Meta has been testing a “human concierge” for its new personal AI assistant, Muse, which entails having human contractors quietly handle some of the phone calls placed via the digital agent, according to internal company posts seen by Reuters.

The Facebook and Instagram owner told employees about the test last week, shortly after the company publicly launched a phone-calling feature for Muse.

The Muse agent, which has topped US app download charts in the two weeks since its debut, is designed to autonomously perform tasks like sending emails, shopping and booking travel on a person’s behalf.

With the phone-calling feature, a Muse user can tell the agent to dial phone numbers for US businesses and talk with people on the other end to handle errands like booking haircuts, checking whether a store has an item in stock or getting quotes from different contractors.

Some employees have raised privacy concerns about humans handling those calls, warning that the approach could result in sensitive information being shared unintentionally with contractors in call centers, the internal posts showed.

“It’s baffling to me why we think this feature is worth the risk,” one employee ⁠wrote in an internal post. “We are one bug away from unnecessary information being leaked to human callers.”

Another employee, who had asked Muse to call an internet and cable provider to negotiate his bill, said a transcript showed the human contractor had made a racist reference during the call.

In response, a vice president in Meta’s Superintelligence Labs unit acknowledged that “it was a miss” to start testing contractor-placed calls without proper disclosures and said the company had “rolled back this feature” for now.

Meta was still “hillclimbing AI calling,” she added in her post, noting that some tests indicated having humans make the calls could get their success rate up to the 95% to 98% range, instead of the “lower percentage of AI calling.”

She also apologized to the employee who reported the mention of race during a Muse call and promised that the “human agent” involved would not work on any Meta projects going forward.

Human labor behind AI promises

The Muse test offers a window into how tech companies sometimes rely on human labor behind the scenes to deliver on the promises of autonomous AI, as they bump up against reluctance to engage from those on the receiving end of robocalls.

It harks back to an earlier experiment Meta conducted a decade ago, when the company, then known as Facebook, trialed a digital assistant called “M” in its ⁠Messenger service. Media reports later estimated that around 70% of the agent’s tasks were actually carried out by humans.

Asked about the Muse “human concierge” plan, a Meta spokesperson said the response from employees had been “overwhelmingly positive” and that the purpose of the test was to “get feedback so we can implement safety and privacy protections and improve features before we release them publicly.”

“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,” said the spokesperson, Daniel Roberts.

Meta’s Muse agent has had more than 2.5 million downloads since its release, according to data provided by market intelligence firm Sensor Tower.

The company’s stock has soared more than 20% since ⁠the launch, adding more than $200 billion in market capitalization to touch its highest level in more than seven months as of last close.

Shares edged down 0.6% on Tuesday, however.

Undercutting a key selling point

Meta heavily emphasized the AI assistant’s safety protections during its launch, telling would-be users that each agent would have its own secure “virtual machine,” or a dedicated computer in the cloud designed to keep people’s sessions isolated, and would keep sensitive information like passwords in separate secure storage.

Employees ⁠critiquing the human concierge plan inside Meta cited those security measures in their feedback, warning that it risked undercutting a key selling point for the agent.

Meta began having employees test out the phone-calling capability in August and gradually started rolling it out to Muse users in the days after the app’s debut, the internal posts showed.

Some users have reported problems with those calls, however, including one employee who said he tried calling ⁠his insurance company and “they keep hanging up on Muse when they hear it is AI.”

Meta enabled the human concierge feature, also referred to as “human agent calls,” for half of its employees last week to overcome those issues, according to the internal posts.

“Muse is now able to hand requests to a trained agent, who places the call and works it through,” the post announcing the feature to employees said.

Employees who did not want to be included could join an opt-out group, according to the post.