Meta tests human concierge for Muse phone calls, then rolls it back

Meta quietly routed some Muse outbound calls to call-center contractors, then paused the human-concierge dogfood after employees flagged privacy gaps, missing disclosures, and a launch safety pitch that promised dedicated VMs and secure storage.

Meta has been testing a “human concierge” for Muse in which contractors quietly place some of the phone calls users ask the agent to make, Reuters reported in an exclusive by Katie Paul on September 22. The company told employees about the test last week, shortly after Muse’s public phone-calling feature went live, and enabled “human agent calls” for about half of Meta staff, with an opt-out group for anyone who did not want in.

Muse has topped US app download charts since its debut and is built to handle emails, shopping, and travel. The calling lane dials US businesses for errands such as booking haircuts, checking stock, or collecting contractor quotes. Internally, Meta framed the human layer as a way to finish those calls when pure AI calling failed — some tests put human-handled success in the 95% to 98% range against a lower AI-only rate, Reuters said.

Employees pushed back hard on privacy. One called the feature “one bug away” from leaking unnecessary information to human callers. Another, after asking Muse to negotiate a cable bill, said a transcript showed the contractor made a racist reference on the call. Staff also warned that routing requests to call-center humans undercut Muse’s launch safety pitch — each agent in its own secure virtual machine, with passwords in separate secure storage. 404 Media (Jason Koebler) separately reported a confidential pre-launch internal post, PR fear that headlines would say the AI was not good enough without humans, and testers who were not told a human had placed the call until after it happened.

A Superintelligence Labs vice president acknowledged “it was a miss” to start the test without proper disclosures and said Meta had rolled the feature back for now while still “hillclimbing AI calling.” She apologized over the racist-reference incident and said that agent would not work Meta projects again. Spokesperson Daniel Roberts told Reuters employee response was “overwhelmingly positive,” that dogfooding was meant to improve safety and privacy before any public release, that Meta is working with merchants, and that the calling feature would roll out only when ready and with proper disclosures.

The episode echoes Facebook’s earlier Messenger assistant “M,” which media reports later estimated was about 70% human-handled. AI Tech Daily’s live Muse wires cover a separate macOS privilege hotfix and Wall Street’s revenue-engine read — this brief is the human-concierge disclosure miss and rollback, not a rewrite of those.

Meta tests human concierge for Muse phone calls, then rolls it back

Sources