Meta's Muse: the human concierge behind the AI agent

The phone call your AI agent places for you is, in some cases, placed by a person. A Reuters exclusive by Katie Paul, September 22, 2026: Meta is testing a "human concierge" for Muse, its new personal AI agent. Internally the program is called "human agent calls," and human contractors quietly handle some of the phone calls the digital agent appears to make. The test is enabled for half of Meta's employees. The app they are dogfooding launched September 8, crossed 2.5M downloads in two weeks, and sits at #1 free on the US App Store, ahead of ChatGPT.

Put those numbers next to each other and this stops being an internal HR curiosity. A mass-market consumer agent, the fastest-downloaded app of its category, leans on human labor for part of its signature capability, and the company's own employees heard about it roughly a week before the press did.

• The news: Reuters reports contractors quietly place some Muse phone calls, under the internal name "human agent calls." Scope: half of Meta's employees.
• The framing: Meta calls the internal response "overwhelmingly positive" and says the purpose is to "get feedback." Improvement, not concealment, is the claim.
• The open question: Reuters does not say whether the in-app summary and transcript disclose that a contractor, not the model, placed the call.
• The precedent: Google's 2018 Duplex demos drew the same question when recordings revealed a human voice behind the assistant.
• The structural point: an open stack's fallback is visible by construction, because the operator is the human watching. A closed product has to choose to tell you.

what reuters reported

Timeline first. Muse's phone-calling started as internal dogfooding in August and went public alongside the app's September 8 debut. Roughly a week before the Reuters story ran, shortly after that public launch, Meta told its employees about the concierge test. The scope: the feature was enabled for half of Meta's employees. The internal post read, "Muse is now able to hand requests to a trained agent, who places the call and works it through."

The reaction inside was not uniform praise. Meta's own employees raised confidentiality risks: contractors hear what gets said on calls placed in a user's name, including potentially sensitive information. A Meta spokesperson answered that employee response was "overwhelmingly positive" and framed the program's purpose as getting feedback, the vocabulary of improvement rather than disclosure.

Read the internal post's wording again: "a trained agent." The sentence is accurate and slippery at once, because agent is exactly the word the industry uses for the software. Here it means a person.

how the calling works

The capability is easy to describe. Muse dials US businesses: booking haircuts, checking whether a store has stock, getting quotes from contractors. After the call, the agent provides a summary and a transcript in the Muse app. Dogfooding began in August, and a gradual rollout to users followed the app's September 8 debut. Two weeks later: 2.5M downloads, #1 free on the US App Store, ahead of ChatGPT.

The reason a concierge exists is reliability. Businesses hang up on robots, negotiations stall, models mishear, and a task that ends in "sorry, nobody answered" is a task the product failed. The engineering answer, for now, is to hand the fumbled calls to a person who can work them through. That is a legitimate pattern; every mature automation stack has a fallback lane. The controversy lives entirely in one word of the reporting: quietly.

the question the transcript may or may not answer

Here is the part the reporting leaves open, stated as precisely as the facts allow: Reuters does not say whether the in-app summary and transcript tell the user that a contractor, not the model, placed the call. Nothing in the story settles whether disclosure is part of the product.

Both possibilities are live. Meta may label contractor-assisted calls in the app, in which case the quiet is a launch-communication artifact and the fix is documentation. Meta may not, in which case every transcript that reads as machine work is a small miseducation: the user learns that the model booked the contractor quote, when what they actually experienced is the model plus a trained human. The difference matters because trust calibration, the thing that decides whether someone hands an agent a medical booking or a salary negotiation, is built from exactly these transcripts.

Meta's employees named the sharpest version of the risk before the public did. A call placed in your name is a conversation you are not a party to, conducted by someone you did not choose, and the contractor hears all of it, including whatever sensitive information the task requires.

this road has been walked before

One precedent paragraph, offered as history rather than accusation. Google's 2018 Duplex demos drew the same question when recordings revealed a human voice booking appointments behind the AI assistant. The crowd that marveled at the uncanny booking call and the crowd that asked who exactly made it were hearing the same demo. And the pattern is older than agents by decades: automation has always had people somewhere behind the curtain, from telephone operators to the monitored-call lane. The recurring failure is not that humans are in the loop. It is that users find out by accident.

closed products choose, open stacks reveal

The thesis, stated without consolation prizes: the reliability frontier of consumer agents is part-human right now, and that is not a scandal. It is the honest state of the art. The trust question lives in disclosure, and there the two ways of building agents diverge structurally.

In an open stack, the kind rotative runs on, the fallback is visible by construction. The operator is the human in the loop: every handoff, every action, every call lands somewhere the user can read it, because visibility is the architecture rather than a feature promised on a settings page. Nobody has to remember to tell you a person was involved. You watched the whole time, or you can.

A closed product has no such forcing function. It has to choose to disclose, in the transcript, in a policy, in a toggle, and each of those choices stays optional until someone makes it. That asymmetry is bigger than Meta: every company shipping or planning agent phone calls is making the same choice, quietly or otherwise. And the population inside the experiment already numbers in the millions. 2.5M downloads in two weeks means 2.5M people are running this trust test whether or not they know its shape.

what would resolve it

Three things to watch, none of them predictions. First, a disclosure statement: if Meta says contractor-assisted calls are labeled in the app, the open question closes and the story shrinks to a launch communication miss. Second, scope: whether the test moves from half of employees toward the gradual user rollout, and whether disclosure ships with the expansion or trails it. Third, whether anyone with authority asks the question out loud. The rules about who may hear a call placed in your name predate agents by a long stretch, and they were written for exactly this shape of situation.

the fine print

The reporting here is Reuters's (Katie Paul, September 22, 2026), built on internal Meta posts, employee accounts, and company statements; we did not independently verify the internal post or the deployment scope. "Half of Meta's employees" is an employee deployment scope, not a user count, and employee dogfooding is not general availability. The disclosure question is genuinely open: this piece does not claim Muse hides contractor involvement, only that the reporting does not settle it. Download and ranking figures are as reported on September 22; app store ranks move daily, and downloads are not the same thing as active users. Muse calling covers US businesses today, and a test can change shape or scope without notice. The Duplex comparison is history, cited as pattern rather than accusation. The open-stack contrast reflects how rotative operates its own tooling, which we can verify directly, because we run it.