Meta quietly ran a test in which call-center contractors placed some of the phone calls requested through its Muse AI agent, according to internal company posts seen by Reuters. The trial, enabled last week for half of Meta's employees, has since been rolled back.

Muse, launched on September 8, is built to send emails, shop, and book travel on its own. Its calling feature dials U.S. businesses to schedule haircuts, check whether a store has an item in stock, or obtain quotes from contractors. Meta began testing calling with employees in August and gradually extended it to Muse users after the app's launch. The episode zeroes in on a question that follows any agent-style product: when an AI hands a task to a person behind the scenes, who on the other end of the line gets told?

Human agents lifted call success rates to 95% to 98%

The AI-only calling feature had rough edges. Some businesses hung up on the agent, and one employee said an insurance company repeatedly ended the call once it learned Muse was an AI.

Last week, Meta enabled "human agent calls" for half of its. Staff could join an opt-out group, and Muse could then pass a request to a trained agent, who placed the call and worked it through. A vice president at Meta Superintelligence Labs told employees that tests showed the success rate of human-placed calls could be boosted to 95% to 98%, a level AI-only calling had not reached, she said. For the businesses receiving those calls, there was no signal that a person rather than the agent was on the line — the disclosure gap the vice president later conceded was a miss.

Privacy concerns and a racist remark preceded the rollback

Employees warned that sensitive information users provide to Muse could be leaked to call-center contractors. Some pointed to Meta's launch pitch that each Muse agent runs in its own isolated virtual machine. One employee who had Muse negotiate a cable bill said the call transcript showed the contractor made a racist remark. For a product that asks users to delegate email, shopping, and travel, the test showed how quickly trust questions extend beyond the AI itself to the people and vendors in its workflow.

The vice president acknowledged it was a miss to start contractor-placed calls without proper disclosures and said Meta had rolled the feature back for now. The contractor in question, she said, would not work on Meta projects again.

Meta spokesperson Daniel Roberts said the test was intended to gather feedback so the company can build in protections for safety and privacy before any public release. "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," Roberts said.

Muse has surpassed 2.5 million downloads. Since its launch, Meta's stock has climbed more than 20%, adding more than $200 billion to its market value. Whether the calling feature returns — and under what disclosures for businesses and users — now stands as the near-term test of how Meta ships an agent designed to act on people's behalf.