Meta rolls back Muse human-concierge test as downloads pass 2.5 million

Meta Platforms, Inc. (Nasdaq: META) has rolled back a test that allowed human contractors to handle some telephone calls initiated through its new Muse personal artificial intelligence agent after employees raised privacy and disclosure concerns. Reuters reported the experiment on 22 September after reviewing internal company posts and obtaining a response from Meta Platforms. The disclosure arrived as Muse passed 2.5 million downloads and reached the top of Apple’s US app rankings within two weeks of its 8 September launch.

Muse is designed to perform tasks rather than only answer questions, including booking travel, shopping, sending emails, comparing service quotes and negotiating bills. Its phone feature can contact US businesses to arrange appointments or check stock, bringing the agent into conversations with people who may not expect an artificial intelligence system. Meta Platforms tested a human-concierge layer because contractors could complete some calls that automated systems struggled to finish.

Internal tests cited by Reuters suggested contractor-assisted calls could achieve success rates of 95% to 98%, compared with a lower rate for artificial intelligence-only calling. Meta Platforms acknowledged through internal communications that the feature had been introduced without adequate disclosure and said it had been withdrawn for now. The company said publicly that it would continue improving the capability and would release it only with appropriate safeguards and transparency.

What did Meta test behind Muse’s artificial intelligence phone calls?

The test allowed Muse to transfer selected requests to trained human agents who would place calls and work through the task. Half of Meta Platforms employees were reportedly enabled for the experiment, while employees who did not want to participate could opt out. The approach was intended to overcome situations in which businesses ended calls after recognising that they were speaking with an automated agent.

Employees warned internally that contractors might receive information users believed was confined to an artificial intelligence system. One employee who used Muse to negotiate with an internet and cable provider reported that a transcript contained a racist reference made by the contractor, according to Reuters. Meta Platforms apologised internally and said the individual involved would no longer work on its projects.

The episode does not show that all Muse calls involved people or that the released service routinely exposed user data. It does show that the boundary between automated and human execution was not sufficiently clear during the test. For a product marketed around privacy and secure computation, uncertainty over who performs a task can become more damaging than a normal early-stage technical error.

Meta Platforms said Muse operates in a dedicated secure virtual machine and keeps sensitive information such as passwords in separate storage. That architecture can isolate digital sessions, but it cannot by itself control what a human contractor hears or sees after a task is transferred. The company must therefore extend privacy design beyond software to contractor access, training, monitoring, retention and consent.

Why does human assistance challenge Muse’s autonomous AI proposition?

Customers may accept a supervised service if the human role is disclosed before sensitive information is shared. The problem arises when an experience presented as artificial intelligence quietly depends on people, because the user makes privacy decisions under an incomplete assumption. Clear consent must explain whether a contractor can hear the call, view context, access transcripts or take action using personal details.

Human assistance can be commercially useful while models improve. It can rescue failed tasks, identify difficult conversation patterns and generate training data for future automation. The operational cost, however, is much higher than pure software and can limit margins if a large proportion of requests requires manual intervention.

Meta Platforms has encountered this tension before. Its earlier Messenger assistant, M, used extensive human support before being discontinued, showing how difficult it is to automate open-ended personal tasks across unpredictable businesses. Muse has more capable models and infrastructure, but telephone negotiation still involves accents, interruptions, policy exceptions and counterparties that may refuse to engage with a bot.

The responsible commercial path is not necessarily to eliminate people immediately. Meta Platforms could present human escalation as an optional premium service with explicit consent, limited data access and auditable controls. That model would be less magical than a fully autonomous agent, but it could be more reliable and easier to defend to regulators and users.

Regulators may also treat the disclosure question as a consumer-protection issue rather than a product-design preference. A caller, merchant or user could reasonably want to know when a conversation is recorded, automated or transferred to a contractor in another location. Standardised notices and permission controls may slow some tasks, but they would give Meta Platforms a clearer basis for scaling the feature internationally.

Can 2.5 million Muse downloads become durable subscription revenue?

Initial adoption is unusually strong. Sensor Tower data cited by Reuters showed more than 2.5 million downloads and a number-one position in the US Apple app store, ahead of established assistants including ChatGPT and Claude. Meta Platforms can also distribute Muse through WhatsApp and eventually other products, giving it access to an ecosystem with about 3.6 billion daily active users.

The basic service is free, while Meta Platforms plans subscription levels of roughly $20 per month and as much as $100 for heavier users. That structure allows rapid adoption before the company asks customers to pay for higher usage or more valuable tasks. Conversion will depend on whether Muse saves enough time or money to become a recurring utility rather than a short-lived experiment.

Agentic commerce creates another revenue path. Muse can complete purchases through Stripe’s Link checkout and use protected, single-use payment credentials, potentially placing Meta Platforms between consumers and merchants. The opportunity is substantial, but every transaction increases the consequences of an incorrect order, hidden fee, privacy failure or unauthorised action.

Costs will determine whether subscriptions become attractive. Model inference, secure virtual machines, customer support, payment protections and human intervention can consume a meaningful share of monthly fees. Meta Platforms needs to disclose enough operating evidence for investors to judge whether Muse is a high-margin software service or an expensive concierge subsidised by the wider advertising business.

What does Meta Platforms’ share surge reveal about investor expectations?

Meta Platforms shares gained more than 20% between the Muse launch and 22 September, adding more than $200 billion of market value and reaching their highest level in over seven months, according to Reuters. The stock then traded near $753.46 around 16:56 UTC on 23 September, up about 2.3% for the session and valuing the group at approximately $1.93 trillion. Investors are clearly assigning substantial option value to the possibility that Meta Platforms can turn its social reach into a leading consumer-agent platform.

That valuation response is much larger than any revenue Muse has yet disclosed. Reuters Breakingviews estimated that the added market value implied about $24 billion of additional revenue when measured against Meta Platforms’ recent enterprise-value-to-revenue multiple. The calculation is illustrative rather than a forecast, but it shows how much successful subscription and commerce conversion is already embedded in sentiment.

Shareholders also have reasons for caution. Meta Platforms reported weaker free cash flow amid heavy artificial intelligence spending earlier in 2026, agreed to an $18 billion settlement related to teen-safety litigation and has accumulated roughly $80 billion of Reality Labs losses. Muse must therefore produce more than downloads if it is to justify another large capital and trust commitment.

The human-concierge test is an early governance warning rather than proof that Muse will fail. Meta Platforms acted by withdrawing the feature and promising better disclosure, but the next version must show that safety controls advance as quickly as functionality. The company’s network can deliver adoption at breathtaking speed; preserving that advantage requires users to know whether they are dealing with software, a person or both.