Meta Platforms is testing a human concierge layer for Muse, its personal AI agent, with contractors handling some phone calls when the assistant cannot complete them autonomously. Users may not always know when a call has been transferred to a call-center agent. Muse launched in the United States through a dedicated app, the web and WhatsApp, with plans for AI smart glasses, and reached the top of Apple's free U.S. app chart ahead of ChatGPT. Meta shares rose 11.4% to $741.25 on Sept. 21, adding about $192.3 billion in market value. The test comes as Meta develops an assistant that can execute tasks such as filling forms, sending emails, booking travel, making purchases and negotiating for users. Employees have raised concerns about sensitive information reaching contractors. Meta spokesperson Daniel Roberts said feedback was "overwhelmingly positive" and that the test is intended to improve safety, privacy protections and features before public release. Muse's rollout has also involved reported safeguard failures, including exposure of a user's iCloud photos and unexpectedly halted tasks. Meta says a separate monitoring AI checks external connections, requires approval for emails and purchases, and cannot directly view passwords or card numbers. Amazon has blocked Muse from buying products on its site, highlighting the dependence of AI agents on cooperation from outside platforms.