Meta Tests Human Callers for Muse AI Agent

Meta is currently experimenting with a hybrid approach for its Muse AI agent, introducing human concierges to handle specific phone calls initiated by the system. According to an exclusive report by Reuters, which cited internal company postings, Meta has been testing this model where contractors intervene in some of the calls placed through the agent. This trial was communicated to staff last week, following the initial launch of Muse’s calling capabilities.

The experiment involved enabling the ‘human agent calls’ feature for 50% of Meta’s employees, who were given the option to opt out if they preferred not to participate. Internal communications explained that “Muse is now able to hand requests to a trained agent, who places the call and works it through.” However, the initiative faced scrutiny as some employees raised concerns that sensitive user data might inadvertently be exposed to these call-center contractors, even though such exposure was not the intended outcome.

Despite these privacy concerns, Daniel Roberts, a Meta spokesperson, told Reuters that employee feedback had been “overwhelmingly positive” regarding the test. He emphasized that the primary goal was to refine safety and privacy protections before any public release. Roberts stated that Meta would “only roll it out when it’s ready and with the proper disclosures,” although he did not clarify what those disclosures would entail or which stakeholders would receive them.

Muse debuted in the United States on September 8 as a digital assistant designed to send emails, make purchases, and book travel on behalf of users. Its new calling feature forwards requests to businesses to facilitate tasks such as booking haircuts, checking inventory, or obtaining quotes, subsequently providing a transcript of the interaction. The app has already garnered significant traction, reaching 2.5 million downloads according to Sensor Tower data.

At launch, Meta heavily emphasized its commitment to privacy, noting that each agent operates within its own isolated virtual machine. While this technical safeguard addresses software security, it does not account for scenarios involving human operators using headsets. This is not the first time Meta has navigated this intersection of automation and human labor. In 2015, Facebook tested M, a Messenger assistant supported by human trainers. That project remained in limited beta and was shut down in 2018 because it failed to exceed approximately 30% automation rates. Today, Meta positions Muse as a round-the-clock agent that may occasionally rely on human assistance for complex calls.