SpaceX says its AI voice agent resolves about 70 percent of Starlink calls with no human. Before you buy a tool like that, here is how to find out what "resolved" really means for your customers.
By Warren Schuitema, AI Systems Consultant, Matchless Marketing / Matchless AI
Published: October 16, 2026 | Last Updated: October 2026
Every AI customer service pitch comes with a big number.
70 percent resolved. 80 percent handled. Thousands of calls a day with no human needed. It sounds like you could hand off your phone and your inbox tomorrow and get your evenings back.
So how do you know if an AI customer service agent is actually working? You decide what "resolved" means before you turn it on. Then you measure that, and only that. For a small business, a good working definition is simple. The customer's problem got fixed, the fix was done right, and they did not have to call back about the same thing within a week. If a tool can't show you that number, it hasn't shown you that it works.
The rest is details. Here is how to get them right.
Key Takeaways
- On October 14, 2026, SpaceXSI launched Grok Voice 3.0 and said Starlink uses Grok Voice to answer about 20,000 calls a day, with roughly 70 percent resolved without a human.
- The company did not say how it defines "resolved," what time period it measured, or how many callers came back.
- A call that ends without a human proves the call ended. It does not prove the problem got fixed.
- Customers still want a way out. In a Gartner survey, 64 percent said they would prefer companies not use AI for customer service, and their top worry was not being able to reach a person.
- Before you buy, write your own definition of resolved, test the hard cases, and put the tool on a 90-day trial with clear pass or fail numbers.
A Big Number Feels Like Proof
I get why these numbers land.
If you run a small business, your phone is a leash. You are on a client call and it buzzes. You are at your kid's practice and it buzzes. Somebody needs a password reset, a refund, a delivery update. Each one takes five minutes, and each one pulls you out of whatever you were doing.
So when a vendor says their AI handles 70 percent of calls on its own, your brain does quick math. Seven out of ten interruptions, gone.
Here is the tension. That number might be true, and you could still end up with more work.
I spent 16 years in manufacturing and supply chain. One of the first things you learn on a production floor is that "shipped" and "good" are two different words. A part can leave the dock on time and still come back as a warranty claim three weeks later. If you only count what shipped, your numbers look great right up until the returns pile up.
AI customer service has the same trap. Calls that end are easy to count. Problems that stay fixed take more work to count. And the difference between those two numbers is where your time, your money, and your customers' patience go.
SpaceX Just Put a Number on It
On October 14, 2026, SpaceXSI launched Grok Voice 3.0, a voice AI model that businesses can build phone agents on. TechRepublic covered the launch the next day.
The headline claim. SpaceXSI says Starlink call centers use Grok Voice to answer about 20,000 customer calls a day, in several languages. It says roughly 70 percent of those are resolved with no human involved. It also says about one in four sales calls ends in a purchase before the caller hangs up.
What the agent can do. According to the company, it can verify a caller's identity, troubleshoot hardware, issue account credits, and set up replacement equipment. That goes well past answering FAQs.
The price. The voice API costs $0.08 per minute, the same as the last version. The Neuron worked out that a five-minute call costs about 40 cents in time charges, before other fees.
The benchmarks. Grok Voice 3.0 scored highest on Artificial Analysis' overall speech-to-speech index. But on a test built around customer support calls with accents, background noise, and interruptions, it scored 71.6 percent and trailed OpenAI's voice model at 74.5 percent.
What's missing. TechRepublic noted that the company did not share the measurement period, the call sample, or its definition of resolution. The Neuron listed more open questions. Were callers' problems actually fixed? How many called back? Did callers who gave up and hung up count as resolved? How often did people ask for a human, and did they get one?
None of that means the number is wrong. Starlink may run a great operation. It means a small business owner cannot copy that number into a buying decision. You need to know what is underneath it.
The Research Says Customers Still Want a Way Out
The vendors have their numbers. The research on customers tells a more careful story.
Most customers are wary. Gartner surveyed 5,728 customers in December 2023 and released the results in July 2024. 64 percent said they would prefer that companies not use AI for customer service. 53 percent said they would consider switching to a competitor if they learned a company was going to use it. The top concern was that AI would make it harder to reach a person. Gartner's Keith McIntosh put it this way: "Many customers fear that GenAI will simply become another obstacle between them and an agent."
Big companies are walking back their cuts. In June 2025, Gartner predicted that by 2027, half of the organizations that expected to cut their customer service staff significantly because of AI will drop those plans. In a Gartner poll of 163 customer service leaders that March, 95 percent said they plan to keep human agents. Gartner's Kathy Ross warned that AI does not fix everything and described the goal as "digital first, but not digital only."
One famous example already reversed course. Klarna said in February 2024 that its AI assistant could do the work of 700 customer service agents. It handled about 2.3 million conversations in its first month. By May 2025, CEO Sebastian Siemiatkowski told Bloomberg the company was hiring human agents again. As reported by Entrepreneur, he said AI support was cheaper but lower in quality, and that "it's so critical that you are clear to your customer that there will always be a human if you want."
Put those together. The AI is getting better fast. Customers are still nervous. And at least one company that went all in found that the cheap number and the good number were not the same number.
So measure it the right way before you trust it.
Count the Problems That Stay Fixed
In manufacturing, we had a measure called first pass yield. It asks a simple question: out of everything we made, how much came out right the first time, with no rework and no returns? It is a tougher number than "units shipped," and it is the one that tells you if the process actually works.
AI customer service needs its own first pass yield. I call it the Stays-Fixed Test. A call only counts as resolved if all three of these are true:
- The problem got fixed. The customer got what they called for.
- The fix was done right. The credit went to the right account. The replacement went to the right address. The appointment landed on the right day.
- It stayed fixed. The customer did not come back about the same issue within seven days.
Run any vendor number through those three questions and it gets more honest.
Here is a quick example with made-up numbers, just to show the math. Say an AI agent handles 100 calls and the dashboard says 70 were resolved. You check, and 15 of those 70 callers called back within a week about the same problem. Five more had a credit or booking that went in wrong. Your real number is 50 out of 100.
That changes the cost math too. If the agent costs about 40 cents per call, 100 calls cost $40. Divided by 70 "resolved" calls, that looks like 57 cents each. Divided by 50 calls that actually stayed fixed, it is 80 cents each. And that is before you count the time you spent cleaning up the other 20.
TechRepublic made the same point in plainer terms. Measure the cost per successfully resolved call, including setup, monitoring, and human backup, instead of the per-minute price.
I Put My Own AI Agents on Probation
I learned to think this way about AI by running my own business on it.
I run Matchless with a workforce of AI agents that handle a lot of the day-to-day work. Each one has a written job description. And every one of those job descriptions starts the agent on a 90-day probation, with written success criteria, and a daily recap of what it did.
Nobody on a real team gets hired and trusted on day one. They earn it. My agents work the same way. For 90 days, I am watching whether the work holds up after it is done. The daily recap means I see what happened every day instead of finding out a month later.
I judge my home systems the same way. My Weekly Meal Planner replaced HelloFresh in our house. I did not decide it worked because it made meal plans. I decided it worked because the Sunday planning spiral stopped. My Digital Wall Calendar replaced a $50 a month subscription. It passed because the nightly "what does tomorrow look like" conversation went away and stayed away.
That is the Stays-Fixed Test at home. The question is never "did the tool do something?" The question is "did the problem stop coming back?"
If you want a broader look at where AI agents help today and where they still need a human checking the work, my post on whether AI agents actually handle your everyday tasks goes deeper.
A Small Business Can Measure This Without a Data Team
You might be thinking this sounds like a big company problem. Dashboards, analysts, call center software.
A small business can do this too, and it has one big advantage. You know your customers by name. You can tell when the same person calls twice. A spreadsheet and 20 minutes a week will tell you more than most vendor dashboards.
You only need to track four things:
- Handled. How many calls or chats did the AI take from start to finish?
- Stayed fixed. Of those, how many customers did not come back about the same thing within seven days?
- Handoffs. How many people asked for a human, and how fast did they get one?
- Fix-ups. How many times did you or your team have to correct something the AI did?
That's it. If "stayed fixed" is close to "handled," the tool is working. If there is a big gap, you have found where your time is leaking.
And keep the human door wide open. Gartner's research and Klarna's reversal point the same way. Customers relax when they know they can reach a person. Make "talk to a human" easy to say and fast to get. An AI agent that traps people will cost you customers no dashboard will ever show you.
If you are still deciding whether AI should answer your phone at all, I wrote about the missed-call side of this in my post on AI phone agents for small business.
7 Steps to Test an AI Customer Service Agent Before You Trust It
You can do all seven of these with a spreadsheet and the tool's free trial or demo.
- Write your definition of resolved first. Before you talk to a vendor, write one sentence: "A call is resolved when the problem is fixed, the fix is correct, and the customer does not come back about it within seven days." Ask every vendor how their number compares.
- Count your starting point. For two weeks, tally the calls, emails, and messages you get now. Note what each one was about and how long it took. This is your baseline, so you can tell if the AI made things better.
- Pick the top three request types. Most small businesses get the same few questions over and over. Start the AI on only those. Leave the rest with a human until the first three pass.
- Decide what the AI can change on its own. Can it issue a refund? Change an appointment? Give a discount? TechRepublic's advice is to spell out which actions the agent can take, which need approval, and when it must hand off to a person. Start narrow.
- Test the ugly cases. Call it yourself, or ask a friend to. Try a failed identity check, a billing dispute, an angry customer, a request that breaks your policy, and a plain "I want a person." Watch what happens.
- Track the four numbers every week. Handled, stayed fixed, handoffs, and fix-ups. Ten minutes on a Friday. Listen to or read five real conversations while you are at it.
- Run a 90-day probation, then decide. Set a pass mark before you start. For example: stayed fixed is at least 85 percent of handled, and every customer who asked for a human got one the same day. At 90 days, keep it, adjust it, or cancel it based on those numbers.
Frequently Asked Questions
What does "resolved" mean for an AI customer service agent?
It depends on who is counting, and vendors often don't say. Some count any call that ends without a human. A stronger definition has three parts: the problem got fixed, the fix was done correctly, and the customer did not come back about it within about a week. Ask every vendor which definition they use.
Is a 70 percent AI resolution rate good?
It might be, but the number alone can't tell you. You need to know how resolved was defined, how long it was measured, and how many customers called back. A 70 percent rate where many callers return is worse than a lower rate where problems stay fixed. Judge it by repeat contacts and correct outcomes.
How much does an AI voice agent cost per call?
Prices vary by vendor. As one example, SpaceXSI's Grok Voice 3.0 API costs $0.08 per minute, so a five-minute call runs about 40 cents in time charges. Setup, phone service, monitoring, and human backup add more. The number to watch is cost per call that actually stayed fixed.
Do customers mind talking to AI customer service?
Many do. In a Gartner survey of 5,728 customers, 64 percent said they would prefer companies not use AI for customer service. Their top concern was trouble reaching a person. Customers are more comfortable when a human is easy to reach, so always offer a fast, clear way to talk to someone.
How long should I test an AI customer service tool?
Give it about 90 days with clear pass or fail numbers set up front. That is long enough to see repeat contacts, seasonal swings, and the odd cases that only show up now and then. Track handled calls, problems that stayed fixed, handoffs to a human, and fixes you had to make yourself.
Measure What Comes Back
Back to that big number on the vendor's slide.
70 percent might be real. It might even be great. But it measures what left the dock. You are the one who gets the call when something comes back.
So run every AI customer service tool through the same test I use on every system I build.
Time. Does it give you back hours, or does it trade phone calls for cleanup? Your "fix-ups" number will tell you.
Money. What does each problem that actually stayed fixed cost you? Do that math first.
Energy. Do you stop worrying about the phone, or do you start worrying about what the AI told your customers? If you are checking every conversation at month three, it isn't passing.
The tools are getting better every month. That is good news for every small business owner who wants their evenings back. Just hold them to the same standard you would hold a new hire. Clear expectations, a fair trial, and a decision based on whether the problems stay fixed. That is A Better Way.
Keep Learning
- AI Phone Agents for Small Business: How to Stop Losing Revenue Every Time You Step Away From Your Desk The case for putting AI on your phone in the first place.
- Do AI Agents Actually Handle Your Everyday Tasks, Or Is That Still Hype in 2026? Where AI agents help today and where they still need a human.
- What the AI Labs' Agent Security Crisis Means for Your Business How to check what your AI tools can reach and change.
- How Do You Turn a Client Call Into a Ready-to-Send Action Plan in Under 10 Minutes? Using AI on the calls you still take yourself.
Related: once an AI tool is live, the next question is whether AI search is sending you new customers. Here is how to track AI search traffic from ChatGPT and Google AI.
Model changes are the other half of that testing habit. When OpenAI retires the model under a tool, the owners who wrote down what "working" looks like can test the swap in an afternoon. My guide to getting ready for the GPT-4o API shutdown walks through the AI Parts List I use to track every model before a deadline lands.
About the Author
Warren Schuitema is an AI Systems Consultant with Matchless Marketing and Matchless AI. He spent 16 years in manufacturing and supply chain, where he learned that the number that matters is the one that stays good after it ships. He runs his business day to day with a workforce of AI agents, each on a 90-day probation with written success criteria. At home, his Digital Wall Calendar replaced a $50 a month subscription and his Weekly Meal Planner replaced HelloFresh. He helps small business owners and parent entrepreneurs build practical AI systems that give back Time, Money, or Energy, and he is a partner in the Goal Achievers Summit.