Meta’s Human Concierge Test Raises a Privacy Question for Calling AI Agents

Meta is testing a “human concierge” option for its Muse assistant’s calling feature, according to internal company posts reviewed by Reuters. Under the test, contractors may handle some calls that users ask Muse to place. Employees raised concerns that sensitive information could be shared with call-center workers without callers realizing a person was involved.

Meta told Reuters that the trial is intended to gather feedback and improve safety and privacy protections before any public release. The company said employees could opt out of testing. Muse’s call feature is designed to contact businesses for tasks such as booking appointments or checking prices, then provide users with summaries and transcripts.

A service boundary users should be able to see

Human help can make an agent more useful when a call requires negotiation, clarification, or a judgment that software cannot reliably make. But a human handoff changes the privacy expectations. People may share addresses, health details, schedules, or payment context because they believe they are speaking only to an automated system.

A user-facing product should identify when a human is listening or acting, explain what information will be available to them, and provide a meaningful way to approve or decline the handoff. Summaries and transcripts can help users audit what happened, but they do not substitute for informed consent before sensitive data is exposed.

The test for AI calling products

Calling agents will need to prove that they can distinguish routine tasks from sensitive ones, handle unexpected questions, and stop when a caller asks for a person or when the conversation leaves the user’s instructions. Companies should publish clear policies for recording, retention, contractor access, and complaint handling.

Meta’s experiment highlights a broader design challenge: autonomous products often depend on human labor behind the scenes. Making that boundary visible is part of building an agent users can trust.

Source: Reuters report on Meta’s Muse concierge test.