TL;DR
- Muse, Meta's assistant, already places calls to U.S. businesses for the people who use it (Reuters, September 22). In internal posts, a Meta employee testing it complained that his insurer wouldn't stay on the line once it heard AI.
- That insurer didn't hang up on a machine. It hung up on a customer who sent one.
- An assistant's call has more in common with somebody's executive assistant than with a scammer, and it gets handled the same way: on one condition.
- A machine gets told what's already public. What belongs to a person gets told only to that person.
- The decision that matters is the third one: an appointment an assistant asks for doesn't exist until the customer confirms it from their own phone.
- There's no published count of these calls yet. Deciding now costs half an hour; skip it and whoever picks up that day decides.
- You'll leave with your front desk's script: three lines, in the words of the person who answers, taped next to the phone.
The short answer. An AI assistant that calls a business is a program that talks on the phone for a person, to book an appointment or ask about a price. It isn't new: Google was booking restaurant tables this way in 2018. What's new is that in September 2026 Muse, Meta's assistant, started calling U.S. businesses for its users. The question is what your front desk does when one of those calls comes in.
On Monday we asked you for one thing: ask whoever answers your phone what they'd do if an AI assistant called to book an appointment for a client, and write down the answer without correcting it.
If you did, you have on paper where your front desk actually stands on something that's already started. If the answer was "I'd hang up," that's exactly what the insurer did. If it was "I don't know," that's the most honest answer there is. And if it was "I'd help them," the next question is what they'd hand over.
This piece picks up after that answer. What it takes is a three-line script and no new tool, and almost no business has one written down, because until now these calls were rare.
1. Hanging up on the assistant is hanging up on the customer
What Reuters published on September 22 has three parts, and the one that made the headline is the least useful for a business.
The headline part: to make the calls go better, Meta had contract workers place some of them, in a test with its own employees and without proper disclosure. Meta has pulled that test for now, and in an internal post a vice president in its Superintelligence Labs admitted that starting without proper disclosure was a mistake. That's Meta's stumble, and it stays Meta's.
The part that matters more: an employee who used Muse to call his insurer wrote that "they keep hanging up on Muse when they hear it is AI." Read that from the insurer's side. On the other end of the line was a customer who wanted something from them, who chose not to dial himself and handed the errand to his assistant. The insurer hung up, and the customer knew exactly what happened: he wrote it down inside his own company, and Reuters published it. The one who has no idea how often it happened is the insurer, because the side that hangs up keeps no record.
Hanging up by reflex feels careful, and leaves no trace of what it cost
In the moment, hanging up feels like the safe call. But a dropped call leaves nothing behind. Nobody logs it, nobody counts it, and a year from now nobody can say how many there were.
A day later, at Meta Connect, the company said Muse is getting its own email address and that you'll be able to ask it for an appointment from your glasses while you're driving or cooking. Meta's bet: more and more of the calls your customers make to you will be made by their assistant.
2. This call has more in common with an executive assistant than with a scammer
Until recently your front desk handled two kinds of calls: the customer, or someone calling for them. A daughter booking her mother's appointment. An executive's assistant booking for the boss. Your team has always known how to do that, and does it well: ask who it's for, write it down, confirm.
Now there's a third kind: a customer's AI assistant, calling on that person's behalf.
Filing it under scams puts it in the wrong box
The handiest box for "a strange voice on the phone" is the scam box. But an assistant's call looks a lot more like one from somebody's executive assistant. There's a real person behind it, with a real need, who chose not to make the call themselves.
That doesn't mean helping blind. You ask first who it's for, same as you would with anyone's assistant.
What to do. Decide with your front desk that a call from an assistant gets handled, on one condition: it says who it's calling for. If it doesn't say, you ask. If it can't tell you, you end the call politely. The opening line is short: "Happy to help. Who are you calling for, and what does that person need?"
If your business decides never to take these calls, that works too, as long as it was decided ahead of time and your front desk has a line for it that leaves the door open for the actual person.
3. A machine gets told what's public, and nothing more
The reasonable worry is what the machine walks away with.
Your business already has the answer, even if nobody wrote it down: an assistant gets told what's already public. What belongs to a person gets told only to that person.
Hours, address, services, prices already on your website, general availability on a given date: anyone can hear that, because anyone can read it. Whether someone is or isn't your customer, their history, their balance, a change to an existing appointment or a new phone number: none of that goes to anyone but that person, and a machine isn't that person even when it speaks for them.
The risk sits in answering the wrong question
A front desk that hangs up on every assistant and also confirms to any human voice that so-and-so is a patient has the problem backwards. The line doesn't run between people and machines. It runs between public information and someone's information.
If your business handles health or financial information, the privacy rules you already follow on the phone with any stranger still govern, and this piece isn't legal advice.
What to do. When the assistant asks for something that belongs to a person, the line is the same whether or not that person is in your records: "That's something we only handle directly with the person. If they call or message us, we're glad to help."
4. The appointment doesn't exist until the customer confirms it from their phone
Of the three decisions, this is the one that makes the other two work, and it's the one almost nobody has written down.
An assistant can get it wrong. It can write down Tuesday when the person said Thursday. It can ask for an appointment its owner no longer wanted. And now and then, the caller may not be who they claim to be. There's one answer that covers the first three without guessing anything:
An appointment an assistant asks for stays pending until the customer confirms it from their own phone.
If the person is already your customer, the text goes to the number in your records. If they're new, it goes to the number the assistant gives, and it's worth being honest: that text doesn't prove who they are, only that someone at that number wants the appointment.
Look at how much that settles without anyone guessing: nobody has to tell a machine's voice from a person's, and nothing new gets installed. One text to the customer's phone does it, and that text is what turns a request into an appointment.
What to do. Close the call with: "We've got the request down for [day] at [time]. What number should we text [name] at to confirm it? As soon as they reply, it's booked." And send the text the minute you hang up.
5. What we don't know, and why you can decide anyway
Nobody has published how many AI assistant calls a business gets today, or how many it'll get next year. Nobody knows either whether they'll all identify themselves as AI. Meta has been opening the calling feature to its users gradually.
If you don't decide, whoever picks up that day will
What is known is the cost of each path. Deciding now is half an hour with your front desk and a printed sheet. Not deciding costs something too: the decision gets made live, with no script, by whoever answers the day it rings. Sometimes they'll hang up, sometimes they'll give out something they shouldn't, and nobody will write it down.
If a month from now none have come in, you spent half an hour and the script is ready for when they start. If several came in, you know how many times it happened.
6. If the AI on the phone is yours, say so in the first sentence
This runs the other way too. If your business answers or places calls with an AI assistant, say so in the first sentence: "Hi, you've reached the virtual assistant for [your business]."
In the same story Meta tripped over the reverse case: it had contract workers place some calls that went out from its assistant, without proper disclosure. Hiding it just pushes the cost to the day the customer finds out.
And if you're thinking about putting an assistant on your own phones, before you pick one, go over what an AI sales assistant does and doesn't do and what an assistant should be allowed to do before you switch it on.
The sheet-by-the-phone conversation
Naty: When I had my spa in Venezuela, I wasn't the one taking the inquiries. And from the owner's side it's easy to forget: the person answering shouldn't have to guess what you think. If something new comes in and nobody said what to do, hanging up is the most natural thing in the world.
Todd: That's why the sheet works even though it's so short. It doesn't give that person new information. It gives them permission. It says: the owner already decided this, you just follow it.
Naty: And it takes a weight off. If the call goes strange, it wasn't their mistake: they did what was written down.
Todd: And if you look at the bet straight on, it's small. If they take the call with the script and it wasn't real, the worst you're left with is a pending request nobody confirms. If they hang up and it was a customer, that customer is already gone.
What you're doing today
With the answer your front desk gave you on Monday, you do one thing: the three-line script, written in your front desk's own words and taped next to the phone. Didn't ask on Monday? Ask before you sit down. The uncorrected answer is what you're about to improve.
If your front desk doesn't like one of the three, change it on the spot. The line that goes on the wall is the one that person will really say.
The guide has everything ready to print: the full script, the table of what a machine gets told and what it doesn't, the confirmation text to copy, what to do when the customer doesn't reply, and the log sheet. Send the word and it arrives in that same chat.
To close
The story of the week was a huge company tripping over its own phone calls. What's left for a service business is smaller and more useful: people have started handing off the calls they used to make to you, and your front desk is the door those calls come through.
All that door needs is three lines on paper: take the call on one condition, share only what's public, and the appointment doesn't exist until the customer confirms it from their phone.
Write them today. The day the first one rings, whoever picks up will already know what to say.
One last thing, for whoever answers your phone: if one of these calls comes in before the sheet is up, log it anyway. It's the first number of your own you'll have on this.
Keep reading
- The email and the phone call stopped proving who's there - Monday's roundup, where the question for your front desk started
- Give a program a job it can't finish, and no way to say so, and it will find another way out - what an assistant can do on its own inside your business, and who checks its work
- AI Chatbots and Agents for Business - what an AI assistant does when it's yours
- SEO vs. GEO vs. AEO: which one your business needs in 2026 - the full map
- Meta testing a "human concierge" for its new personal AI agent, Muse, Reuters, September 22, 2026 - the original story
Sources
- Meta testing a "human concierge" for its new personal AI agent, Muse, Reuters via BNN Bloomberg, September 22, 2026 - https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/22/meta-testing-a-human-concierge-for-its-new-personal-ai-agent-muse-reuters-exclusive/
- Everything we announced at Meta Connect 2026, Meta, September 23, 2026 - https://www.meta.com/blog/meta-connect-2026-everything-we-announced/
- I answered the restaurant phone when Google's human-sounding AI called, CNBC, June 26, 2018 - https://www.cnbc.com/2018/06/26/google-duplex-preview-i-answered-the-phone-when-googles-ai-called.html
- FCC Confirms that TCPA Applies to AI Technologies that Generate Human Voices, FCC, February 8, 2024 - https://www.fcc.gov/document/fcc-confirms-tcpa-applies-ai-technologies-generate-human-voices
