways to build a classifier, and many production systems combine them.
| Approach | How it works | Trade-off |
|---|---|---|
| Trained classifier | A model learns from labeled example messages for each intent | Fast and inexpensive at high volume; needs new examples and retraining whenever intents change |
| Similarity matching | Messages and intent examples are converted to embeddings and matched by closeness | Needs few examples; struggles to separate intents that sound alike |
| Language-model classification | A language model is given intent names and descriptions and asked to choose | Handles unfamiliar phrasing well; slower, costlier and less predictable without constraints |
Whatever the method, the hardest part is not choosing among known intents but recognizing a request that fits none of them. A classifier forced to pick from a fixed list will often pick something, confidently, when the correct answer is that the request is new. Systems therefore add an explicit out-of-scope category and test it as carefully as the intents themselves.
Why intent classification matters
Researchers who built a benchmark of 150 intents across 10 domains found that classifiers handled requests within their scope well but struggled with those outside it. A widely used language model reached 96.9% accuracy on in-scope queries yet correctly flagged only 40.3% of out-of-scope ones, and the best system tested reached 66% even after training on additional out-of-scope examples.
The requests most likely to fall outside a predefined list are often the ones that matter most: a complaint phrased as a question, a sign of financial hardship, a dispute raised in passing. When those are forced into the nearest routine category, the customer gets a routine answer to a non-routine problem. That is why the out-of-scope path, and its handover to a person, deserves as much design attention as the intents the system was built for.
Where intent classification is used
- Call and chat routing. Deciding which queue, team or automated flow receives a request.
- Email and ticket triage. Sorting incoming messages by topic and urgency.
- Speech-enabled phone menus. Turning what a caller says into a menu choice.
- Compliance monitoring. Flagging complaints, disputes or hardship mentions in recorded conversations.
- Decision points inside agents. Recognizing what the customer has just asked for so the agent can choose its next step.
Classification labels a single message. On its own it misses context from earlier in the conversation and does not decide what to do about the request. Systems that need both rely on a broader understanding layer and on separate decision logic.
What good intent classification looks like
- Training data from real traffic. Examples come from actual customer messages, including typos and transcription errors.
- An out-of-scope class. Requests that match nothing are recognized as such.
- Calibrated confidence. A score of 0.9 really means about nine in ten are right, so thresholds can be trusted.
- Confirmation for consequential intents. Cancellations, disputes and payments are confirmed before action.
- Drift monitoring. Accuracy is tracked as customer language and products change.
- Human review of low-confidence cases. Uncertain messages feed back into new training examples.
How to measure intent classification
- Precision and recall per intent. How often a label is right, and how often each intent is caught.
- Confusion pairs. Which intents are most often mistaken for each other.
- Out-of-scope recall. The share of unsupported requests correctly identified as unsupported.
- Calibration. How closely confidence scores match observed accuracy.
- Performance by channel and speaker group. Accuracy on voice versus text, and across accents and languages.
When a customer's words carry legal weight
Some customer statements create obligations as soon as they are made, including when they are made by voice. A classifier that misses them does not suspend those obligations. The main US points follow, using payments and credit as the example; telecoms, utilities and retail have cancellation and complaint rules that work the same way. In the United Kingdom, the financial regulator also expects firms to be able to spot signs that a customer is vulnerable.
- Errors reported by phone. Under Regulation E, a consumer can report an error on an electronic fund transfer orally, and the institution's investigation obligations begin from that notice. It may ask for written confirmation within 10 business days, but only if it says so, and gives the address, when the customer calls.
- Stopping recurring payments. Regulation E also lets a consumer stop a preauthorized transfer by notifying the institution orally or in writing at least three business days before it is scheduled. The institution may require written confirmation within 14 days.
- Credit card billing errors. Regulation Z's billing error protections, by contrast, are triggered by a written notice. A system that recognizes a card dispute in a call should direct the customer to that written process rather than treat the call as notice.
Frequently asked questions
What is the difference between intent classification and natural language understanding?
Intent classification is one task within natural language understanding. Understanding also covers extracting details, resolving references to earlier statements and reading context and tone.
How much training data does a classifier need?
For a trained classifier, typically dozens to hundreds of varied examples per intent. Similarity matching and language-model approaches can start with far fewer, though they still need test data to prove accuracy.
Can a language model replace an intent classifier?
Often, for flexibility with unusual phrasing. Many teams keep a trained classifier for high-volume, well-defined intents because it is faster, cheaper and more consistent.
What happens when confidence is low?
A well-designed system asks a clarifying question, offers the most likely options or hands the conversation to a person, rather than acting on a guess.
What is out-of-scope detection?
Recognizing that a request does not belong to any intent the system supports, so it can be handled safely instead of being forced into the closest category.
Related terms
Beyond the label.
Encore AI's agents do more than sort requests: they run the conversation that follows, using a playbook compiled from how your top performers win and approved by your compliance team.
Scale your best to be the best.
Book a Demo