The Hard Parts of a Voice Travel Concierge Aren't the Voice

Amazon's walkthrough for adding a voice travel concierge to an airline app begins with a plain premise: a traveler can change a seat or check a delay by speaking, without leaving the app or tapping through screens. The instructive part is the list of things that must work before that sentence is true. Audio streams in both directions, and the system has to hold the thread of a conversation across many turns — a memory problem more than a speech-recognition one. The agent also has to reach existing backend systems without being tightly coupled to them, and it has to scale when traffic spikes before a holiday weekend. None of that is the model. It's streaming, session state, integration seams and capacity planning, which is usually what decides whether a voice layer feels like an assistant or like a phone tree that learned to talk.