Grok's Voice Agent Builder: Build an AI Phone Agent in 2 Minutes

You can now build a voice AI agent from a plain-English description, in about two minutes, without writing a line of code. That's the pitch behind xAI's Voice Agent Builder, which came out of beta on July 1 and picked up 21 new voices on July 6. If you've ever wanted your own AI phone agent, this is worth twenty minutes of your time. Here's what it actually does, what it costs, and where it falls short.

What you're actually building

Forget the usual voice-bot setup, where you stitch together a speech-to-text model, a language model, and a text-to-speech model, and then pray the latency stays under a second. Grok's builder runs on a single speech-to-speech model instead, which is why responses come back fast enough to feel like a real phone call rather than a walkie-talkie exchange.

The workflow is simple. You write out how you want the call to go, in ordinary sentences: greet the caller, ask what they need, check the order-status database, offer to text a summary. Then you attach whatever it needs to do that job, like documents for it to reference, tools it can call, and guardrails on what it shouldn't say. Two minutes later you have a working agent with a live phone number attached to it.

Voices, and the cloning question

Out of the box you get 80-plus built-in voices, now including 21 flagship options added this week. You can also clone a voice from two minutes of audio, which is either the most useful feature here or the one you should think hardest about before you use it.

If you're building a small-business phone agent, voice cloning means your AI can sound like your actual receptionist instead of a generic bot voice. Callers already familiar with that voice get a more natural experience. But keep the two-minute requirement in perspective: the FTC has warned that scammers can produce a convincing clone from as little as three seconds of audio, far less than what xAI asks for here. That's not a knock on this particular tool, it's a reminder that the underlying capability is already loose in the world, and a public-facing business voice is easy raw material for it.

What it costs to run

Pricing is usage-based: $0.05 per minute of audio, voices included, no separate platform fee on top. Every account comes with one free provisioned phone number, and you're billed an extra penny a minute for that line. If you already have a business number through a telephony provider, you can connect it directly over SIP instead of using the free one.

Run the math before you commit a workflow to it. A voice agent fielding 500 five-minute calls a month runs about $150 in audio costs alone. That's cheap next to a part-time phone rep, but it adds up fast if your use case involves long, meandering calls rather than quick, resolved ones.

Who should actually use this

If you're running a small business and drowning in repetitive phone calls, appointment confirmations, order status checks, basic FAQ, this is a real option worth testing this week. The setup time is genuinely low, and the free number means you can pilot it without touching your existing phone system.

If you're building something more complex, a multi-step sales qualification flow, a support agent that needs to escalate to a human mid-call, spend time on the guardrails step before you go live. A voice agent that sounds natural is more convincing when it's wrong, not less, so the fallback behavior matters more here than it does with a text chatbot.

The multilingual support, more than 25 languages with in-call switching, is a legitimate advantage if your callers don't all speak English. Slator's rundown of the launch flags this as one of the stronger language-coverage claims among no-code voice tools right now.

It's the same company that fired the safety engineer who raised concerns about Grok's weapons capabilities, so treat the guardrails step as load-bearing, not optional. If you want to see how Grok itself performs before wiring a phone number to it, chatbot.gallery's Grok profile has the full breakdown.

Two minutes to build something is also two minutes to build something you haven't fully thought through. Test with real callers, not just yourself, before you point a live number at customers.

Want the next launch like this one in your inbox before it's old news? Subscribe to About.chat Weekly — one email, every week, no fluff.