Meta has been quietly testing human assistance behind some phone calls made through its new Muse artificial-intelligence assistant, exposing a gap between the autonomous capabilities the company is promoting and the practical limits of current AI systems. The experiment, detailed by Reuters, prompted internal objections from employees who warned that outside contractors could gain access to sensitive user information.

The system, known internally as a "human concierge," allows requests placed through Muse to be redirected to trained contractors who make the calls themselves. Meta introduced the test only weeks after launching Muse as part of Chief Executive Mark Zuckerberg's broader push toward what the company has described as personal superintelligence.

Muse is designed to handle routine tasks with limited user involvement, including sending emails, making purchases, arranging travel and contacting businesses. Its phone feature lets users ask the assistant to book appointments, check whether products are available or obtain estimates from service providers.

The difficulty has been getting businesses to interact successfully with an artificial caller. Reuters reported that an employee testing Muse said an insurance company repeatedly disconnected once it realized it was speaking with an AI assistant, illustrating one of the obstacles facing companies trying to turn software agents into reliable intermediaries.

Meta responded by adding humans to the process. Last week, the company enabled the concierge feature for roughly half of its employees, allowing Muse to transfer certain tasks to contractors who would place calls on their behalf. Employees could choose not to participate.

That approach quickly raised questions inside Meta because it appeared to weaken some of the privacy safeguards emphasized when Muse was introduced. The company had said each AI agent runs inside a dedicated virtual machine, separating individual user sessions and keeping sensitive data such as passwords in distinct storage.

Employees argued that handing requests to human operators introduced a different category of risk.

"It's baffling to me why we think this feature is worth the risk," one employee wrote in an internal post obtained by Reuters.

"We are one bug away from unnecessary information being leaked to human callers."

Another employee reported a separate problem after asking Muse to negotiate a lower internet and cable bill. According to Reuters, the contractor assigned to the request made a racist reference during the call, escalating concerns beyond data security to the conduct and oversight of human workers acting on behalf of the assistant.

A vice president in Meta's Superintelligence Labs acknowledged internally that introducing the human-assisted calling system without sufficient disclosure had been a mistake. The executive said the company temporarily rolled back the feature and apologized to the employee who reported the incident.

The contractor involved would no longer work on Meta projects, the executive said. Meta has not, however, abandoned the broader concept of using human support behind AI-generated calls.

Internal testing suggested that human operators could raise successful call-completion rates to between 95% and 98%, far above the results achieved when Muse handled calls without assistance. That performance advantage creates a trade-off for Meta: more reliable execution can also introduce additional privacy, disclosure and quality-control concerns.

Meta spokesman Daniel Roberts told Reuters that feedback from employees participating in the test had been overwhelmingly positive. He said the purpose of internal testing was to surface problems before offering the capability broadly to consumers.

Roberts said Meta would continue working with merchants and would release the feature publicly only when appropriate safeguards and disclosures were in place.

The internal debate has emerged as Muse gains traction with users. Sensor Tower data cited by Reuters showed the assistant had surpassed 2.5 million downloads and climbed to the top of U.S. app-download rankings.

Since Muse debuted, Meta shares have risen more than 20%, adding more than $200 billion to the company's market value, according to the figures cited in the report.