Meta has been testing a human concierge for its new personal AI assistant, Muse, allowing human contractors to handle some phone calls that users ask the digital agent to make, according to internal company posts reviewed by Reuters.
The test came shortly after Meta launched Muse’s phone-calling feature, which allows users to instruct the AI assistant to contact US businesses for tasks such as booking appointments, checking product availability and requesting contractor quotes.
Muse is designed to perform tasks on behalf of users, including sending emails, shopping and booking travel. The assistant has quickly gained traction in the US, recording more than 2.5 million downloads since its September 8 launch, according to Sensor Tower data cited by Reuters.
Under the Muse AI human concierge test, the system could hand a request to a trained human agent who would place the call and complete the interaction. Meta enabled the feature for about half of its employees, with an opt-out option available, according to internal posts.
Mitchell Starc Says South Africa Tour Could Bring Testing Conditions
The arrangement prompted privacy concerns among some employees, who warned that sensitive information could potentially be exposed to call-centre contractors.
One employee questioned the risk of allowing humans to become involved in calls that users expected an AI system to handle. Another employee reported that a transcript of a call to an internet and cable provider contained a racist reference allegedly made by the human contractor handling the conversation.
A Meta vice president acknowledged that launching the contractor-based calls without adequate disclosures had been a mistake and said the company had rolled back the feature for the time being.
The executive said internal testing suggested human agents could raise the success rate of phone calls to between 95% and 98%, compared with a lower success rate for AI-only calling.
Meta spokesperson Daniel Roberts said employee feedback on the experiment had been “overwhelmingly positive” and that the test was intended to help the company develop safety and privacy protections before any broader release.
“We’re working with merchants to continue improving this potential calling feature,” Roberts said, adding that Meta would only introduce it more broadly when it was ready and with appropriate disclosures.
Privacy challenge for autonomous AI
The Muse AI human concierge experiment highlights one of the challenges facing companies developing autonomous AI agents: systems may need human intervention when automated interactions fail to work reliably.
Meta began testing Muse’s phone-calling capability internally in August before gradually expanding it to users following the assistant’s launch. Some employees reported that businesses had ended calls after discovering they were speaking with an AI system.
The human fallback was intended to address some of those difficulties, but employees said it could conflict with Meta’s privacy assurances.
Meta had emphasised security protections when launching Muse, including a dedicated virtual machine for each agent session and separate secure storage for sensitive information such as passwords. Employees raising concerns about human callers argued that contractor involvement could undermine those protections.
The situation also recalls Facebook’s earlier M digital assistant experiment, which relied heavily on human workers to complete tasks that the system could not handle autonomously.
For now, Meta has rolled back the human concierge test while continuing to develop its AI calling technology. The company says any future wider rollout will include appropriate disclosures and additional safety and privacy measures.
