Learn AI /Computer Vision and Language Technology /Intent detection
Intent detection
- Intermediate
- Computer Vision and Language Technology
- Technical readers
- Programme & data teams
Also called: intent classification.
Back to the milestone you came from
In five seconds
What it is
A system that assigns a user message or action to a goal category, such as asking for help, making a booking or reporting a problem.
In plain words
It is a way of sorting messages by what the sender is trying to do, such as ask for help, change a booking or report a problem.
An everyday example
A small clinic receives messages in one shared inbox. A staff member quickly sorts each message into trays labelled “urgent medical advice”, “appointment change”, “billing question” and “general information”. They do not need to know every detail of the sender’s situation to place most messages in the right tray. Once sorted, each tray goes to a different team for the next action.
A human sorter can use broad background knowledge, ask clarifying questions, and notice unusual context. An intent detection model usually relies on patterns from its training examples and may fail on new wording, mixed requests, sarcasm or messages that fit no existing category. The analogy also suggests one person making a careful judgement, while many deployed systems produce a fast statistical prediction.
Remember thisIntent detection assigns a message to the most likely goal category. Useful systems also recognise low-confidence cases and send them for clarification or human review.
Understand it
Understand it
Why it matters
For a non-specialist, intent detection matters because it shapes how organisations respond to people. It can speed up triage, route requests to the right team, and make services easier to use. But it also creates practical risks: if the label set is poorly designed, if training examples are unbalanced, or if confidence is ignored, people can be misrouted or misunderstood. Knowing what the system actually does helps you ask better questions: What are the intent categories? What happens when none fit? How is low confidence handled? Who checks errors? Those questions affect service quality, fairness and accountability more than impressive-sounding claims about AI.
From message to intent label
Show how a short message moves through an intent detection system and becomes a category used for routing or the next step.
01
1. Input message
A person sends a short text such as “Please move my meeting to next week”.
Draw a message bubble on the left with the example text. Add a small caption: “new user message”.
02
2. Compare with learned patterns
The model analyses words and phrases against patterns learned from labelled examples.
Draw a central box labelled “intent model”. Inside, show small token-like text fragments flowing into a pattern grid. Add a note: “trained on labelled examples”.
03
3. Score candidate intents
The system gives several possible labels scores, such as “reschedule” 0.78 and “cancel” 0.14.
To the right of the model, show a short ranked list with three intent labels and horizontal score bars of different lengths.
04
4. Choose next action
The top label is used to route the request, ask a clarifying question, or trigger a workflow.
On the far right, show the winning label in a highlighted pill and an arrow to an action box such as “send to scheduling workflow”. Add a warning icon and text: “if confidence is low, ask a follow-up question”.
A left-to-right flow diagram with one example message entering a model, then a ranked list of possible intents, then the chosen label and resulting action. Read it from input to prediction to use. Small notes show that labels are predefined and confidence can be low when wording is unclear.
Hovering or tapping each step reveals a plain-language note with the role of that stage and one possible failure point, such as unclear wording or missing categories.
Describe this without the diagram
The diagram shows a four-step flow from left to right. First, a message bubble contains the text “Please move my meeting to next week”. Second, the message enters an intent model trained on labelled examples, shown as text fragments being compared with learned patterns. Third, the system outputs a ranked list of possible intents with scores: reschedule is highest, cancel is lower, and general query is lower still. Fourth, the highest-scoring label is used for an action, such as sending the request to a scheduling workflow. A note explains that if no label scores highly, the safer design is to ask a clarifying question rather than assume.
With reduced motion: A static four-column diagram shows the same process without animation: input message, model box, ranked intent list with score bars, and chosen action. All arrows are fixed and visible at once.
Intent detection takes a piece of language, and sometimes a user action, and predicts what goal it most likely represents. The input is usually a short message such as “I need to change my flight” or “Can I reset my password?”. The output is a label, such as “booking change” or “account support”.
In practice, the system is given a set of predefined categories. It compares the new message with patterns learned from earlier examples. Those examples are often called training data: past messages that people have already labelled with the correct category.
This does not mean the system fully understands the person. It is performing a classification task: choosing one label from a list, or sometimes several labels if the design allows that. Some systems also return a confidence score, which is a number showing how strongly the model favours one label over others.
Intent detection is often used before another step. For example, the predicted intent may route a message to the right team, trigger a form, or choose which question to ask next. It is commonly confused with sentiment analysis, which looks for emotional tone such as positive or negative, not the user’s practical goal.
How “Sorting messages by what the sender needs” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Shared inbox messages | Incoming user text or chat turns |
| Labelled trays | Predefined intent categories |
| Staff member sorting by clues in the wording | Model predicting the most likely intent from patterns in text |
| Different teams handling each tray | Downstream routing or action based on the predicted intent |
A worked example
A duty officer at a regional hub must sort new service desk messages before circulating the morning handover.
Starting from: Message: “I can sign in, but the portal says I do not have permission to upload the progress report.”
- The classifier converts the message into internal features and compares them with labelled examples for intents such as account access, system outage, upload failure and request for guidance.
- It estimates scores for the available classes: account access 0.72, upload failure 0.19, request for guidance 0.07 and system outage 0.02.
- The account access score exceeds the routing threshold, so the message is assigned to that queue. If no score had met the threshold, the system would have requested clarification or sent the message for manual review.
Result: The message is labelled “account access” with a score of 0.72 and routed accordingly. The original text remains available to the officer, who can correct the label if the apparent permission problem has another cause.
In a United Nations workplace
During the technical evaluation of three bids, an ICT officer in a Resident Coordinator's Office tests each proposed service desk classifier against the same set of labelled sample requests before the evaluation matrix goes to the procurement unit at close of business.
The test shows whether each system can assign new messages to operational goals such as requesting access, reporting an outage or asking for guidance. It should also show when confidence is too low for automatic routing and the message should instead be flagged for clarification or human review.
Before you try thisBid contents and procurement-sensitive correspondence must not be entered into an external tool that has not been approved for that material. Use synthetic or properly anonymised test messages, apply the same test set and thresholds to all bids, and record misclassifications in the evaluation file rather than relying on a polished vendor demonstration.
Where you may hear it
- A vendor demo during a procurement meeting
- A service desk handover at a regional hub
- An ICT governance board reviewing chatbot performance
- A country office meeting on routing requests from a shared inbox
Avoid confusion
Avoid confusion
Often confused with
Sentiment analysis
A method that sorts text, speech transcripts or sometimes images into attitudes such as positive, negative or neutral.
Topic modelling
A method that scans many documents and groups frequently co-occurring words into recurring themes, helping people see what a collection is mostly about.
Named entity recognition
A text-processing method that finds mentions such as people, organisations, places, dates or amounts and labels each mention by type.
Common misconceptions
Common misconception: Intent detection tells you what a person really meant in a deep or complete sense.
Correction: It usually predicts the best-fitting label from a predefined list. That can be useful, but it is narrower than full interpretation. If the list is incomplete or the wording is ambiguous, the prediction may be wrong even when the message makes sense to a human reader.
Common misconception: It is the same as sentiment analysis.
Correction: Sentiment analysis looks for emotional tone, such as positive, negative or angry. Intent detection looks for the practical goal, such as making a complaint, requesting a refund or asking for information. One message can have the same intent with very different sentiment, or the same sentiment with different intents.
Common misconception: A high score means the model is correct.
Correction: A confidence score only shows how strongly the model prefers one label over the alternatives it has been given. If the correct intent is missing from the label set, or the training data were poor, a high score can still point to the wrong category.
Quick check
Which output best matches an intent detection system?
Intent detection outputs the likely goal category of the message. Rewriting text is a generation task, extracting names is named entity recognition, and positive or negative tone is sentiment analysis.
Go deeper
Go deeper
The simple version
It is a way of sorting messages by what the sender is trying to do, such as ask for help, change a booking or report a problem.
The technical version
Intent detection is a supervised classification task in natural language processing. A model is trained on labelled examples where each utterance, query or text span is mapped to one intent class. At inference time, the model converts the new text into internal features or embeddings and estimates scores or probabilities over the available classes. The system then selects the top class, or a set of classes in a multi-label design, often with thresholds for fallback or clarification. In deployed systems, intent detection is commonly combined with entity extraction, dialogue state tracking or workflow routing. It differs from sentiment analysis, topic modelling and named entity recognition: those tasks target emotional tone, broad themes and named items, not the user’s operational goal.
Reviewed external resources
EncyclopaediaWikipedia search: Intent detection (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers see the range of related articles and terms, and suits someone starting with broad orientation rather than a single definitive source.
search results, not one pagelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary provides short definitions of machine learning terms that often appear in explanations of intent detection, and suits readers who need quick clarification while reading.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives a practical introduction to core machine learning ideas behind tasks like intent detection, and suits readers who want structured beginner-friendly technical grounding.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Intent detection (opens in a new tab)YouTube search
This YouTube search results page helps readers find video explanations with different levels of depth, and suits people who learn better by watching than by reading.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Intent detection (opens in a new tab)Reddit search
This Reddit search results page shows how practitioners and learners discuss the term in real-world questions, and suits readers who want to see common confusions and informal explanations.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Intent detection assigns a message to the most likely goal category. Useful systems also recognise low-confidence cases and send them for clarification or human review.