If you have ever tried an AI phone answering service and given up on it, the reason probably was not that the AI said the wrong thing. It was the rhythm. You'd start talking, it would talk over you. You'd pause to think, it would assume you were done and jump in. You'd say "no, wait —" and it would keep going. It felt like a walkie-talkie: one person at a time, press to talk, and any overlap breaks it.
On July 8, OpenAI shipped GPT-Live, a voice model built on what engineers call a full-duplex architecture. That term is the whole story, and it is worth understanding because voice is the one AI interface your customers will encounter without choosing to.
Half duplex versus full duplex, in plain terms
Older voice AI worked in strict turns. It listened, decided you were finished, converted your speech to text, generated a reply, converted that back to speech, and played it. Each of those steps added delay, and the "decided you were finished" step was a guess — usually just a timer counting silence. Talk with pauses in the middle of a sentence, like most people do, and the timer fires early.
Full duplex means the system listens and speaks at the same time, the way a phone line has always carried both directions at once. The model is processing your audio while its own audio is playing. That is what makes it possible for it to stop mid-word when you cut in, to say "mm-hm" while you're still talking, and to handle you changing your mind halfway through a sentence.
Why this matters for a business that answers phones
The failure mode of half-duplex voice AI was not wrong answers — it was callers hanging up. People tolerate an AI that needs a question repeated. They do not tolerate being interrupted, especially when they are already annoyed enough to be calling. If you evaluated an AI receptionist a year ago and rejected it on feel, the thing you rejected has materially changed, and it is worth a second trial call before you write off the category.
The honest caveat
Better conversational rhythm is not better judgment. A full-duplex model that interrupts gracefully can still quote a price you don't offer or promise an appointment slot you don't have. The architecture fixed the timing problem; it did not fix the knowing-your-business problem. That still comes from what you connect it to and what you let it say without a human.
The takeaway
When you evaluate voice AI now, test it rudely. Interrupt it. Pause in the middle of a sentence. Change your mind. The systems worth paying for are the ones that handle being talked over, because your callers will.