Intent Detection

A typical intent detection pipeline has three pieces. First, an inventory of intents the system supports, for example "product question," "size question," "shipping question," "order status," "return," "complaint," and "small talk." Second, a classifier that maps an incoming message to one or more of those intents. Third, a router that decides what to do based on the predicted intent: retrieve products, look up an order, escalate to a human, or stay in the conversation.

Classic systems used hand-written keyword rules. They were easy to debug but missed paraphrases and broke on misspellings. Modern systems use one of two approaches. The first is sentence embeddings: every message is turned into a vector and compared to anchor examples for each intent. The second is a large language model with a structured-output prompt that returns the intent label directly. The LLM approach handles ambiguity well and can return multiple intents per message, which matters because a single message often combines a question and a complaint.

For example, "is this still in stock and when will it ship to Bangalore" contains three intents: stock check, shipping ETA, and location-specific delivery. A good system tags all three so the reply can answer them in one turn.

For Shopify merchants, intent detection is what turns a chatbot from a clever toy into a useful assistant. Knowing the shopper's intent lets the system choose the right tool: search the catalog, look up the order, surface a return policy, or escalate. Without it, the bot defaults to either generic answers or repeated requests for clarification.

Intent data also feeds analytics. A merchant who can see that 30% of pre-purchase questions are about shipping, 20% are about sizing, and 10% are about ingredient lists has a clear list of product page improvements to make. Shop Me, for example, surfaces top intents in its growth dashboard so merchants can use chat data to fix the rest of the storefront.