The Voice at the Other End of the Line

The call came in on a Friday evening, during the rush. A host at a restaurant in Austin picked up, heard a woman ask for a table for four at 7:30, heard her mispronounce the name of the street the restaurant sits on, and hung up. It was the third time that week.

The voice paused in exactly the right places. That was how she knew.

The agent doing the calling belongs to Meta. Alexandr Wang, the company’s chief AI officer, said on Nov 13 that paying subscribers can now send Meta’s Muse assistant to make phone calls on their behalf inside the United States, handling reservations, appointments and the kind of queueing that consumes an afternoon.

The capability was tested with a small group in September. It is now open to subscribers across the country, at about $20 a month. The pitch is straightforward and slightly strange: the assistant that has spent two years answering questions will now speak to strangers for you, in a voice, over a phone line, to businesses that did not ask to be part of the arrangement.

Meta is not the first company to attempt this. Google demonstrated a reservation-making assistant years ago and was criticized for the ethics of a machine impersonating a person. OpenAI has built voice into its products without pushing it toward outbound tasks. What is different now is distribution: Meta can put the capability in front of billions of users through apps they already have open, and it can charge for it.

According to two people familiar with the rollout, the company is limiting the feature to U.S. subscribers at first and has built in a disclosure that identifies the caller as an assistant when the business asks. Those people said the disclosure is not automatic on every call, and that the design reflects a judgment that most businesses hang up on anything that sounds like a robot before the conversation starts.

TechCrunch reported that many merchants do exactly that. A restaurant that receives several automated calls a day learns to recognize the cadence, and the cadence is not hard to recognize. Synthetic speech has improved at the level of individual sentences and remains recognizable across a full conversation, particularly when the subject is a table for four and the background is a kitchen.

That is the technical problem dressed as a social one. An agent that can book a table is not an agent that can negotiate one, and the difference shows up in the failure cases. A human calling about a reservation can hear that the restaurant is fully committed and ask about Tuesday instead.

An agent working from a fixed instruction will ask about Tuesday only if it was told to. The result is a call that completes and a booking that does not happen.

The economics of the feature are less obvious than the price suggests. Twenty dollars a month is the same price as the subscription it sits inside, which means the calling capability is being used as a reason to keep paying rather than as a product with its own margin. Every call costs money to place and more to complete, and a subscriber who uses the feature heavily is a subscriber the company loses money on. People close to the product said the company expects usage to be occasional, which is what every company expects before it ships something convenient.

For the businesses on the receiving end, the calculation is different and worse. A reservation line exists to fill seats, and a call that fills a seat is welcome regardless of who makes it. A call that fills a seat with a party that never arrives is not, and automated booking has a well-documented problem with no-shows because the person who asked for the table was never the person who agreed to come.

Some restaurants have started screening. Others have moved reservations to platforms that require a card. A restaurant owner in Chicago said her staff now asks for a phone number and calls back before confirming, a workaround that costs more labor than the original booking saved. She said the automated calls have not stopped, they have just become someone else’s problem.

Meta’s bet is that the friction is temporary. If the agents get good enough to pass, businesses will stop screening, because screening costs money and accepting does not. That has happened before with other automated channels: email, then web forms, then chat widgets, each of which was resisted and then absorbed. The difference with voice is that the phone line is the last channel where a human is expected, and the expectation is doing real work in keeping the volume low.

There is also a legal edge. California requires bots to disclose that they are bots in certain commercial interactions, and several other states have drafted similar rules. A caller that identifies itself only on request may satisfy the letter of those statutes and not the intent, and the intent is that a person answering a phone should know what they are talking to.

Mr. Wang has described the feature as part of a broader shift in which assistants act rather than answer. That shift is real and it is arriving unevenly: agents can now complete purchases, file forms and place calls, and the institutions they interact with were designed on the assumption that a person was on the other end.

The adjustment will not be negotiated. It will happen one hang-up at a time.

The Austin host has stopped answering numbers she does not recognize before 5 p.m. Her reservation book is thinner than it was in September. Somewhere in Menlo Park, a dashboard shows a completion rate that looks healthy, because a call that ends without a booking still counts as a call that ended.

  • Related Posts

    • November 14, 2026
    • 51 views
    Anthropic Packages Claude for the Factory Floor

    In a machine shop that supplies aerospace brackets, the quoting process has looked the same for two decades. An engineer opens a spreadsheet, pulls costs for material and machine time,…

    • November 14, 2026
    • 50 views
    Chipmakers and Cloud Providers Form an Alliance Around SGLang

    At the SGLang summit on Nov. 12, the slide that mattered was the one with eight logos on it. LMSYS, the research group behind the Chatbot Arena model leaderboard, used…