Two years ago, a voice agent that could hold a natural conversation was a demo. Today it answers real phone lines for clinics, dealerships and coaching institutes — but only when the engineering underneath is right. The gap between a great demo and a great deployment is not the model; it's everything around it.
The latency budget is the product
Humans notice a pause of more than about 800 milliseconds. That budget has to cover speech recognition, the language model, retrieval from your knowledge base and speech synthesis — combined. Production systems get there by streaming every stage: the agent starts formulating a reply while the caller is still finishing the sentence, and starts speaking while the rest of the answer is still being generated.
Barge-in, or the art of being interrupted
Real callers interrupt. A production agent has to stop talking the instant the caller speaks, discard the rest of its planned sentence, and listen — the same way a good human receptionist does. Agents that plough on through an interruption feel robotic no matter how natural the voice is.
Languages are not a checkbox
In India especially, callers switch languages mid-sentence. An agent serving Bengaluru sees Kannada, Hindi and English in one call. What works is detection and switching on every turn, not a language picker at the start of the call.
The result, when it's done right: every call answered on the first ring, around the clock, in the caller's own language — with the difficult engineering invisible to everyone but the team that shipped it.