Meta’s Muse calling agent sometimes hands the conversation to a human call centre

Meta is testing a human fallback for Muse, its assistant that calls businesses to make reservations, book appointments or handle customer service. Internal material reviewed by 404 Media says Muse can hand a request to a trained human agent who places the call and completes it. Meta has not said how frequently this occurs or what kinds of failures trigger the transfer.

The arrangement is not inherently deceptive: human escalation is a normal way to make an unreliable automated service useful. The problem is consent. Employees said at least one internal tester learned only after the call that a person had handled it. A contractor may therefore see the user’s request, relevant chat context and potentially sensitive information under circumstances the user understood as a private machine interaction. Training contractors in data handling reduces risk; it does not substitute for an access-control policy or prior disclosure.

Meta says this is confidential pre-launch testing, that employee feedback is being used to add safety and privacy protections, and that any public release would come with appropriate disclosures. That qualification is material: the report does not establish that ordinary users are currently being routed to undisclosed humans.

Muse’s economics and capability claims nevertheless need to separate three things: calls completed entirely by AI, calls in which people intervene, and calls that fail. A polished overall completion rate can hide expensive human labour and very different privacy exposure. The honest product would ask permission before escalation, show what information the human will receive, minimize that packet and let the user decline without losing the rest of the assistant.