In mid-September 2026, Meta turned on a feature for half its employees that routed some phone calls placed by its new Muse AI assistant to trained human contractors instead of its own AI — without telling the people on the other end of the line they weren't talking to AI. **If the best-funded AI company on earth couldn't make its own assistant's phone calls fully autonomous without a quiet human backstop, that's the exact question any business should ask an AI receptionist vendor before signing a contract: how much of this is actually AI?**
What Did Meta Actually Do With Muse's Phone Calls?
Muse, Meta's AI assistant, launched in the US on September 8, 2026, and had reached roughly 2.5 million downloads within weeks (Superpower Daily, 2026). Muse can place calls on a user's behalf — booking a haircut, checking store inventory, getting a quote — and hands back a transcript afterward. Starting in mid-September, Meta enabled a "human agent calls" option for about half its employees, with an opt-out: when those employees asked Muse to make a call, a trained contractor sometimes placed and handled the call instead of the AI (The Next Web, 2026). At least one tester only discovered a human had made the call after it was already over (Superpower Daily, 2026). Reuters broke the story on September 23, 2026, after internal posts showed employees objecting — one wrote "please PLEASE do not release this as-is" — over sensitive information reaching contractors who hadn't been part of the user's original choice to talk to an AI (VKTR, 2026). A Meta Superintelligence Labs VP called it "a miss" and said the feature had been rolled back; a company spokesperson said feedback was "overwhelmingly positive" and that Meta would only launch with "proper disclosures" (VKTR, 2026).
Why Didn't Meta Just Let the AI Handle the Calls Itself?
Performance, by Meta's own internal numbers. An executive described the company as still "hillclimbing AI calling," with tests showing human-placed calls succeeding 95% to 98% of the time, against a markedly lower success rate when the AI handled the call alone (VKTR, 2026). That's a company with more AI research talent and compute than almost anyone, looking at its own data, and deciding a live person was still the safer bet for a call that actually needed to land correctly. For a business shopping for an AI receptionist on the strength of a vendor's demo reel, that's worth sitting with: if Meta's internal numbers said the AI alone wasn't reliable enough yet, a smaller vendor's unverified accuracy claim deserves the same skepticism.
Has an AI Company Quietly Used Humans Behind the Scenes Before?
Yes, and it's the same company. Meta — then Facebook — launched "M," a Messenger assistant, in 2015 with human trainers helping it handle requests behind the scenes. Three years later, about 70% of tasks still required a human to step in, meaning the assistant had only reached roughly 30% real automation, and Facebook shut M down in 2018 (VKTR, 2026). More than a decade and several generations of far more capable models later, the same company ran into the same gap on a new product: an AI demo that works and an AI system that can be trusted to handle a real call alone, unsupervised, every time, are still two different things. That gap is exactly what an AI receptionist buyer is trying to measure when they ask a vendor what their system actually does versus what their marketing page claims.
How Do You Verify an AI Receptionist Vendor Is Actually Using AI?
Ask the question directly, and ask for proof, not a reassurance. A Kalispell dental office or a Whitefish retail shop evaluating a vendor rarely gets to see what's happening on the other end of a demo call — which is exactly the position Muse's beta testers were in before Reuters' reporting forced the disclosure. The table below is a starting checklist for that conversation.
| Ask the vendor | What a genuinely AI system can show you | Red flag |
|---|---|---|
| Can I hear a live, unscripted call right now, not a recorded demo? | Says yes and dials one on the spot | Needs to "schedule" a live demo for another day |
| What share of calls are fully AI-handled versus reviewed or placed by a person? | Gives a specific number and explains how it's measured | Answers with "mostly automated" or won't give a figure |
| Can I see an unedited transcript or recording from a real customer call? | Shares a real example, redacted for privacy only | Only offers written case studies or screenshots, no audio |
| What exactly happens when the AI doesn't know the answer? | Describes a specific escalation or handoff rule | Claims the AI always resolves the call itself |
When Is a Human-Assisted System Actually the Honest Choice?
Plainly: there's nothing wrong with a system that blends AI and people, as long as it says so. A vendor still in early beta that keeps a trained person listening in as a safety net, a concierge-style service openly marketed as "AI-assisted, human-reviewed," or a business that deliberately keeps a person on legal or medical calls where judgment matters more than speed — all of that is a reasonable, disclosed design choice. What drew the internal objections at Meta wasn't that humans were involved; it was that Muse was marketed and experienced as an autonomous AI assistant while a person was quietly standing in for it, with no way for the caller or the user to know (VKTR, 2026). The line isn't AI versus human. It's disclosed versus hidden.