Meta’s stock has added nearly two hundred billion dollars in market value on the assumption that its new AI agent will scale. Yet early internal trials show that some Muse phone calls are still being handled by contractors through a “human concierge” feature — a detail Wall Street has largely treated as routine product development. It is central to the investment case because investors have priced in automation before knowing how much of the service is genuinely automated or what each completed task will ultimately cost Meta to deliver.

A Fast Launch With Big Expectations

Muse launched on Sept. 8 and quickly moved to the top of the U.S. app charts. It can send emails, arrange travel, shop online, complete forms and call businesses on behalf of users. More than 2.5 million downloads followed within roughly two weeks, and Meta shares jumped 11% in one session as investors started treating Muse as the consumer product that could finally explain the company’s enormous AI expenditure.

It’s no wonder people are excited. Meta has around 3.6 billion daily users across Facebook, Instagram, WhatsApp and Messenger. It is believed to be eyeing Muse subscriptions in the $20 to $100 a month range, which could produce billions of recurring revenue even with a modest conversion rate. The market has already done that calculation. The part it cannot yet calculate is what each paying customer will cost Meta to serve. I previously wrote that rapid adoption can create an extraordinary company while leaving investors uncertain about how much of the economics remain after computing, infrastructure and competition take their share.

Meta Stock Has Already Moved Past The Launch

Muse quickly shifted the focus on Meta. Before the launch, shareholders could see the AI expenditure but had difficulty identifying the consumer product that would justify it. AI improved advertisements, recommendations and engagement, yet those benefits disappeared into Meta’s existing operations. Muse gave investors a separate product with visible adoption and a possible subscription model.

The timing helped. Meta’s advertising business is growing quickly, providing management with the cash flow to finance experimentation. Second-quarter revenue increased 28% to $60.8 billion, although the operating margin fell to 31%. Meta spent $31.1 billion on capital investment during the quarter and expects 2026 capital expenditures of $130 billion to $145 billion, largely to support servers, data centers, networks, and its AI buildout.

It would be wrong to assign the entire recent share-price increase to one application. Technology stocks were recovering, Meta’s core business remained strong and enthusiasm for AI had returned. Muse was still the event that allowed investors to place a revenue opportunity against the spending. Analysts raised targets, the stock surged and investors effectively gave a massive valuation to a product less than two weeks old.

The opportunity could justify it. Meta already owns the distribution that most AI companies would spend years and billions of dollars trying to acquire. Muse can sit inside WhatsApp or be promoted across Instagram and Facebook without Meta paying to acquire each new customer. That advantage is real, but it answers the question of how Meta reaches users. It says little about the cost of completing the work those users request.

The Cost Behind Meta Stock

Reuters reported that Meta is testing a human-concierge feature in which contractors handle some Muse calls. The trial has been offered to half of Meta’s employees on an opt-out basis, with the company describing it as a way to learn about safety and privacy before public release.

Nothing is inherently suspicious about using people to train or improve an unfinished product. The sensible bullish view is that human intervention will decline as the system learns. That may prove correct. Investors simply do not have enough information to assume it.

The simpler jobs should become very automated. Muse can write an email, locate a restaurant or compare simple things. The economics are less obvious when a user asks it to dispute a hotel payment, rebook a missed trip or cancel a service whose provider doesn’t want consumers to quit. These tasks involve negotiation, a lack of knowledge and corporations that lack motivation to collaborate. And someone must be accountable if the agent bookings the wrong date, buys the wrong stuff or gives out information the buyer thought was private.

I examined the same operational problem with Google Gemini . An AI agent does not need malicious intent to create a real cost. It only needs permission to act and an incorrect understanding of what it has been asked to do.

That is where the income statement begins to look different from the product demonstration. A task may require repeated model calls, access to several outside services, and a person to resolve the exceptions. Customers paying $100 a month will probably be the heaviest users, so the most valuable subscribers could also be the most expensive to serve.

Another problem is Meta’s interaction with merchants. Shopify has welcomed Muse, which has added Shop Pay and sees the agent as another channel to its merchants. Amazon has banned Muse from buying on its site, citing concerns about unlawful activities, privacy, and user experience.

Meta cannot force the commercial internet to accept its agent. Some companies will welcome the additional transactions, while others will protect their customer relationships and data. Agreements, technical integrations, and possible revenue-sharing arrangements will influence the eventual economics. An app-download chart does not capture any of that.

Why Meta Stock Still Has The Best Distribution

The unanswered cost questions do not make Muse a failure, and Meta is likely to have the best starting position of any company in consumer AI because it is already woven into the daily lives of billions of people. It understands how they communicate, which brands they follow, and which material keeps them engaged. WhatsApp also gives Muse a familiar home that avoids asking users to adopt another unfamiliar interface.

Meta can afford a lengthy learning period. The advertising business generates the money needed to hire contractors, build infrastructure, and absorb early mistakes. A start-up would need to demonstrate viable economics before its capital disappeared. Meta can improve Muse while its existing operations carry the cost.

Human involvement may help that process. Contractors can identify the requests that confuse the software, the moments when users become uncomfortable, and the safeguards needed before an agent acts. If those lessons reduce intervention over time, the current expense could become a useful investment in automation.

The attraction for Meta extends beyond subscriptions. If Muse becomes the place where a consumer begins a purchase, books a trip, or contacts a company, Meta gains another opportunity to understand commercial intent. That information could strengthen its advertising business and give it influence over transactions that currently begin with Google, Amazon, or a retailer’s own application. The strategic value may eventually be greater than the subscription revenue.

That possibility helps explain why Wall Street moved so quickly. Meta has repeatedly shown that distribution can rescue products that were not first to market. It does not need to invent every important technology. It needs to place a useful version in front of more people than its competitors can reach.

The Number Wall Street Needs

Downloads will keep the story alive for a while, but they will not settle the investment case. Paid conversions and subscriber retention will matter, although one figure may tell investors more than either: the cost of a successfully completed task.

That number would capture the computing expense, the percentage of jobs completed without help, and the cost of resolving mistakes. A falling cost per task would show that Muse is learning and that subscription revenue can scale. A stubbornly high cost would suggest that Meta has built a useful service with economics closer to a concierge business than a software platform.

Meta may get there. Its distribution, balance sheet, and existing advertising machinery provide advantages no new entrant can easily reproduce. The company has already shown that people want to try Muse, and the current human testing may help turn an impressive launch into a dependable product.

Investors have nevertheless awarded Meta close to $200 billion before seeing those economics. Meta stock can sustain that increase if automation rises, human intervention falls, and Muse becomes a profitable gateway to online activity. The next stage of the story will not be decided by how many people download the application. It will be decided by what Meta earns each time Muse completes the job.