Meta Routes Some Muse Phone Requests to Human Callers in Pre-Launch Tests

A call placed through an AI assistant may end up in a person’s hands. During internal testing, Meta’s Muse can pass phone requests to trained human callers who contact businesses on a user’s behalf, according to internal communications obtained by 404 Media. Some employees say testers did not learn about that handoff until the call was over.

One request, two ways to make the call

Meta expanded Muse’s beta for outbound calls to U.S. businesses on September 16. The idea is to ask the agent to handle an errand, such as booking a restaurant table or a haircut, then get back a transcript and summary. An internal post described Muse as handling the conversation and completing the request.

That same post said Muse can instead hand a request to a trained person, who places the call and works through it. Meta described this human-caller setup as confidential, pre-launch and ready for employee testing. How often requests reach people, and what triggers a handoff, remains unclear.

A completed errand, but a late disclosure

Finishing the errand is one measure of whether the feature works. Knowing who will hear the request is another. In the employee discussions reported by 404 Media, one worker said a tester discovered only after a call that a human had made it. The tester had shared information believing the caller would be the AI, the worker said.

Employees worried that people might give Muse sensitive details without realizing a human contractor could receive them. A medical appointment was raised as a possible example, not a documented exposure. One employee also worried that reliance on callers would make the AI appear unable to do the job alone.

please PLEASE do not release this as-is and at least make this a user preference if we absolutely must launch this. Please.

Meta employee, in internal comments reported by 404 Media

Training is not the same as a choice

Meta told employees that contractors had received extensive training to keep user data safe. An employee pushed back, arguing that training was not itself a security mechanism. Those are different questions: how a caller handles information after receiving it, and whether a user gets to decide beforehand if a person should receive it at all.

A Meta spokesperson told 404 Media that employee testing is meant to gather feedback and improve safety and privacy protections. Meta characterized the response from employees as overwhelmingly positive. It said it was working with merchants on the potential feature and would roll it out only when ready, with proper disclosures.

That promise leaves a product decision open: whether users will be able to decline a human handoff and still make a call. In the employee’s account, a tester learned the caller was human only after sharing information. A choice offered afterward cannot change what the tester already disclosed.