Conversational AI

The field moved through three broad generations. Rule-based systems matched patterns to scripted responses. Intent-based systems classified an utterance into one of a fixed set of intents, extracted entities, and ran a handler — the architecture behind most enterprise chatbots of the last decade. Language-model systems reason over the conversation directly and call tools, which removed the requirement to enumerate every intent in advance.

Voice imposes constraints that text does not. A spoken reply cannot be skimmed, so length is a real cost rather than a preference. There is no back button, so a misunderstanding must be repaired in dialogue. There is no rendering of lists or tables, so information must be sequenced. And latency is bounded by conversational expectation rather than by patience.

The corollary is that a good chat assistant does not become a good voice agent by adding speech at either end. Answers must be shortened, confirmations made explicit, options offered two or three at a time rather than as a list, and every branch must terminate in something the agent can actually say. The prompt is different because the medium is different.