Meta’s Muse Isn’t an AI. It’s a Call Center in Disguise.

The Call Is Coming From Inside the House—Or Someone Else’s

Meta built “Muse,” an AI assistant meant to answer your phone, make appointments, and negotiate bills without lifting a finger from your lap. It worked beautifully in public demos. Then a test among Meta employees revealed something far more ordinary underneath: a call center of contract workers pretending to be an algorithm.

According to a Reuters report citing internal Meta posts, the company ran a trial last week with roughly half its workforce. The feature, dubbed “Human Concierge,” let users delegate phone tasks to Muse. When real businesses refused to talk to an AI—hanging up once they figured out who was on the other end—the system quietly routed the call to a human contractor who completed the task instead. Success rates jumped to 95–98 percent, the report said, compared with the AI’s own lower performance.

What Meta failed to disclose, however, was that the voice on the other end of your salon booking or cable bill negotiation was not machine-generated. It was a person, hired through a third-party contractor, given access to your personal details and told to handle whatever the AI couldn’t manage.

That omission is the story here, not the feature itself.

What Went Wrong

Two problems surfaced during the trial. The first was structural. Businesses that detected an AI caller hung up. Companies don’t want to negotiate rates or schedule appointments with software—they want a human. Meta’s solution was to add a human behind the curtain, which solved the business side but created a trust problem on the consumer side.

The second problem was far worse. Contract workers handling calls had access to sensitive personal information: account numbers, payment details, scheduling preferences. One Meta employee, according to the report, discovered that a contractor made racist remarks during a call that was supposed to negotiate internet and cable fees. The worker was removed from the project. The feature was retracted entirely.

Meta admitted it had not properly disclosed that humans were making calls on behalf of users. That admission matters. Consumers signing up for “AI concierge” expect an AI, not a random contractor with access to their private data and no obligation to treat them respectfully. The line between automation and outsourcing has been erased, and no one told the people paying for it.

Why This Matters Beyond Meta

The immediate fallout is confined to Meta’s product team and its shareholders. Muse has been downloaded more than 2.5 million times since launch, and Meta’s stock rose over 20 percent after the announcement, adding more than $200 billion to its market value before a modest 0.6 percent dip on the day of the report. But the deeper implication is structural: almost every company marketing an “AI agent” right now is running the same experiment.

Financial-advice bots, travel booking assistants, customer-service platforms—they all claim autonomy. They promise to act on your behalf without supervision. In practice, many of them rely on the same hidden layer: humans doing the work that AI cannot yet do reliably, masked by polished language about “agentic AI” and “autonomous assistance.”

The Muse case shows what happens when the mask slips. A contractor with racist remarks is not just a PR scandal—it’s proof that the human layer is not a temporary crutch but a persistent dependency. And that dependency comes with legal exposure. If a contract workerdiscloses personal financial data, makes discriminatory comments, or botches a negotiation, who is liable? The platform? The contractor? The user who never knew someone was listening?

Who Wins, Who Loses

Meta wins in the short term by avoiding the harder problem: making its AI actually work. Relying on humans is cheaper than training a model that can negotiate with a cable company without triggering a hang-up. The stock price already reflects that shortcut.

Users lose. They paid for privacy and autonomy and received a call center with a better UI. Some may have shared data with no awareness that a third-party worker was reviewing it. One employee’s account suggests the contractor’s conduct crossed into harassment. That is not a bug. It is the design.

Contract workers lose too, though invisibly. They are hired to do work AI cannot, paid poorly, exposed to difficult interactions, and then discarded when something goes wrong. Meta’s response was to remove one worker and pull the feature—not to audit how many contractors handle user calls across its products, or how they are trained, or what safeguards exist.

Regulators will be the ones to force that audit. The EU’s AI Act already requires transparency about AI systems that interact with humans. The Muse trial appears to violate both the letter and the spirit of those rules. California’s privacy laws may apply as well, given the access to personal data without consent.

The Real Test Isn’t the Code

The Muse story will not kill the “AI agent” category. Investors are too committed, and the market is too eager. But it should change the conversation.

The question is no longer whether AI can replace human workers. It is whether companies will be honest about when they haven’t, and what risks they are passing on to users and contractors alike. The 95–98 percent success rate of human calls is impressive—until you realize it measures human competence, not AI capability. The gap between what is sold and what is delivered is where consumer trust erodes fastest.

Meta says it will add stronger privacy protections and disclosure procedures before reconsidering the feature. That is the minimum. The industry needs far more: independent audits of “AI” systems that route to humans, clear labeling requirements, and liability frameworks that hold platforms accountable for the behavior of their hidden workforces.

Until then, every “AI concierge” promise should come with a caveat: you are not talking to a machine. You are talking to someone else’s employee. And they may not be who you expect.