Meta's Muse calls businesses for you. Some callers were human

Last Wednesday, Ryan Fox, a principal engineer on Meta's Muse agent, posted from @wailord that Meta had widened the Muse beta for outbound calls to US businesses. "Phone calling was one of our top requests," he wrote. About half an hour later, Alexandr Wang, who runs Meta Superintelligence Labs, quoted him from @alexandr_wang to announce the expansion. The pitch was simple: tell Muse to book a haircut, check whether a store has something in stock, or collect quotes from contractors, and the agent makes the call.

On Tuesday, two reports showed another part of the plan. Jason Koebler at 404 Media and Katie Paul at Reuters had both seen internal Meta posts about a "human concierge." In the test, Muse could pass a request to a "trained human agent, who places the call and works it through." Contractors in call centers handled some of the calls.

What Meta actually tested

According to both reports, this was an internal trial and not part of the public app. Reuters says Meta switched the feature on for half its employees last week and let them opt out. 404 Media reports that at least one tester found out a person had placed the call only after it was over.

Employees pushed back in the internal threads. One wrote, per Reuters, "We are one bug away from unnecessary information being leaked to human callers." Another reported that a contractor made a racist reference on a call to negotiate an internet bill. A third, quoted by 404 Media, worried the story would read as "their AI is not good enough so they still need humans."

Meta says the test was meant to gather feedback and that the feature will only launch "with the proper disclosures." According to Reuters, a vice president in Superintelligence Labs said it was "a miss" to start the test without disclosure, and that Meta had "rolled back this feature."

Why Meta still needs people on the phone

Reuters also has the most useful detail. In some tests, human callers pushed the success rate to 95% to 98%, up from a "lower percentage of AI calling." Meta didn't say how much lower.

That won't surprise anyone who has built voice automation. Phone calls go wrong in more ways than web forms do: hold queues, phone trees, transfers, a receptionist with a follow-up question, a manager who has to call back. An agent can stall at any of those points, so a human fallback is a sensible engineering choice. Gary Marcus pointed out that Facebook tried a similar mix with its M assistant from 2015 to 2018, with people quietly doing much of the work, and that project was shut down.

I don't mind the hybrid. Plenty of good services put a person behind the software for the hard cases. My problem is that nobody told the testers. People ask an agent to book a doctor's visit or haggle over a bill, and they assume software is handling the details. If a contractor might be handling them instead, users need to know that before the call, not after.

If you're the one picking up

If you run a salon in St. Charles or a roofing company in Geneva, your question is simpler. More agent calls are coming, and the caller might be a synthetic voice, a contractor reading from a request, or the actual customer. You won't always be able to tell which.

A few habits help no matter who is on the line:

• Answer the way you would for any first-time caller: hours, price ranges, the next open slot.
• Don't read out card numbers, account details, or anything about another customer to a caller you can't verify.
• Confirm bookings directly with the customer by text or email, so a garbled agent call doesn't leave a ghost appointment on your calendar.
• Put prices, hours, and a booking link online, where an agent can find them without calling at all.

If you're shopping for an AI receptionist or AI calling tool yourself, ask the vendor one direct question: when the AI can't finish a call, who takes over, and does the person on the other end know? A good answer names the human fallback and says it's disclosed. A vague answer usually means nobody at the vendor has thought it through.

Our free process audit at New Face Design looks at how calls and bookings move through your business today, and which parts software can handle well versus which still need a person.

Fox said phone calling was one of the features Muse users asked for most. Meta's own test suggests that, for now, getting those calls right still takes people.