Intent Classifier

An intent classifier is a mechanism that first figures out what the user wants, then routes the conversation to whichever agent or system is set up to handle that request. Think of a government office's front counter. A visitor states their business, and the clerk immediately sorts them: "That's tax matters, third floor," or "Passports, second floor." In an AI system, a model or a set of rules plays that clerk's role, looking at an incoming message and sorting it into categories like "this is an order lookup" or "this is a refund inquiry," then handing each one off to a separate program dedicated to that task.

This approach has the advantage of keeping things simple, since each dedicated handler only has to deal with its own job. But shuffling the visitor from department to department can break the thread of the earlier conversation, and ambiguous requests can leave the classifier unsure where to send them. Because of this, some newer designs drop the classifier altogether and instead have a single front-desk staffer stay with the visitor the whole way through, pulling out whatever file is needed along the way. Rather than cutting the conversation into pieces, it stays as one continuous flow, and specific capabilities are pulled in only when needed.

Articles have noted that Anthropic's published commerce agent design notably lacks an intent classifier. Instead of splitting the conversation and handing it off to specialized agents by department, the design has a single model stay with the conversation from start to finish, pulling in skills only as needed. This doesn't mean intent classifiers are a flawed technology in general — it simply means this particular design chose a different approach.