Key takeaways:
- Intent, entity, and context work together. The intent is the goal, the entity is the detail, and the context is the conversation history. Catch all three and the bot acts on the first try.
- Classification and recognition aren’t the same. Classification sorts messages into predefined intents; recognition is how the bot reads meaning from natural language in the first place.
- The 2026 choice is NLP vs. LLM—and usually both. Traditional NLP is cheap but needs training data; LLM/zero-shot is flexible but needs guardrails. Most bots run a hybrid.
- An LLM agent skips per-intent training. With Tidio’s Lyro, you connect your help content and it recognizes intents out of the box.
A chatbot intent is the goal behind a user’s message—what the person is actually trying to do, like track an order, reset a password, or compare plans. Intent is the foundation of every useful bot reply: read it right and the bot helps; read it wrong and the conversation stalls. This guide covers what intents are, how they differ from entities, how classification and recognition work (including LLM-based detection), real examples by industry, and how to train your own.
Use live chat to support your customers in real-time
What is a chatbot intent?
A chatbot intent is the purpose behind what a user types or says. It’s the bot’s read of why someone sent a message, separate from the exact words they used. “Where’s my package,” “I haven’t gotten my order yet,” and “track shipment” are three phrasings of one intent: check order status.
Intent rarely travels alone. To act on a message, a bot also reads the entities (the specific details) and the context (what came before). Together they’re the working trio behind every accurate response.
| Term | What it captures | Example |
|---|---|---|
| Intent | The goal of the message | cancel_order |
| Entity | The specific details the bot needs to act | order #1234, “tomorrow” |
| Context | What was said earlier in the conversation | the product named two messages ago |
Take the message “Cancel my order #1234 tomorrow.” The intent is cancel_order. The entities are the order number (#1234) and the date (tomorrow). The context is everything said earlier—which product, which account. A bot that catches the intent but misses the entities makes the customer repeat themselves. One that tracks all three acts on the first try.
Types of chatbot intents
Most chatbot intents fall into six buckets: informational, transactional, navigational, support, feedback, and small talk. The mix depends on what your bot is for. A support bot leans on support and informational intents, while a sales bot sees more transactional ones.
- Informational—the user wants facts. “What are your store hours?”
- Transactional—the user wants to do something. “Book a table for two at 7 pm.”
- Navigational—the user wants to get somewhere. “Where’s my account page?”
- Support—the user wants a problem fixed. “My app keeps crashing.”
- Feedback—the user wants to share an opinion. “The new checkout is confusing.”
- Small talk—the user is being social. “Good morning!”
Two more categories matter in practice. The first is the fallback intent (also called a no-match intent): the catch-all a bot triggers when it can’t match a message to any known intent with enough confidence. Instead of guessing, it asks a clarifying question or hands off to a human. A well-tuned fallback is what keeps a bot from confidently giving the wrong answer.
The second is separating chitchat from business intents. A friendly “how are you?” shouldn’t be routed the same way as “I want a refund.” Good bots answer the first briefly and prioritize the second.
Read more: see the full breakdown of chatbot types and which one fits your use case.
Chatbot intent classification: how it works
Intent classification is the process of sorting each incoming message into a predefined intent category. The bot reads the text, compares it against patterns it has learned, and assigns the most likely label—order_status, password_reset, pricing_question—so it can pick the right response or action.
Under the hood, classification runs through a short pipeline:
- Input. The user’s raw message arrives.
- Preprocessing. The text is cleaned and tokenized—lowercased, punctuation handled, sometimes spell-corrected.
- Classification. A model scores the message against every known intent and picks the top match.
- Confidence check. If the top score clears a set threshold, the bot proceeds; if not, it triggers the fallback.
- Route or respond. The bot answers, runs an action, or passes the conversation to an agent.
Teams usually group intents into a few working categories:
- Transactional intents that trigger an action (place order, cancel, pay).
- Informational intents that return an answer (hours, policy, specs).
- Support intents that start a troubleshooting flow or open a ticket.
- Routing intents that hand off to the right team or human.
Intent recognition and detection: traditional NLP vs. LLM-based
Intent recognition is how a chatbot works out what a user means from natural language. There are two main approaches in 2026: traditional NLP models trained on labeled examples, and large language models (LLMs) that infer intent with little or no training data. Most production bots now use a mix of both.
Traditional NLP
The classic setup defines each intent up front, gives it a set of training phrases, and trains a machine-learning classifier to match new messages to those intents. Tools like Dialogflow and Rasa work this way. It’s fast, cheap to run, and predictable. The catch: it only recognizes intents you’ve explicitly trained, and it struggles with phrasings it hasn’t seen.
LLM and zero-shot detection
LLMs flip the model. Instead of training a classifier per intent, you describe the intents in a prompt and let a GPT-class model decide which one fits—often with zero or a handful of examples (zero-shot or few-shot). The upside: no training set to build, and it handles messages it has never seen, including typos, slang, and mixed languages. The downside: it costs more per message, adds latency, and can hallucinate an intent that isn’t there, so you need confidence checks and guardrails.
| Approach | Strengths | Weaknesses |
|---|---|---|
| Traditional NLP | Fast, low cost, predictable | Needs labeled data; misses unseen phrasings |
| LLM / zero-shot | No training data; handles new phrasings | Higher cost and latency; needs guardrails |
| Hybrid | Cheap routing plus LLM fallback | More moving parts to maintain |
Hybrid and RAG routing
In practice, many teams combine the two. A lightweight classifier handles the high-volume, well-defined intents cheaply, and anything it isn’t confident about falls through to an LLM.
Retrieval-augmented generation (RAG) is a setup where the bot fetches relevant facts from your own content and feeds them to the LLM before it generates a reply. It helps to see the three jobs separately: classification labels the intent, retrieval pulls the matching facts (a help article, a product spec, an order record), and generation writes the answer. Because the model summarizes the documents it was handed instead of recalling them from training, replies stay grounded and current. A common pattern is to classify the intent, retrieve the matching knowledge, then let the LLM phrase the response.
Intent recognition for enterprise chatbots
At enterprise scale, the question isn’t which model is smartest. It’s which setup stays accurate and controllable across millions of messages. Four things decide that: classification accuracy on your real traffic, confidence thresholds that trigger fallback instead of wrong answers, clean handoff to human agents, and monitoring you can act on.
Lyro, Tidio’s AI agent, is an LLM-based example: you point it at your help center and product info, and it recognizes and answers customer intents without per-intent training, then hands off to a human when it hits its limit.
Read more: how NLP chatbots understand language, and an in-depth Lyro review.
Boost customer satisfaction with a professional AI agent
Chatbot intent examples (by industry)
Chatbot intents look different by industry, but they follow the same shape: a user message, the intent behind it, the entities the bot extracts, and the action it takes. Here are common examples across ecommerce, customer service, finance, insurance, travel, and healthcare.
| Industry | Example user message | Intent | Entities | Bot action |
|---|---|---|---|---|
| Ecommerce | “Where’s my order?” | order_status | order ID | Pull tracking, share ETA |
| Ecommerce | “Do you have this in size M?” | product_inquiry | product, size | Check stock, reply |
| Customer service | “How do I reset my password?” | password_reset | account email | Send reset steps or link |
| Customer service | “The app keeps crashing.” | tech_support | product, issue | Troubleshoot or open ticket |
| Banking | “What’s my account balance?” | balance_inquiry | account | Retrieve and show balance |
| Insurance | “What does my policy cover?” | policy_info | policy number | Pull coverage details |
| Travel | “Book a flight to London on July 10.” | flight_booking | destination, date | Search flights, start booking |
| Healthcare | “Schedule a dental appointment next week.” | appointment_booking | service, date | Offer slots, confirm |
Read more: deeper guides on ecommerce chatbots, customer service chatbots, and finance chatbots.
Intent datasets and training data
An intent dataset is a collection of example user messages, each labeled with the intent it represents. It’s what a traditional classifier learns from. A rough rule of thumb: start with 15–30 varied example phrases per intent, then expand using real messages once the bot is live.
You don’t have to start from scratch. A few well-known public datasets are worth studying:
- CLINC150—150 intents across 10 domains, with out-of-scope examples for testing fallback.
- Banking77—77 fine-grained intents inside a single banking domain.
- Kaggle hosts several smaller intent-recognition sets that are good for prototyping.
To build your own, pull real questions from your chat logs, help desk tickets, and search queries, then label each one with the matching intent. Cover the messy versions, not just the clean ones: typos, half-sentences, and the three different ways people ask the same thing. That variety is what makes a classifier hold up in production.
How to train chatbot intents (5 steps)
Training intents is an ongoing loop, not a one-time setup. Five steps cover it.
1. Define clear intent categories
Start with the goals your users actually have. Mine your chat logs, support tickets, and FAQ for recurring requests, then group them into distinct intents like order_status, returns, and shipping_policy. Keep them separate enough that the bot won’t confuse two.
In Tidio, smart Conversations Views detect intent automatically by topic. To add one, click the + above the Views menu in your inbox, name the view, pick a topic, and hit Create.
2. Collect and label training data
Gather real user messages and tag each with its intent. Cover a wide range of phrasings per intent: “where’s my order,” “order not here yet,” and “track my package” all map to order_status. The more natural variety you feed in, the better the model generalizes.
In Tidio, you can map several intents to the same view. Use Edit on a view—say Shipping policy—to group related intents like shipping price and shipping change under it.
3. Train the model (or skip training with an LLM)
With a traditional classifier, you feed the labeled data to an NLP model and tune it until it reliably maps messages to intents. With an LLM-based setup, there’s often no training step at all: you connect your knowledge base and write prompts that define the intents, and the model handles the rest. Tidio’s Lyro takes the second route—add your FAQ and help content, and it’s ready to recognize intents without a training set.
4. Incorporate user feedback
Once the bot is live, watch where it misreads intent. Add a quick post-chat rating, review the misclassified messages, and feed corrections back into the training data or prompts. New intents will surface that you didn’t anticipate.
5. Monitor and update
Track the metrics that show whether recognition is holding up: fallback rate, resolution rate, and customer satisfaction. When a new pattern shows up—a product launch, a recurring complaint—add or refine intents to match. Recognition decays if you leave it alone.
Read more: a full walkthrough of training a chatbot and the chatbot analytics worth tracking.
FAQ
A chatbot intent is the goal behind a user’s message—what the person is trying to accomplish, such as tracking an order, resetting a password, or comparing plans. The bot identifies the intent first, then chooses how to respond or which action to run.
The intent is the goal of the message; the entity is the specific detail the bot needs to act on it. In “Cancel order #1234,” the intent is cancel_order and the entity is the order number. One says what to do, the other says with what.
Intent classification is the process of sorting each incoming message into a predefined intent category. The bot scores the message against every intent it knows, picks the best match, and checks its confidence before responding or handing off to a human.
Chatbots recognize intent in two main ways: a traditional NLP model trained on labeled example phrases, or a large language model that infers the intent from the message directly. Most production bots combine both, using a classifier for common requests and an LLM for the rest.
Yes. LLMs can classify intent with little or no training data, often from a prompt that lists the intents (zero-shot or few-shot). They handle phrasings they’ve never seen, but cost more per message and need confidence checks to avoid inventing an intent that isn’t there.
Common examples include order_status, product_inquiry, and returns in ecommerce; password_reset and tech_support in customer service; and balance_inquiry in banking. Tools like Lyro recognize these automatically from your help content.
