Meta opens Muse business calls to US subscribers paying from $20 a month

Meta said last month that Muse had more than 10 million users. That company-reported total does not establish how many pay or use calling. Keeping calls exclusive to subscribers gives a concrete household task a role in the paid offering. By contrast, Wang said this week that Muse’s recently added dedicated email-address feature would soon reach free users.

Google's illustration accompanying its October 2020 Duplex update provides historical context for automated business calls. Source: Google.

Calling is already a competitive category

Automated business calls have a longer history than this rollout. Google’s October 2020 account of Duplex, its conversational AI technology introduced in 2018, described booking appointments and reservations and waiting on hold. Google said its Assistant had completed more than a million bookings since launch and that it always disclosed an automated system when making a call. Those historical company claims establish a precedent, not a benchmark of Muse’s current performance.

Today, Google’s Gemini-powered Call for Me offers a more immediate comparison. It uses the user’s phone number, provides live transcripts and lets the user take over, while remaining limited to eligible Pixel owners in the US, according to the feature comparison. Access expanded to Pixel 10 phones this week, the account says.

That comparison also describes startup Instinct’s Concierge calling service, Grok Bot’s calls to US businesses and consenting individuals, and outgoing calls through third-party connectors such as AgentPhone for Anthropic’s Claude. These are different access arrangements, not evidence of equivalent performance. Meta has not announced a similar extension of Muse calling to personal contacts.

Why Meta tried human callers

Muse was designed to handle tasks such as sending emails, shopping and booking travel. A September Reuters report republished by The Business Times, based on internal company posts, documented an obstacle to its calling feature: one employee said his insurer repeatedly hung up when it heard Muse was AI.

The posts showed Meta had enabled an experimental human concierge for half its employees, allowing requests to pass to contractors who placed the calls. Employees could opt out. A Superintelligence Labs vice-president said some tests suggested human callers could bring success rates into the 95–98% range, above an unspecified AI rate. The report supplied no sample size, task mix or comparable AI percentage. These figures explain the stated rationale for the experiment; they do not measure the current subscriber rollout.

Employees warned that contractors could receive sensitive information unintentionally. One employee also said a transcript showed a contractor making a racist reference during a bill-negotiation call. The vice-president apologised, said the contractor would no longer work on Meta projects, and acknowledged that testing without proper disclosures had been a mistake. She said the feature had been rolled back.

The concern cut across Meta’s product pitch. The company had promoted a separate cloud-based virtual machine for each agent to isolate sessions, with passwords kept in separate secure storage. Those protections did not answer employees’ questions about information passed to human callers.

Meta spokesperson Daniel Roberts told Reuters that employee response had been overwhelmingly positive and that testing sought feedback to improve safety and privacy before public release. He said the potential human-calling feature would roll out only when ready and with proper disclosures. The report also recalled Meta’s earlier Messenger assistant, M, whose tasks media reports said often depended on people—a relevant precedent for distinguishing an automated interface from the labor behind it.

What would establish the value of a paid call

An account of a demonstration shared by Matt Schlicht, a Meta Superintelligence employee, describes a recreation of his agent calling a Volvo dealership about appointment availability. Schlicht said it waited on the line for 15 minutes. This was an employee’s example, not an independent test of typical performance. The shared call did not explicitly identify the caller as AI, although it disclosed that it was calling on Schlicht’s behalf and transcribing the conversation.

That example leaves the broader disclosure practice unresolved. The September contractor experiment likewise does not establish who handles calls in the current rollout.

The evidence needed now is how often subscriber calls finish the requested task, how often users or contractors must intervene, what information reaches those people, and how much usage a completed task consumes. The launch accounts provide no comparable results on those measures. Those answers would let subscribers judge whether the monthly fee buys dependable delegation as well as access to a caller.