What intent recognition is

A shopper types something blue and breathable for running. The system has to understand: this is a
product recommendation request, the category is running footwear, the parameters are blue and a
breathable material. That is intent recognition working alongside entity extraction.

An intent answers what the user wants to do, not what exactly they are looking for:

  • find_product — wants to find an item
  • compare_products — wants to compare options
  • check_availability — wants to know what is in stock
  • track_order — wants an order status
  • initiate_return — wants to return something
  • get_help — needs general help

How intent classification is built

Modern systems use several approaches.

Fine-tuned models: a small model such as BERT or DistilBERT is trained on labelled dialogues from
a specific store. Fast and cheap at inference, but it needs labelled training data.

An LLM with prompting: the query and the list of possible intents go to a language model. No
labelling required, new intents are easy to add, but inference costs more.

A hybrid: the LLM for rare and difficult cases, a small model for the routine ones — the best
cost-to-quality point.

Query: "when will my order arrive"
→ Intent: track_order (confidence: 0.94)
→ Entity: [implicit: order_id required]
→ Action: ask the shopper for the order ID

Why intent recognition is critical in e-commerce

In a shopping assistant, one wrong intent breaks the whole scenario. A shopper who gets an
irrelevant answer does not rephrase — they close the chat.

Wrong intent Consequence
Transactional read as informational A ready buyer gets an article — conversion stalls
Return read as product discovery Someone with a return sees new product promotion — irritation
Out-of-scope not detected The system answers outside its competence — hallucination

Tip: always design an explicit out_of_scope intent with a clear scenario: I cannot help with
that, connecting you to an agent. Its absence is the cause of most failed conversations.

Confidence and fallback

Most classifiers return a probability per class rather than just a winner. Low confidence below a
threshold is a signal to act:

  • Ask a clarifying question: do you want to find a product or check an order?
  • Hand over to an agent: when the shopper is stuck and the stakes are high (a return, a complaint)
  • Ask for a rephrase: I am not sure I understood, could you say more?

The optimal threshold depends on the task: transactional intents deserve a higher bar, because the
cost of being wrong is larger.