As of this week, the question of whether AI can handle real phone calls at real scale is settled. According to The AI Marketers' October 15, 2026 roundup, Starlink's call centers now route 20,000 calls a day through xAI's Grok Voice 3.0, and roughly 70% of those calls close without a single human agent ever joining. The AI checks who's calling, troubleshoots the hardware, issues credits, ships replacement equipment, and closes one in four sales calls before the caller hangs up. On its own.
That is not a pilot. That is an operating system.
I've been building AI receptionists for local businesses for a couple of years now. An orthodontic practice, a med spa, a couple of home service companies. Every single time, the first objection I hear is some version of "our clients expect a real person." I used to spend ten minutes on that objection. Now I'm going to spend about ten seconds: Starlink's customers were expecting a real person too. Seventy percent of them got their problem solved anyway and apparently didn't complain enough to make the news.
What actually happened
xAI released Grok Voice 3.0, an AI voice model designed for businesses running support and sales lines. The source reports that independent benchmark firm Artificial Analysis currently ranks it number one on its Speech-to-Speech Index, and it responds in under a second. The price is $0.08 per minute through xAI's Voice Agent API, which works out to $4.80 an hour. That's the same price as Grok Voice 2.0. You get a materially better model for the same money.
The Starlink deployment is the headline. Twenty thousand calls a day across Starlink's support operation, 70% resolved end to end by the AI, with no human ever involved. The AI doesn't just answer questions; it takes action. It issues credits. It ships hardware. It closes sales. That last number is worth sitting with: one in four sales calls converted before the caller disconnects.
The math xAI gives you is simple. Multiply last month's inbound call minutes by $0.08. That's roughly what Grok Voice would have charged to answer every one of them.
The benchmark problem (and why it matters to you)
Here's the concept I want you to take from this piece, something I'm calling the Benchmark Ratchet. Once a public company posts a real-world number, that number becomes the floor. Not for Starlink. For everyone. Your customers don't know Starlink's stats, but their patience for hold times, missed calls, and "let me check on that for you" has been quietly shaped by every interaction they've had with a system that just... worked. The expectation ratchets up whether or not you were watching.
I saw this in auctions years ago. The moment eBay made real-time price discovery available to anyone with a laptop, the idea that a dealer's word was the market died overnight. The Benchmark Ratchet. Once the public number exists, you're being graded against it.
Seventy percent resolution without a human. That's the number. Your front desk is being measured against it right now.
What it means for your business
You do not run Starlink. You probably answer somewhere between 20 and 200 calls a week. So let's translate this to your scale.
The calls your staff handles every day are almost certainly dominated by a short list: hours, scheduling, pricing, directions, status updates, insurance questions, whether you're taking new patients or new clients. McDonald's found the same thing at the drive-thru: the same five or six questions, over and over, every shift. An AI voice agent trained on your specific FAQ handles those without putting anyone on hold, without missing a call at 7pm on a Tuesday, and without requiring a salary, benefits, or PTO.
The $0.08 per minute pricing from xAI is the going rate for raw API access, meaning you'd need a developer or an AI implementation partner to wire it into your phone system. Turnkey platforms for small and mid-sized businesses run higher but bundle the setup. TPC Scottsdale built a working version of this on a platform designed for exactly that use case. You don't have to start from an API.
The honest cost comparison is not AI versus free. It's AI versus a front desk employee who costs $3,500 to $4,500 a month fully loaded, is unavailable after hours, takes sick days, and handles maybe 60% of calls without escalating anyway. A Zenoti survey found 71% of med spa clients are already comfortable with AI handling their inquiries. The "real person" objection is aging faster than you think.
Under the hood
Grok Voice 3.0 is a speech-to-speech model, meaning it processes audio in and returns audio out without converting to text in the middle. That's why the sub-one-second latency is real and not a marketing stat: there's no transcription step adding delay. The Artificial Analysis Speech-to-Speech Index ranking reflects latency, accuracy, and naturalness scores across standardized test sets, not a self-reported benchmark. What's new in 3.0 compared to 2.0 is improved accuracy (xAI's claim) and that independent top ranking. The pricing is unchanged, which is unusual and probably not permanent. The Starlink deployment also shows something technically significant: the agent isn't just answering questions, it's taking actions inside backend systems, issuing credits and shipping hardware. That's an AI agent operating with real permissions on live systems, not a glorified FAQ bot. If you're evaluating platforms, that distinction matters. An answering service is cheap. A system that can actually complete a transaction is where the business case compounds.
Get The Root delivered weekly
We respect your privacy. Unsubscribe at any time.
What to do this week
Four moves, in order of effort:
- Run the math. Pull last month's call log or ask your phone carrier for total inbound minutes. Multiply by $0.08. Write that number down. That's your comparison point for every vendor conversation you have next month.
- List the top five calls. Have whoever answers your phone write down the five questions they answer most. If more than three are factual and repeatable, you have an AI receptionist use case. Full stop.
- Check what after-hours calls you're missing. Most phone systems log missed calls by time of day. If you're losing calls between 5pm and 9am, that's pure lost revenue with a known price tag, and an AI agent runs those hours for free.
- Read the TPC Scottsdale case if you want a local example of how this gets implemented without starting from scratch. It's a template, not a one-off.
Skip it if
Your business genuinely runs on high-touch, emotionally complex calls, think crisis counseling, bespoke legal consultations, or anything where the relationship is the product from the first hello. The Benchmark Ratchet still applies to your industry eventually, but you have more runway than a dental scheduler or a home services dispatcher. For now, you can watch.
But if your phone rings with the same five questions on a loop, you are paying a person to do a job a machine will do better, cheaper, and at 2am. Starlink just put a number on what that looks like at scale. The only question left is how long you want to keep pretending yours is a different situation.
So: what did your phone staff actually resolve last month without escalating? If you don't know that number, you're running a benchmark you can't see against a competitor you can't beat.
Sources
Frequently Asked Questions
See Where Your Business Ranks
Get a free competitive intelligence report. We analyze your website, scan your competitors, and test your AI search visibility, ready in 60 seconds.
Run My Analysis →Did this guide help?
Make our content appear first in your Google Search and AI Overview results by adding us as a preferred source.
Requires signing in to your Google account.
Add Wise Roots as a Google Preferred Source
