Meta Quietly Tested Humans Behind Muse's AI Calls — Then Pulled the Plug

Meta tested human contractors handling Muse phone calls for half its employees last week — before quietly rolling back the feature after internal privacy concerns surfaced.

What the test actually was

Meta enabled a "human concierge" feature — where human agents, rather than AI, handled Muse phone calls — for half of its employees last week. The idea was to see whether human-handled calls could outperform AI-only call handling for tasks Muse's agent takes on for users.

The privacy concern that ended it

Employees participating in the test warned that sensitive information could be unintentionally shared with call-center contractors handling the calls — a red flag inside a company that has spent years building trust (and defending scrutiny) around how it handles user data. A Superintelligence Labs vice president called the initial test "a miss" and confirmed Meta rolled the feature back.

The awkward part: humans actually worked better

The rollback comes with a complication. Meta said human-made calls achieved a 95% to 98% success rate, notably higher than the rate for AI-handled calls — meaning the company pulled back a feature that was, by its own numbers, outperforming the AI-only alternative it's built its product around. It's a reminder that even as Muse racks up adoption milestones (surpassing 2.5 million downloads in its first two weeks), some of the hardest problems in building a truly autonomous AI agent are still unsolved.