Intent detection
Quick definition
Intent detection is the task of identifying what a person wants to achieve from what they write or say, so that a system can respond or route the request. Typical intents are tracking an order, requesting a refund or asking about a product.
On this page
How intent detection works
Classic chatbots relied on an intent classifier trained on example phrases for each intent: “where is my parcel”, “my order hasn't arrived” and “tracking number?” all labelled track_order. A slot extractor then picked out details such as the order number. Research datasets such as BANKING77 (Casanueva and colleagues, 2020), with 77 fine-grained banking intents, show how hard fine distinctions are even for good models.
With large language models, intent detection is often implicit. The model reads the message, decides what the person wants and chooses which tool to call. Some teams still keep an explicit classification step to route conversations, tag them for reporting or trigger specific workflows.
Why it matters in customer service
Knowing the intent decides what happens next: answer from the FAQ, look up an order, open a return, send the conversation to billing, or call a person immediately for a complaint. Aggregated over thousands of conversations, intents also tell a business what customers really contact it about — the input for improving pages, policies and products.
In e-commerce the most frequent intents are usually order status, delivery times, returns and exchanges, product questions and payment problems.
Common pitfalls
Real messages often contain several intents (“where is my order, and can I change the size?”). Fixed intent lists age quickly as products and policies change, and a long tail of rare requests never fits any label. Confidence thresholds that are too low send customers down the wrong path; thresholds that are too high escalate everything. Systems should allow a fallback — ask a clarifying question or hand over — rather than force a guess.
Intent handling in AI SmartTalk
AI SmartTalk does not ask you to define an intent list. The assistant's language model interprets each message and chooses among its tools — knowledge search, product search, order lookup, appointment booking, handoff. For routing and reporting, conversations can be tagged, and SmartFlow workflows can branch with an AI condition on what the customer said. Questions the assistant could not answer are listed so that the team can fill the gaps.
Frequently asked questions
What is the difference between an intent and an entity?
The intent is what the person wants (track an order); entities are the details needed to do it (the order number, the email address).
Do LLM-based assistants still need intent classification?
Not to answer, because the model interprets the request directly. An explicit classification remains useful for routing, analytics and triggering fixed processes.
How many intents should a support bot have?
With a classic classifier, start with the dozen intents that cover most volume and add more as data shows them. With an LLM agent, focus instead on the tools and knowledge each frequent intent requires.