Agents that answer the phone

The phone is the most expensive queue in a small business. It rings while the person who answers it is with a customer, and every missed call is a job someone else books. Voice agents have been almost good enough for years. In 2026 the models stopped sounding like the obstacle.

OpenAI's Realtime API came out of beta in August 2025, which turned low-latency speech-to-speech from a demo into something you could put behind a phone number. In May 2026 it added GPT-Realtime-2, a voice model with reasoning-class capabilities. In July it introduced GPT-Live, a full-duplex model that can listen and speak at the same time, followed by GPT-Live-1 in the API in September. The generated audio now carries a SynthID watermark where supported.

What full duplex changes

A turn-taking model waits for you to finish. A full-duplex model can back-channel — 'mm-hm', 'right' — while you talk, and can be interrupted mid-sentence without collapsing. On a good line, that is the difference between talking to a system and talking to a receptionist. It also means the agent can talk over someone, misread a pause as a turn, or keep selling after the caller has said 'no'.

Then there is the part no model release fixes: names. A caller says 'Jagiello', the transcriber hears 'Jagelow', and the confirmation text goes to nobody. Voice interfaces hide their uncertainty. A screen can show three possible spellings; a phone call cannot.

Treat the phone as an eval problem

Before a voice agent answers a real number, record a set of scripted calls with consent and grade them: was the name right, was the address right, did it stop when asked, did it hand over when unsure, how long was the dead air? Then run the same set every time the model or prompt changes. This is the same discipline as a text agent, with two extra dimensions — latency and audio — and a much lower tolerance for invention, because a caller cannot scroll up to check what was said.

Reliability research is a useful cold shower here. A February 2026 paper evaluating 14 models across two agent benchmarks found accuracy rising faster than reliability: agents get better at the tasks we measure, without becoming proportionally more dependable across repeated runs. A phone agent that is right 90% of the time per turn is not right 90% of the time per call.

Keep a human one ring away

The deployments that work share a shape. The agent answers, books and answers factual questions from an approved record. It does not take card numbers, give medical or legal advice, or promise a price it cannot see. It says it is an AI in the first sentence. It offers a person before the caller asks. And it logs a transcript the team can read the same day, because the first two weeks are where you learn what callers actually say.

Voice is finally a product decision rather than a research project. The businesses that benefit will be the ones who measure the calls, not the ones who buy the most lifelike voice.