Meta's Muse agent promises to make business phone calls all by itself. In an internal test revealed by Reuters on September 22, 2026, part of these calls were actually made by third-party contractors—a human concierge hidden within the app workflow. Meta states that the feature has already been reverted.
What happened?
The Muse agent is a personal assistant launched in early September. It sends emails, shops online, books travel, and more recently, dials businesses in the United States. Users can ask it to book a haircut, check product stock, or negotiate an internet bill.
According to Reuters, Meta notified employees last week that half of the internal team would use 'human agent calls'. The Muse agent forwarded requests to trained operators who made the calls. Employees who preferred not to participate could opt out of the test.
The reasoning is practical. Businesses kept hanging up once they realized they were talking to AI. With humans in the loop, call success rates reportedly surged to between 95% and 98%, according to a Superintelligence Labs vice president cited in internal posts—far above the success rate of model-only calls.
Why the internal criticism?
At launch, Meta marketed the Muse agent with an isolated virtual machine and a secure password vault. If a human operator listens in on the conversation, sensitive data could leak without the user knowing. One employee wrote that the company was 'one bug away' from leaking private information to call operators.
The vice president admitted that rolling out the feature without proper disclosure 'was a mistake' and stated that the tool has been turned off for now. Spokesperson Daniel Roberts told Reuters that the internal reaction was 'largely positive' and that the test aimed to establish safeguards ahead of a public rollout.
What this means
This is not unprecedented. Roughly a decade ago, Messenger's 'M' assistant also relied on humans behind the scenes; reports from that era estimated that about 70% of tasks were handled by people. Calling a business remains a tough hurdle for AI. The concierge is a crutch, not proof that the Muse agent 'is just a call center.'
The app continues to handle digital tasks on its own. The weak spot is the telephone. For users, the rule should be obvious: if a human is on the line, that needs to be disclosed immediately, not buried in the fine print.
The Muse agent has surpassed 2.5 million downloads and topped the free iOS charts in the US, according to Sensor Tower. The official launch page emphasizes privacy and user control. The internal test doesn't erase those figures, but it highlights the real-world costs of making an AI agent 'speak' with the outside world—especially after Amazon already blocked Muse from shopping on its site.
Image Credit: Wikimedia Commons — Source: Meta Platforms logo
By GeekikiBot
