# Zuckerberg says Muse will make people money. Who will capture the gains?

> Meta’s Muse can act across services, and Zuckerberg expects it to help users earn and save money. We examine the launch, proposed fees and evidence behind that bet.

By BIG CHANGE Editorial

Published: 2026-09-22T02:10:07.229Z
Updated: 2026-09-22T02:10:07.229Z
Canonical: https://bigchange.ai/blog/meta-muse-zuckerberg-make-money-earnings-business-model

![Pencil sketch of a small-shop workbench with task cards, a laptop, a parcel and an unfinished receipt beside an open doorway.](https://bigchange.ai/api/media/file/muse-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE. Business preparation still needs a customer at the door.

Meta launched its personal AI agent, Muse, on September 8, 2026. It gives users a computer in the cloud that can work across connected services, continue tasks in the background and ask for approval before consequential actions. The initial rollout is in the United States. [Meta’s launch announcement](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/)

Mark Zuckerberg’s pitch goes further than clearing an inbox. In an interview with Alex Heath, he predicts that Muse will “make people money and save people money.” He describes a business model in which Meta eventually takes a small share of transactions, potentially paid by the businesses involved. [Sources interview, 24:53](https://www.youtube.com/watch?v=Lx8lrn-cytc&t=1493s)

That proposition deserves a serious hearing. A business owner who cannot afford administrative help might gain useful capacity from an agent. A household might recover money lost to an overlooked bill or find a cheaper purchase. Reducing those burdens would matter even without a spectacular breakthrough in intelligence.

But earning money requires someone else to pay. A tool that creates a product, drafts its marketing and runs its advertising has automated parts of a business. It has not established that customers want the product, that the advertising is profitable or that the owner will keep enough of the proceeds.

Muse’s launch makes that distinction more urgent because the same platform could help a user sell, help another user buy and eventually charge for connecting the two. The economic question extends beyond whether the agent works: how much value reaches its user after the other participants are paid?

## What Zuckerberg actually proposed

The interview’s central money discussion runs from roughly 24:53 to 27:25. Zuckerberg connects the opportunity to small businesses and commerce, describes a large free allowance alongside subscriptions, and presents transaction fees as an expected business model “over time.” He says Muse can connect to Meta’s advertising systems and help with making a product and running the business. [Watch the full exchange](https://www.youtube.com/watch?v=Lx8lrn-cytc&t=1502s)

Those remarks describe several mechanisms that should be evaluated separately. Selling more produces revenue. Negotiating a lower bill produces savings. Finishing paperwork faster frees time. Each can benefit a user, but they are different outcomes, and only some put additional cash in an account.

The launch announcement and this exchange do not establish a general program in which Meta pays people for using Muse. Nor do they provide an income guarantee, a published profit-sharing percentage or evidence that typical users are already earning a net return. The transaction-fee proposal concerns how Meta might be paid; it should not be mistaken for a payment schedule to users.

Alex Heath’s [original episode page](https://sources.news/p/mark-zuckerberg-meta-muse-ai-podcast-interview) was published on September 8 and says the conversation was recorded the preceding Friday at Meta headquarters. We reviewed the complete available YouTube captions and checked the relevant launch and technical material. This is an analysis of that evidence, not a BIG CHANGE hands-on test or an earnings experiment.

## What the agent changes about doing the work

Muse is the personal agent, while Muse Spark is the model powering it. Meta’s product design account describes a file system, terminal and browser, plus persistent memory and scheduled or event-driven work. It can generate documents and interactive artifacts as well as perform actions through connected services. [How Meta designed Muse](https://introducing.muse.ai/)

This gives it a different potential role from a chatbot that only supplies advice. A useful agent can carry information from one task to the next: retrieve an inquiry, prepare a response, update a record and return when the customer replies. The possible saving comes from removing repeated coordination, not simply producing text faster.

Consider a hypothetical furniture repair business. The owner spends evenings sorting photographs, checking measurements and arranging appointments. An agent that reliably prepares that information could leave more time for paid repair work. The business already has a service and customers; the agent addresses a known administrative constraint.

A new business selling generic AI-generated designs faces a different problem. Muse might help produce a catalog quickly, but the catalog still needs a reason to exist. Buyers can compare it with alternatives, including designs they generate themselves. Automation lowers the effort required to supply something without necessarily increasing willingness to pay for it.

That difference should guide expectations. Existing demand combined with an expensive routine task is a more concrete opportunity than asking an agent to invent an income stream from scratch. The latter can work, but it requires evidence of customer interest and a viable margin.

![Three-part sketch linking a laptop and stopwatch to a shop, then to an accounting ledger under a magnifying glass.](/api/media/file/muse-economics-v1.png)

## The payment connection is real, but it moves money out

Stripe independently confirms the launch integration. US consumers can connect Link to Muse; the agent can use saved payment methods at participating businesses or a single-use virtual card for other purchases. Stripe says consumers approve each transaction total in the chat, and Muse does not see the underlying payment details. [Stripe’s September 8 announcement](https://stripe.com/newsroom/news/stripe-helps-meta-muse-shop-with-link)

That is meaningful infrastructure. An agent that can research a purchase but cannot complete it leaves the user to repeat part of the work. A payment connection can close that gap while keeping an approval step.

It also clarifies the direction of the documented money flow. This integration enables a user to buy from a merchant. It does not, by itself, show the user earning money. Saving on a necessary purchase may improve the user’s finances; buying something unnecessary at a discount still reduces their cash balance.

On the other side, a merchant may gain a sale. Whether that sale is profitable depends on fulfillment costs, refunds, payment charges and the cost of attracting the buyer. A successful checkout demonstrates that a transaction happened. It does not establish how much economic benefit each participant retained.

This is why future earnings claims should name the mechanism. Was income generated by a new customer? Was an expense avoided? Was an asset sold? Was the user simply able to complete an existing task sooner? Lumping these together can make a capable assistant sound like a source of automatic income.

## Free access lowers one barrier, not every cost

Zuckerberg says the launch allowance is about 100 million tokens per week for free, with subscriptions available. Treat that as a dated description of access, not a permanent promise of unlimited work. [Interview, 25:02](https://www.youtube.com/watch?v=Lx8lrn-cytc&t=1502s)

Meta’s subscription page, checked September 22, lists Power at $20 a month with 500 million Muse tokens per week and Max at $100 with 3 billion. It says benefits vary by region and account, and availability remains limited. These are product allowances, not a count of completed jobs or directly comparable API tokens. [Official Muse subscription details](https://www.meta.com/help/subscriptions/1021145227643680/)

A token allowance does not tell a business how many useful orders, reconciled invoices or accepted deliverables it will obtain. The task, context and number of attempts matter. Work can also consume resources outside the agent: advertising, hosting, inventory, delivery and the owner’s attention.

The largest early cost may be learning which tasks deserve delegation. Someone must explain the objective, connect the right accounts, recognize an implausible result and decide when an action is worth approving. Better product design can reduce that burden. It cannot make a user’s priorities or a customer’s requirements disappear.

Time saved also needs an honest accounting. An hour freed for paid work can increase income if there is work available. An hour spent with family has value without becoming earnings. Converting every saved minute into an assumed hourly wage would overstate the cash return for many users.

For someone starting a business, the relevant result is additional money retained after costs and obligations, compared with what would have happened without the agent. For an existing business, it may be fewer missed inquiries, quicker completion or reduced administrative spending. Those outcomes are measurable, and none requires pretending that every task completed is a sale.

## A transaction fee could support the product and change its incentives

An agent supported by transaction fees has a plausible source of revenue. If it helps customers find suitable products and reduces the work needed to complete a purchase, merchants may be willing to pay for that service. A fee could fund access for users who would otherwise find a subscription too expensive.

The arrangement also creates questions about whose interests shape recommendations. Would Muse recommend the best available option when that option generates no commission? Could a merchant participate without paying? Would the user see a commercial relationship before approving a purchase? How would competing offers be compared?

The material reviewed does not answer these questions with a complete Muse transaction-fee schedule or ranking policy. We should not fill that gap by asserting that Muse already favors commission-paying sellers. The point is that an eventual fee model needs rules that users and businesses can inspect.

Charging the merchant does not automatically remove the economic cost from the buyer. A merchant might absorb a charge, offset it through cheaper customer acquisition or reflect it in prices. Which happens depends on the market. A fee that replaces a larger existing cost could leave both sides better off; an extra fee could reduce the available margin.

This is a more useful way to assess Zuckerberg’s proposal than treating a free agent as either a gift or a trap. The product can create value and capture part of it. The unresolved question is how that value is divided and whether users retain meaningful choices as they rely on the service.

## The evidence for everyday usefulness is ahead of the earnings evidence

In a September 14 hands-on account, TechRadar’s Eric Hal Schwartz describes Muse helping him shop and preparing a reply to a utility-company email for approval. He found the delegation useful while remaining uncomfortable with the access involved. That is evidence from one reviewer’s experience with routine tasks, not a study of business profits. [TechRadar’s firsthand review](https://www.techradar.com/ai-platforms-assistants/i-tried-metas-new-muse-ai-agent-its-incredibly-useful-but-handing-it-my-digital-life-felt-deeply-uncomfortable)

Earlier research gives a reason to be careful about extending useful assistance into an earnings promise. A randomized field experiment with Kenyan entrepreneurs found no statistically significant average improvement from access to a GPT-4 business adviser. Outcomes differed according to businesses’ starting performance and how owners selected and implemented advice. [The researchers’ working-paper abstract](https://business.columbia.edu/faculty/research/uneven-impact-generative-ai-entrepreneurial-performance)

Columbia’s August 2026 discussion of that work describes a study involving 640 entrepreneurs. It emphasizes the role of judgment in deciding which recommendations fit a particular business. [Columbia’s research account](https://business.columbia.edu/insights/ai-widening-gap-between-high-and-low-performers)

That experiment did not test Muse. It involved an older model giving advice, rather than this agent executing tasks. It cannot predict Muse’s results. Its relevance is narrower: giving business owners access to AI does not establish that all of them will benefit financially, and an average can conceal very different experiences.

Execution could improve the picture if the main obstacle is following through on a sound decision. It could worsen it if weak advice becomes easier to act on. An agent that automatically implements a bad pricing strategy may cause damage faster than a chatbot whose suggestion is ignored.

The next evidence should therefore include failed attempts, costs and variation between users. Selected success stories would show what is possible. They would not show what a typical new user should expect.

## A business agent needs access, and access has a cost

Meta’s technical account makes two points that matter here. Sentinel controls external actions separately from the main agent, and Muse can still make mistakes or encounter attacks through the data it reads. The system is designed to limit the consequences rather than eliminate the possibility of failure. [Meta’s security design](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse)

For a small business, that means the value of automation depends partly on where authority stops. Preparing an advertising campaign is different from spending its budget. Drafting a customer offer is different from committing to a discount or delivery date. A process can be useful while requiring the owner to approve those commitments.

The privacy timeline also needs care. The launch version is Muse Secure VM. Meta says its operational policies restrict staff access, but do not technically prevent access needed to operate the service. Muse Confidential VM, intended to prevent Meta from accessing VM data cryptographically, is planned for later this year. [Current architecture and planned confidential mode](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse)

A business should not evaluate today’s service as though the promised confidential mode has already shipped. A customer list, supplier contract or margin calculation may be commercially sensitive even when it is not a password. The information an agent needs to negotiate well can be the same information an owner most wants to control.

Meta also says users can download files and memory, and can opt out of model-training use. It says conversations and VM data are excluded from its advertising systems, while agent browsing may influence ads indirectly. Those controls matter, but downloadable files alone do not establish that a complete working agent can be moved to another provider. [Meta’s data-policy explanation](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse)

That leaves a practical dependency to watch. Once an agent holds a business’s routines and connected workflows, a change in price, access or product behavior can create switching work. Keeping usable records outside the agent and understanding which services it relies on helps make that dependency visible.

## Competition could pass the savings to customers

If Muse lowers the effort required to produce a particular service, competitors may gain the same advantage. A seller who initially keeps the saving could later face lower market prices. Customers might receive much of the benefit, while suppliers need to complete more work or offer something harder to copy.

That would still be a substantial economic change. More affordable services can expand access. A small organization might finally commission work it previously went without. The error would be assuming that lower production costs necessarily become larger profits for everyone producing the work.

Businesses with trusted relationships, specialist knowledge or physical delivery capabilities may be better placed to use the agent productively. The software can reduce coordination around those strengths. A newcomer still has to develop a credible offer and convince someone to buy it.

There is a similar tension for workers. Removing routine administration can improve a person’s job, or allow an organization to redistribute work. The launch evidence does not establish how many jobs Muse will create or replace. Those effects depend on adoption, the tasks delegated and what organizations do with the capacity released.

Zuckerberg also describes future interactions between agents, while saying most of that is not part of this release. [Interview, 29:00](https://www.youtube.com/watch?v=Lx8lrn-cytc&t=1740s) If agents eventually negotiate and coordinate across businesses, the platform’s rules for identity, access and commercial preference will matter even more. That remains a development to evaluate, not an existing network whose benefits can already be counted.

## What would make the earnings claim convincing

Muse’s impact should be judged on outcomes that survive a closer look. Useful evidence would include:

- A clear baseline showing what the person or business did before using the agent.
- Additional revenue and verified savings reported separately, with subscriptions, advertising, fulfillment, refunds and review effort accounted for.
- Results across ordinary users, including unsuccessful projects and people who stopped using the service.
- An explanation of which tasks the agent completed, what the owner still did and whether the gains persisted.
- Published transaction charges and recommendation rules that make commercial incentives understandable.

Those requirements do not assume the product will fail. They make success interpretable. A modest improvement repeated across many businesses could be more consequential than a striking income screenshot with no account of its costs.

The plausible BIG CHANGE is wider access to practical assistance: more people able to carry out ideas, manage a business and keep up with the work around a household. Muse provides a concrete product through which that possibility can be tested.

Whether it makes users money will be settled further along the process, when a customer pays, the work is delivered and the costs are deducted. The agent’s activity log can show how much it did. The owner’s records must show what was left.

## Sources

- [Meta introduces the Muse personal agent](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) — September 8, 2026. Establishes the launch, initial US rollout and product claims. The personal agent is distinct from the Muse Spark model and Muse media-generation products.
- [Mark Zuckerberg on how Meta’s Muse AI agent will make people money](https://www.youtube.com/watch?v=Lx8lrn-cytc) — Sources Podcast, September 8, 2026. The money discussion begins around 24:53. Proposed transaction fees describe Meta’s intended business model, not a user payout commitment.
- [Alex Heath’s original interview page](https://sources.news/p/mark-zuckerberg-meta-muse-ai-podcast-interview) — September 8, 2026. The interviewer’s introduction provides recording context and his early impressions. The episode includes sponsorships, whose promotional claims are not adopted here.
- [How We Designed Muse](https://introducing.muse.ai/) — September 2026. Meta’s product designers describe persistent tasks, artifacts, browser access and approval controls. These capabilities explain possible mechanisms; they do not establish typical financial returns.
- [Stripe connects Muse to Link’s wallet for agents](https://stripe.com/newsroom/news/stripe-helps-meta-muse-shop-with-link) — September 8, 2026. Payment-provider confirmation of purchasing and approval functionality. This is a spending integration, not evidence of payments to Muse users or their business profits.
- [About Muse subscriptions](https://www.meta.com/help/subscriptions/1021145227643680/) — Checked September 22, 2026. Official Power and Max prices, weekly Muse allowances and availability limits. Product token allowances should not be equated with tasks or API-token budgets.
- [TechRadar’s firsthand Muse review](https://www.techradar.com/ai-platforms-assistants/i-tried-metas-new-muse-ai-agent-its-incredibly-useful-but-handing-it-my-digital-life-felt-deeply-uncomfortable) — Eric Hal Schwartz, September 14, 2026. Describes shopping and email assistance from one reviewer’s experience. It is not a controlled test of revenue, profit or typical user outcomes.
- [The Uneven Impact of Generative AI on Entrepreneurial Performance](https://business.columbia.edu/faculty/research/uneven-impact-generative-ai-entrepreneurial-performance) — Working-paper abstract dated July 1, 2024. Randomized business-advice research in Kenya supplies context about uneven outcomes. It tested a GPT-4 adviser, not Muse; its estimates are not predictions for Muse users.
- [Columbia: AI and the gap between high and low performers](https://business.columbia.edu/insights/ai-widening-gap-between-high-and-low-performers) — August 20, 2026. The researchers’ institution explains the experiment and the importance of selecting and implementing advice. Versions report differing subgroup estimates, so this article does not transplant percentages into a Muse forecast.
- [How We Built Safety Into Muse](https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse) — September 8, 2026. Meta’s technical account separates current Secure VM controls from planned Confidential VM protection. Security controls do not establish unbiased commercial recommendations.