<p>In voice AI, silence is the interface. Every architectural choice - payload size, round-trip count, tool-call shape, and recovery behaviour - reaches the caller as a pause, which means the technical debt is audible.</p> <p>This talk follows an AI receptionist from a prototype that sounded slow and uncertain to a production system that does not, and it starts with the uncomfortable part: a turn-level breakdown of exactly where the silence went. James will make the case that most of it was not the model and that the fixes were mundane, cheap, and mostly about what the agent is asked to read and which component is trusted to hold state.</p> <p>Before-and-after call audio will allow the audience to hear the difference before it is explained. Attendees will leave with a way to budget silence across a conversational turn and a set of decisions worth revisiting in their own voice AI stack.</p>