Product1 publisher2 min readPublished
Meta routed Muse's rejected phone calls to human contractors
Reuters found that Meta's chart-topping Muse agent escalated calls to paid human concierges when the AI voice could not finish the job. A vice president says the feature is rolled back, and Meta has not published the AI-only success rate.
The Product Desk · Product desk
What happened
- Meta's Muse agent platform launched on September 8 and has topped the US iOS download chart, helped by a beta feature that places phone calls on a user's behalf.
- Reuters reported that some Muse tasks used a "human concierge" feature, letting a person take over the call when the AI voice was not up to the task.
- In internal Meta conversations seen by Reuters, a user said his insurance company repeatedly hung up as soon as the Muse bot disclosed that it was not human.
- An employee tester who read a call transcript said the human operator had made a racist reference, and a Meta vice president said the company had stopped working with that contractor.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Whether a disclosed bot can complete an errand is decided by the person who picks up, so the ceiling on consumer voice agents sits in merchant phone policy. Better models will not move it.
- exposure Anyone whose errand escalated to a concierge had their details spoken aloud to a contractor they never hired. Meta employees raised that risk internally before any public launch.
- decision Anyone buying or building a calling agent now has to specify which success rate is being quoted, because the high-90s number was earned with a person available to take over.
- precedent Meta gave its own staff a way out of human-assisted calls. I'd expect a consumer release to come with the same opt-out.
The one figure a Meta Superintelligence Labs vice president put on the record measures people. Meta's tests showed human involvement could push the phone call success rate into the high 90% range against what the VP called a "lower percentage of AI calling," Reuters reported [10]. Reuters did not report the AI-only number, and Gizmodo said it asked Meta for it [13].
Muse shipped on September 8 [1]. Alexandr Wang posted that the calling beta was expanding on September 16 [2]. Eight days separate the launch from the wider rollout of a feature that had human concierges standing behind it [3][4].
Reuters described the feature as able to "handle errands like booking haircuts, checking whether a store has an item in stock or getting quotes from different contractors" [11]. Haircuts are the easy case. Internal conversations at Meta questioned the wisdom of the escalation, with some employees worried it could feed sensitive user information to a human operator [7]. When human calling was switched on for Meta's own employee testers, they were allowed to opt out [6].
Meta spokesperson Dave Arnold said internal testing, "aka dogfooding, is core to the product development process," and that the company is "working with merchants to continue improving this potential calling feature, and will only roll it out when it's ready and with the proper disclosures" [12]. The disclosure Arnold promises is the same condition that got the bot rejected in the case Reuters described [5]. The vice president told the employee who complained about a contractor that the company had "rolled back this feature" and that the way it had tested calls from human contractors was "a miss" [9].
A success rate measured with a human available describes the escalation path. Anyone piloting a voice agent next week should ask whether the business being called will stay on the line after a bot identifies itself, and in my view that is a property of the merchant's phone policy. Then ask whether the errand carries anything the user would not repeat to a stranger [7].
Booking a haircut clears both tests. A call about a claim, a prescription or a clinic appointment clears neither. Putting a contractor on the line leaves the first failure in place and makes the second one worse. The number to ask any vendor for is the completion rate for calls the model finished alone, broken out by the kind of business on the other end.
What to watch
- Watch for the calling beta to return publicly with the disclosures Arnold promised, and for what merchants are told before it does.
- How Meta reviewed concierge call transcripts, and how many were checked before the racist-reference complaint.
- Whether consumers get the same opt-out from human involvement that Meta's employee testers were given.