“To put this in perspective, this is more than twice the level that Threads started at or where we believe that other AI apps started at as well,” Zuckerberg said.

Muse can do multi-step tasks on behalf of users such as browsing the web, sending emails, and booking travel. One-quarter of Muse adopters use it daily, and the most common use case is shopping, Zuckerberg said, with an average purchase size of over $200. Demand for Muse was so high that Meta had to work with payment processor Stripe to raise credit limits in the first week, though he didn’t specify by how much. Meta had partnered with Stripe to enable Muse purchases from the start. 

“We had to work with Stripe to raise credit limits since the volumes were above what any of us expected,” Zuckerberg said. 

This early shopping activity offers Meta an opportunity to turn its billions in AI investments into another revenue stream. Rather than just selling ads to businesses hoping to reach potential customers, Meta wants to collect fees when Muse helps land a purchase. Zuckerberg said Meta plans to charge businesses a fee when Muse helped them complete a sale. Using the agent will be free for consumers unless they want to use more tokens or more advanced models. 

This can also give Meta a revenue stream beyond advertising, which generated $63.1 billion in the third quarter. Zuckerberg said Meta earns “several hundreds of dollars annually” from its family of apps for each American user, and he expects Muse to generate even more revenue per user as shopping activity grows.

“We expect that Muse will reach an even higher level over time due to the amount of commerce that we already see taking place,” he said.

Meta did not respond to Fortune‘s request for comment, and Stripe declined to comment.

A future without shopping sites?

Meta’s ambitions for Muse go beyond helping customers shop online. The tech giant envisions a network of people’s personal AI agents interacting with businesses’ AI agents instead of navigating through a website. 

Meta CFO Susan Li said the company is working with partners including Shopify and Walmart to develop a standard called the Personal Agent Protocol to define how consumers’ agents can securely interact with a business’s agents.

For retailers, this could hand over control of the shopping experience to Meta. Instead of discovering and browsing websites themselves, customers could delegate those tasks to Muse. That could help businesses find customers, but it also gives Meta greater influence over which sellers consumers discover and where they ultimately spend their money.

Li acknowledged adoption “will likely be uneven” among businesses wary of giving up their customer relationships. But she said at some point, when agents demonstrate they can operate safely, “more and more businesses will be excited to participate.”

“We expect the market, ultimately, to evolve away from agents attempting to navigate every digital experience like a human,” Li said. 

The costly business of AI shopping

Meta faces another obstacle to its vision of agentic shopping: Muse is expensive to run.

The tech giant spent almost $35 billion on capex in the third quarter, driven by AI servers, data centers, and network infrastructure, and it reported negative free cash flow of $2.1 billion, according to its Q3 earnings report. Li said on the call Meta anticipates capex to be $135 billion to $145 billion for the full year, compared to $72.2 billion last year. 

Zuckerberg said Muse’s computing demands are partly why Meta has limited its launch to the U.S. and Canada. The company wants to eventually serve more than 1 billion Muse users, but reaching that scale requires more compute efficiency.

“In the near term, Muse is very compute-intensive to run, but we expect to improve its cost efficiency by more than 10x in the next year,” Zuckerberg said. 

Meta has traditionally focused on attracting users to new products before finding ways to make money from them. But Zuckerberg told analysts Meta expects to start working on monetizing Muse earlier than it typically does with new products, citing the cost of running the agent and the amount of shopping already taking place. 

“We are going to work on monetization somewhat sooner on this than our normal playbook,” he said. “In doing so, I think we will be able to build a model where we get a percent of the value that we are helping to create for people.”