Muse Called Your Hair Salon. Nobody Told the Salon It Was an AI.

"Hi, I'd like to book a table for two on Friday at seven." A receptionist writes it down, name, time, party of two, then asks about dietary restrictions while a polite voice answers every question without hesitation, and nobody in the room knows the caller is not a person, because nobody said so, and the booking goes into the book exactly as if a human had made it.

Since September 16, 2026, that call may have come from Muse. Meta opened a public beta that lets its personal AI agent phone US businesses on a user's behalf, to book restaurant tables, move hair appointments, check whether a store has something in stock, and argue with a cable company. You ask in the Muse app or in WhatsApp, Muse dials, and afterwards you get a transcript and a summary, which means you never hear the call yourself.

What you also never hear is whether Muse introduces itself. Meta has not published the script its calls use, has not said whether a Muse call always discloses that it is AI, has not said which number shows up on caller ID, and has not published any way for a business to opt out: no directory, no exclusion portal, nothing. That silence is the story. Silence is a choice., because the law on this question is not silent at all. It has been written, in several jurisdictions, and it has been on the books for years.

A ruling from before Muse existed

On February 8, 2024, the Federal Communications Commission voted unanimously on a question that sounds, in retrospect, like it was drafted for exactly this product. In Declaratory Ruling FCC 24-17, the commission confirmed that the Telephone Consumer Protection Act's restrictions on "artificial or prerecorded voice" cover current AI technologies that generate human voices, including voice clones of real people. AI voice calls are not banned; existing rules apply to them, and the most important of those rules is consent: a caller must obtain the prior express consent of the called party before placing a call that uses an artificial voice, and a telemarketing call needs prior express written consent.

Identification duties apply too, because under 47 CFR 64.1200(b), artificial-voice messages must identify the responsible entity at the start of the call and offer an opt-out. Statutory damages run $500 per call for a negligent violation and $1,500 per call for a willful one. A separate FCC proposal from August 2024 would have required an AI disclosure at the start of every such call. As of October 2026, that proposal has not become a final rule.

Read the consent rule again, slowly, because this is where the whole arrangement comes apart. Read it again. Consent must come from the called party, and in a Muse call the person who consented is you, the user who asked Muse to dial, while the called party is a receptionist in Ohio who consented to nothing. Nobody on the receiving end consented. That is the gap. Nobody in the chain holds the consent the statute describes, and the statute was written before anyone imagined a consumer delegating the dialing to software. That is not a loophole; it is a category the law has not met yet, and it is sitting in your phone's call log.

Three agents, three answers

Muse is not the only agent with a phone. Google's agentic calling, launched as "Ask for Me" in January 2025, opens its calls by announcing an automated system calling from Google on behalf of a user, according to press reports. Instinct Concierge launched the same day as Muse's beta, September 16, 2026. Neither Meta nor Instinct has published official documentation stating whether or how their outbound calls announce AI usage. Here is the whole comparison, and it fits in a table because there is so little to compare.

One row deserves a second look. Early in testing, Meta used human call-center contractors to place some Muse calls instead of synthetic voices, and internal tests reportedly found the human-placed calls succeeded 95 to 98 percent of the time while AI-only calls did much worse. Meta rolled the human calls back, which is worth thinking about, because businesses were hanging up on AI voices often enough that the product needed humans to get through. The fix was not to tell businesses they were talking to an AI more clearly. It was to send a human instead, and then, once the humans were withdrawn, to let the AI try again on its own, as if the hang-ups had been a staffing problem rather than a message from the people being called.

States and Europe already answered

While the FCC's disclosure proposal sits unfinished, the states have been busy. Utah passed the first state law regulating private-sector generative AI in March 2024, the Artificial Intelligence Policy Act (SB 149), effective May 1, 2024 and amended in May 2025. Its rule is narrow but pointed: if a consumer clearly and unambiguously asks whether they are interacting with AI, the supplier must say so, and people in regulated occupations must disclose proactively and prominently in high-risk interactions. There is a safe harbor: disclose clearly and conspicuously at the outset and the disclosure claims fall away. Fines run up to $2,500 per violation, enforced by Utah's Division of Consumer Protection. Utah also bars the one defense every AI company reaches for first, the argument that the model did it and the company merely built it, by stating plainly that you cannot blame the AI itself to escape liability for what the AI did on your instructions.

California's AB 2905, effective January 1, 2025, amends Public Utilities Code section 2874 so that when a call is placed using an automatic dialing-announcing device, the required announcement must disclose whether the prerecorded message uses an AI-generated voice. Texas SB 140, effective September 1, 2025, requires disclosure that a call is AI-generated within the first 30 seconds. California's older bot law, SB 1001 from 2019, already requires clear and conspicuous disclosure where a bot is used with intent to mislead about its artificial identity in a commercial transaction. Across the Atlantic, Article 50 of the EU AI Act, applicable since August 2, 2026, requires that a person interacting with an AI system be informed they are doing so, unless it is obvious from the circumstances.

None of these laws was written with a personal agent in mind, and every one of them reaches one. A receptionist in Utah who asks "am I talking to a robot?" is owed an honest answer under Utah law. A business in Texas is owed disclosure inside thirty seconds. A business in California is owed it in the opening announcement. Whether Muse's calls satisfy any of that is, at this writing, unknowable from anything Meta has published, which is a strange position for a company that has had since February 2024 to read the FCC's ruling, employ the lawyers who wrote the comments on it, and publish a single paragraph saying what its product does when it picks up the phone.

Work the numbers

Liability math is where abstract rules turn into a decision. Do the math. TCPA damages are $500 per call for a negligent violation and $1,500 for a willful one. Suppose a heavy user lets Muse make one undisclosed business call a week for a year, fifty-two calls in all. If a court ever treated those as violations, exposure runs $26,000 at the negligent rate and $78,000 at the willful rate. That is an illustration, not a prediction, since no court has ruled on a personal-agent business call, no plaintiff has brought one, and the TCPA's consent framework was built for robocallers, not for a homeowner asking software to move a haircut. But the arithmetic matters because the per-call structure means risk scales with volume, and volume is precisely what delegation is for. An agent that makes a hundred calls a month does not have a hundred times the convenience of one call. It has a hundred times the exposure.

The other side, at full strength

Fairness demands the other side, stated at full strength. A restaurant taking a reservation does not care whether the voice on the line is human or synthetic as long as the name, the time, and the party size are right, and a disclosure script that begins "hello, I am an artificial intelligence" mostly produces hang-ups that punish the user for delegating a boring errand. Google announces itself at the start of every call and its calls still get completed, which undercuts the only practical argument against disclosure, namely that honesty would kill the product before it ever got to say hello. If disclosure is harmless to the outcome, though, the case for withholding it collapses to convenience, and convenience is a thin reason to leave the person on the other end of the line uninformed about who, or what, they are dealing with.

Hang-ups are data, because when businesses hang up on AI voices they are telling you something about consent that no product review has yet answered, and routing around that signal with a human contractor is not a fix. It is a confession that the product knows what the answer would be.

One more distinction, because it will confuse readers otherwise. On November 5, 2026, Meta rolled out live voice conversations with an animated avatar inside the Muse app, so you can talk to Muse the way you would call a person. That is a different feature: you are talking to your own agent, nobody else is on the line, and nobody is being misled. The calls in this article go the other direction, out from your agent to a stranger who did not ask for them.

What I could not verify

Honesty requires saying what this piece did not prove. Meta has not published its call script, its caller ID arrangement, or any opt-out process, so absence of publication is not proof of absence of disclosure. Muse may already announce itself on every call, in which case this article's central complaint evaporates and the fix is a single paragraph on a help page. We cannot know from public sources, and that unknowability is itself the finding. No court has ruled on TCPA liability for a personal AI agent's business call, so the liability analysis here is reasoning from the statute, not case law, and the $26,000 and $78,000 figures are illustrative arithmetic, not a forecast of any real exposure. The 95 to 98 percent human-call success figure comes from internal tests reported secondhand, and I could not verify it against a primary source. Texas SB 140's thirty-second rule and California AB 2905's scope are summarized from secondary legal trackers; I did not read the enrolled bill text for either. And the FCC's August 2024 disclosure proposal remains pending as of October 2026, which means the federal disclosure rule that would settle this question still does not exist.

What to do before your first delegated call

Ask Muse to state in its opening line that it is an AI calling on your behalf, then read the transcript afterward and check whether it did. Trust the transcript, not the promise. If it did not, say so in the chat, because the transcript is your only record of what the other side heard. Never let an agent give payment details by voice, and confirm every booking in writing, because a reservation that exists only in a phone call is a reservation that can be disputed by either side. If you run a business and a Muse call arrives, treat it as a real customer request, confirm it in writing, and never take card numbers by voice. If a business hangs up on the AI voice, do not route around it by asking for a human, calling yourself, or switching to a different agent that hides its nature better. That hang-up is the market telling you the disclosure matters, and the right response is a clearer introduction, not a better disguise. And keep the volume low until Meta publishes the script. One call a week is a convenience, while a hundred calls a month is a compliance posture you adopted without reading the terms, and the terms, for once, were written before the product was.

Muse calling businesses for you is genuinely useful, and the machinery underneath it is not the problem. A polite voice that books a table while you cook dinner is a small, real piece of the future personal agents were promised to be. But a phone call is a social act before it is a technical one, and the person who picks up is entitled to know what kind of caller they have, whether that caller is a person with a name and a job or software with a script and a transcript that only its owner will ever read. Until Meta publishes how its calls introduce themselves, the safest assumption is the one the law already made: the caller should say so, at the start, in plain language. That is the rule. And if it does not, the risk is yours, measured at five hundred dollars a call.