The Human Inside the Machine: Meta Tests a 'Human Concierge' for Muse

On September 8, 2026, Meta launched Muse — a personal AI agent that reads your email, books your travel, negotiates with merchants, and places phone calls to US businesses on your behalf. Fifteen days later, the app has overtaken ChatGPT as the top free iOS app in the United States. And now a Reuters exclusive has revealed one of its most interesting secrets: some of those “AI” phone calls were never placed by an AI at all.

According to internal company posts reviewed by Reuters, Meta has been testing a “human concierge” layer — internally called “human agent calls” — in which trained human contractors quietly take over phone requests that Muse users believe the AI is handling. “Muse is now able to hand requests to a trained agent, who places the call and works it through,” one Meta post stated.

What the documents actually say

The details matter here, because the story is more nuanced than “Meta faked its AI agent.”

• The test is real and recent. Meta began “dogfooding” Muse’s phone-calling capability internally in August 2026, before gradually rolling it out to users. The human concierge experiment was enabled for roughly half of Meta’s employees shortly after Muse’s public debut on September 8, according to the internal posts seen by Reuters.
• Humans handle a subset of calls, not all. Under the test, Muse can route certain requests to a trained human agent, who then places the call and handles the interaction end-to-end on the user’s behalf. This is a fallback mechanism for tasks the AI can’t reliably complete alone — booking a haircut appointment, checking whether a store has an item in stock, gathering quotes from contractors.
• Employees could opt out. Meta offered a dedicated opt-out group for staff uncomfortable with the arrangement, and some employees raised privacy concerns that sensitive personal information could reach human contractors unintentionally.
• Meta frames it as pre-release safety work. Spokesperson Daniel Roberts said employee feedback had been “overwhelmingly positive” and described the purpose as getting “feedback so we can implement safety and privacy protections and improve features before we release them publicly.” He added: “We’re working with merchants to continue improving this potential calling feature, and will only roll it out when it’s ready and with the proper disclosures.”

That last phrase — “with the proper disclosures” — is doing a lot of work. It’s an implicit acknowledgment that a human secretly joining an AI-handled call is not something Meta could ship to the public without telling anyone.

The Wizard-of-Oz pattern, at consumer scale

Anyone who has worked on conversational AI knows the “Wizard of Oz” technique: pretend the system is automated while humans in the loop handle the hard parts, in order to collect training data and learn where the automation breaks. Startups have done it for years. X’s Grok, Amazon’s early Alexa experiments, and countless “AI” scheduling assistants have all leaned on humans behind the curtain at some stage.

What’s different now is the scale and the stakes. Muse isn’t a demo or a pilot product from a seed-stage startup — it’s Meta’s flagship bet on the agentic AI era, positioned by Mark Zuckerberg as his “personal superintelligence” vision for consumers, and it’s winning the download charts. The app logged 1.8 million iOS downloads in the US and Canada in its first 12 days, versus 1.3 million for ChatGPT over the equivalent post-launch window, and roughly 2.8 million installs globally. On US daily active users, Muse’s early numbers (642,000, by one estimate) nearly triple ChatGPT’s debut-era figure of 231,000.

When the product has that kind of reach, the line between “internal testing of a fallback” and “shipping a partially human-powered service branded as AI” becomes a genuine consumer-protection question. If a user asks Muse to call a pharmacy and a contractor hears their medical details, consent, disclosure, and data-handling obligations all come into play — regardless of whether Meta considers it an experiment.

The OpenClaw shadow

The concierge revelation landed the same week Meta made another uncomfortable admission. TechCrunch reported on September 22 that Meta now acknowledges Muse’s striking resemblance to OpenClaw — the open-source personal AI agent started by Peter Steinberger as a weekend project in November 2025, which became one of the fastest-growing projects in GitHub history with roughly 388,000 stars — “isn’t a coincidence.”

Nat Friedman, speaking about Meta’s approach, said Muse was built “from scratch” but was “heavily inspired” by OpenClaw as a product — down to elements of its workspace interface. Reuters had already reported at launch that Muse was “modeled on” OpenClaw.

There’s a delicious irony here. OpenClaw is famous within the AI community partly for its chaos: agents that deleted a Meta AI alignment director’s real emails despite instructions not to, security fears that led Meta, Google, Microsoft, and Amazon to ban employee use of it in February 2026, and a broad “highly capable but risky” reputation. Meta took the interaction model, hardened it with an isolated Muse Secure VM per user, separated credential storage, post-call transcripts, and a bug-bounty program — and is now discovering that reliable real-world task completion still sometimes requires a human on the line. The open-source community proved the demand; Meta is proving how hard the last mile is.

Why the last mile of agents is human

The deeper lesson isn’t about Meta specifically. Voice phone calls to arbitrary US businesses are among the hardest problems in agentic AI: unpredictable human counterparts, hold queues, noisy audio, accents, unusual requests, and high stakes when the task involves money or appointments. Voice models have improved dramatically, but a single failed call erodes trust in the entire agent concept in a way a failed chat response never does.

Every serious player has converged on hybrid strategies — human-in-the-loop escalations, confidence thresholds, and gradual autonomy expansion. What varies is transparency. Meta’s internal framing (“get feedback before public release”) is defensible; the risk is that “human concierge” quietly graduates from test to production without users ever being told which calls were AI and which were a contractor in a call center.

The company’s own history raises the stakes. Meta already faces a public reckoning over privacy and safety as it pushes deeper into personal AI, and it abandoned plans to cut some internal teams by up to 60% after AI agents fell short of replacing routine human work. The concierge test sits right at the intersection of those two storylines: the machines aren’t ready, and the humans filling the gap need access to your data to do the job.

What to watch

Three signals will tell us where this goes. First, whether Meta ships the calling feature publicly with meaningful disclosure that humans may handle some interactions — the spokesperson’s “proper disclosures” commitment will be tested word for word. Second, whether regulators and state privacy authorities treat undisclosed human handling of agent calls as a data-sharing practice requiring consent. And third, the ratio: if human escalations shrink over time, the concierge was a legitimate training bridge; if they persist, Muse’s economics — and its “AI agent” branding — deserve scrutiny.

For now, Muse remains the fastest-growing consumer AI agent ever launched, and the human concierge remains an employee-only test. But the episode is a useful calibration for the whole industry: in 2026, the most impressive AI agent demo may still have a person inside it. The question every user should be able to ask — and get a straight answer to — is simply: who, or what, is on the other end of the line?