For about five US cents a minute, you can now rent the one customer-facing job that still needed a person on the line. On the first of July, xAI switched on a Voice Agent Builder: no code, a working voice agent stood up in under two minutes, five cents a minute of audio and a cent more for the phone number.1 A ten-minute call runs to roughly sixty cents. It arrived in the same fortnight OpenAI put GPT-5.6 into general availability, which made it easy to file under model news when it is really a repricing: a receptionist’s hour, sold by the minute.

Voice was the holdout. Chat, email, the FAQ page, all of that went to software years ago. The phone stayed human because a call carries the things text drops: the pause before a question, the worry under a booking, the customer who is really asking whether it is worth it. A durian stall in Geylang on a Saturday afternoon takes more calls than it can answer, and every unanswered ring is an order that went to the next stall. The owner cannot clone himself. He can now clone the part of himself that reads out opening hours.

He can now clone the part of himself that reads out opening hours.

That is the whole move, and it is worth saying plainly. Roughly four in five calls to a small operator are scripted: are you open, do you have Musang King today, can I move my two o’clock, where do I park for the farm venue, my delivery is late. Those go to the machine. The one good person stays on the fifth call, the one where somebody is deciding whether to spend the money. Reading hesitation and turning a complaint into a repeat buyer is the job that pays, and it is the job a five-cent agent is worst at.

The economics only work if you aim the agent correctly. One Southeast Asian vendor reported a five percent lift in conversion after it moved routine inquiries off its people and freed them for the calls that close. The freed hour is the return, and it goes to the calls that close, not the headcount line. The tool is also not plug-and-play here: platforms trained on American speech stumble on Singlish and regional accents, and the deployments that hold up run localised, pulling round-trip latency down from over 1,200 milliseconds to under 500 so the agent does not talk over a caller. A bot that steps on a Malay-inflected sentence loses the room in the first ten seconds.

The machine takes the calls where no money is won, and your one good human takes the calls where it is.

Here is where it fails, and the failure is expensive. For a purchase that turns on taste or judgment, the durian buyer asking whether today’s fruit is worth thirty-eight dollars a kilo, the couple touring a farm venue for a wedding, a five-cent agent will answer smoothly and lose you the sale. It has no palate and no read on the room, and the buyer hears the seam. So the rule writes itself: the machine takes the calls where no money is won, and your one good human takes the calls where it is. The reminder, the stock check, the reschedule, the after-hours “are you open” go to Grok. The tasting, the tour, the complaint that could go either way stay human.

The stall owner who gets this right answers the same volume of calls and spends his Saturday on the ten conversations that turn a stranger into a regular, while the phone quietly handles the ninety that would have burned the afternoon. The cost of finding out is sixty cents a call.

Footnotes

  1. xAI bills the audio and the phone line separately, and the server-side lookups an agent makes to answer a question (web and X search) cost more again, so sixty cents is the floor on that ten-minute call, not the ceiling. ↩