Meta Tests Human Concierge for Muse: When AI Phone Calls Still Need Humans

1. The Promise vs. The Reality of Autonomous AI Telephony

Ever since Google stunned the tech industry with its original Duplex demo in 2018, Silicon Valley has chased the holy grail of conversational agency: a digital assistant that can seamlessly dial up a neighborhood bistro, negotiate an appointment slot, or call an airline support line on your behalf.

With Meta’s recent launch of Muse—the consumer AI assistant built by the newly formed Meta Superintelligence Labs (MSL) headed by Nat Friedman and Alexandr Wang—the dream seemed closer than ever. Muse quickly soared to No. 1 on the U.S. App Store. Yet behind the slick interface, internal testing revealed a stubborn obstacle: real humans at businesses hate talking to AI bots.

When businesses detected synthetic speech pauses, unnatural prosody, or the mandatory automated disclosure ("I am an AI assistant calling on behalf of..."), receptionists and counter staff routinely hung up. Internal telemetry showed bot-to-human call completion rates hovering below 60%.

2. Enter the "Human Concierge": Why Meta Turned to Contractors

To solve the hang-up problem, Meta engineers rolled out an unpublicized feature internally termed "Human Concierge." Under this model, when a user instructed Muse to perform a phone-based errand—such as checking inventory at a local boutique, scheduling a dental visit, or getting a price estimate—the task could be silently routed to a human contractor:

3. The Internal Privacy Backlash: "Who Is Listening to My Calls?"

The feature was enabled for approximately half of Meta’s workforce as an internal dogfooding test. However, because Meta did not explicitly label which calls were handled by software and which were transferred to live humans, employees began sharing alarming discoveries on Meta’s internal Workplace message boards.

Contractors were not merely booking restaurant tables; they were fielding sensitive personal tasks. When employees instructed Muse to negotiate medical bills, dispute credit card charges, or inquire about sensitive personal purchases, contract workers were handed the full prompt text and contact information.

The boiling point occurred when an employee tasked Muse with calling a service provider to resolve a disputed bill. Rather than an AI handling the negotiation, a contractor took the call and allegedly made an offensive racist remark to the service representative—an interaction that bounced directly back to the employee's personal account.

4. Historical Echoes: The Ghost of Facebook M

For veteran AI observers, the Muse concierge incident was an eerie case of history repeating itself. In 2015, Meta (then Facebook) introduced M, a high-touch virtual assistant embedded inside Messenger that could buy gifts, plan travel, and negotiate deliveries.

Facebook M was famously powered by a massive army of human trainers working out of offices in Menlo Park and contractor hubs. The theory was that humans would train the AI until the software could take over. But the long tail of human requests proved impossible to automate, and Facebook quietly killed M in early 2018.

Eight years later, even with state-of-the-art transformer reasoning, speech-to-speech diffusion latents, and sub-second voice synthesis, the messy reality of the physical telephone network remains deeply resistant to pure automation.

5. Current Status: Rollback and the Path Forward

In response to the internal firestorm, Meta leadership pulled the plug on the human concierge test. A company spokesperson confirmed that rolling out the test without clear notification to users was a mistake:

“We tested an experimental concierge feature internally to help improve completion rates on difficult voice interactions. Launching this pilot without clear user-facing disclosures was a miss, and we have disabled the test while we refine our approach to automated agent capabilities.”

Meta emphasized that its long-term ambition remains 100% autonomous speech-to-speech agents. However, the Muse concierge episode underscores an uncomfortable truth for consumer AI: as models attempt to cross the chasm from digital chat screens into real-world social infrastructure, human fallbacks remain a seductive—yet ethically perilous—shortcut.