|
Getting your Trinity Audio player ready...
|
Meta Is Quietly Using Human Workers to Handle Phone Calls for Its AI Agent Muse
Meta’s new AI assistant Muse promises to call businesses on your behalf and handle the awkward parts of adult life, booking a haircut, checking if a store has something in stock, getting quotes from contractors, all without you having to pick up the phone yourself. According to internal company posts reviewed by Reuters, though, a meaningful chunk of those calls aren’t actually being handled by AI at all. They’re being routed to real human contractors working behind the scenes.
Meta informed employees about the test last week, shortly after publicly rolling out Muse’s phone-calling feature to users. The company began having its own staff try out the calling capability internally back in August, a process employees internally call dogfooding, before gradually extending it to Muse users in the days following the feature’s public debut. Last week, Meta enabled what it’s calling the human concierge feature, also referred to internally as human agent calls, for roughly half of its employees. In practice, when Muse can’t complete a phone call reliably on its own, the request gets handed off to a trained human agent who actually places the call and works through it with the person on the other end. Once the call wraps up, Muse’s app shows the user a summary and full transcript of how the conversation went, giving the appearance of an AI-handled interaction even when a human carried most of the actual work.
Employees weren’t required to participate either. Those who preferred not to have their calls potentially routed through a human contractor were given the option to join an opt-out group, according to the internal posts. Not everyone inside Meta has been fully comfortable with the setup regardless. Some staff reportedly raised privacy concerns internally, worried that sensitive personal information shared during these calls could end up passed along unintentionally to contractors working in call centers, people who, unlike a user’s own AI assistant, aren’t necessarily bound by the same expectations around data handling a person might assume when talking to software rather than a stranger.
Asked about the test, Meta spokesperson Daniel Roberts said employee feedback on the feature has been overwhelmingly positive so far. He said the purpose of the internal trial is to gather feedback that will help Meta build in the right safety and privacy protections before the feature reaches a wider public rollout, adding that the company is working with merchants to keep improving the calling experience and won’t release it more broadly until it’s genuinely ready, with proper disclosure to everyone involved in a call.
Meta isn’t exactly breaking new ground with this approach. A decade ago, the company, then still operating under the Facebook name, ran a strikingly similar experiment with a digital assistant called M, built into Messenger and pitched at the time as a genuinely intelligent virtual helper. Media investigations later revealed that roughly 70 percent of M’s tasks were actually being completed by human workers rather than the AI itself, a gap between marketed capability and actual execution that became something of a cautionary tale about how far tech companies were willing to stretch the definition of artificial intelligence in a product’s early days. The Muse test suggests that same tension, between what AI agents can currently do reliably on their own and what companies want to promise users right now, hasn’t gone away, it’s just resurfaced with considerably more sophisticated underlying technology this time around.
The stakes for getting Muse right are higher for Meta than M ever was. The assistant has topped US app download charts in the two weeks since launching, pulling in roughly 730,000 downloads in its first five days and climbing past 2.5 million within thirteen days, according to data from Sensor Tower. Muse sits at the center of CEO Mark Zuckerberg’s broader vision of delivering what he’s called personal superintelligence to the billions of people already using Meta’s platforms daily, making it one of the company’s most closely watched product bets in years.
That ambition has already run into resistance elsewhere in the industry. Amazon confirmed this week that it has blocked Muse from shopping on Amazon.com entirely, citing data security concerns, a move that reflects a broader standoff brewing across the tech industry over who actually controls the shopping relationship and customer data as AI agents increasingly try to shop, book and transact on people’s behalf without a human directly clicking through a checkout page themselves. Shopify, by contrast, took the opposite approach, announcing a partnership Monday to bring its Shop Pay checkout system directly into Muse, giving the assistant a friendlier path into e-commerce even as Amazon shuts its own doors.
Wall Street has taken notice of what Muse could mean beyond retail too. Shares of major banks, insurers and online travel companies slid this week as investors weighed the possibility that AI agents like Muse could erode what’s sometimes called consumer inertia, the habitual tendency of people to keep buying from the same bank, insurer or travel booking site simply because switching feels like more effort than it’s worth. An AI agent willing to shop around, compare options and make calls on a person’s behalf threatens to strip away exactly that kind of habitual loyalty, and investors appear to be pricing in the possibility that Muse’s phone-calling feature, human-assisted or not, represents a genuine long-term threat to industries that have quietly relied on customer inertia for years.
