The company let trained contractors handle phone calls on behalf of its AI agent, but internal worries about sensitive data leakage forced a rethink.
Meta’s new AI assistant can already book your dinner reservation and manage your inbox. Now the company wants it to pick up the phone and talk to actual humans on your behalf, with a twist: sometimes the voice on the other end of Meta’s call would be a real person pretending to be an AI. Or, more precisely, a human stepping in when the AI can’t handle the conversation alone.
The company has been internally testing a “human concierge” feature for its Muse personal AI agent, allowing trained contractors to take over specific outbound phone calls that the AI initiates or struggles to complete. The test rolled out to roughly half of Meta’s workforce around mid-September 2026, shortly after Muse’s phone-calling capability went live. But the experiment didn’t last long in its original form. Privacy concerns prompted Meta to pause the trial.
What the human concierge actually does
Muse can autonomously handle a range of tasks, from managing emails to shopping to making bookings. Its phone-calling feature, which began internal testing back in August 2026, extends that capability into voice conversations with real people, like restaurant hosts, doctor’s offices, or customer service lines.
The human concierge layer exists for scenarios where the AI hits a wall. When a call goes sideways or requires judgment the model can’t provide, a trained human contractor can step in and finish the job.
The feature was made available to about 50% of Meta’s employees during the trial period. According to Meta spokesperson Daniel Roberts, employee feedback was “overwhelmingly positive.”
Why Meta hit the brakes
When an AI assistant makes a phone call on your behalf, it potentially handles sensitive personal data: your name, your schedule, your medical appointments, your credit card details. When a human contractor enters that loop, the privacy calculus changes significantly.
That’s exactly the concern that surfaced internally at Meta. The worry centered on sensitive information leakage, the possibility that human contractors handling escalated calls could be exposed to data they shouldn’t see, or that the handoff between AI and human wasn’t sufficiently transparent to users.
Meta decided to pause the test rather than push forward, opting to review and strengthen its privacy protections before any broader rollout. The Muse application already operates within a secure virtual machine equipped with privacy safeguards, but the human concierge layer introduces complications that a sandboxed VM alone can’t solve.
The company says it remains committed to ensuring full transparency and user consent for sensitive interactions.
Muse’s broader momentum
The concierge pause comes at an otherwise strong moment for Muse. The AI agent launched publicly around September 8, 2026, across iOS, Android, and WhatsApp. It racked up over 2.5 million downloads in the US shortly after release, climbing to the top of app download charts.
