Muse turns Meta’s AI model into a product

Muse is Meta’s personal agent, running on Meta’s own language model. The model gives it its abilities, and the product around it makes them usable for ordinary people. How far Muse gets also depends on trust in Meta and on the shops and banks that let it in.

A large language model is a powerful general-purpose tool, able to take on almost any kind of task. That breadth makes it hard for an ordinary person to get a specific result.

For AI to sell to the mass market, it has to do something useful

That takes a product built on top, such as an assistant or an agent, designed around particular jobs. The model becomes its engine. Muse is Meta’s version.

The product turns what the model can do into actions. Muse sends email, books travel, fills in forms, buys things and keeps working after the app is closed.¹ To do that it connects to the user’s Instagram, WhatsApp and email, to 1Password and to Stripe.¹ Most use is free, and the paid tiers are priced by how much Muse processes.²

Its engine is Muse Spark, and a newer engine, Muse Wonder, is on the way.³

Meta has two reasons to turn its model into a product.

• The first is usability. Meta’s chief executive says configuring and running an agent is beyond ordinary people.⁴ The head of product for Meta’s superintelligence lab describes Muse as the kind of agent developers already run, made easy and scaled to billions.⁵ A reviewer credits Muse with getting the consumer experience right.⁶
• The second reason is money. Margins are higher on the finished consumer product than on the engine inside it. Meta plans to make money from fees on purchases, subscriptions and personalised ads around Muse.⁷

An agent that spends money and sends email for someone needs limits, and Meta builds them into the product

People choose which apps Muse may use,⁸ and it asks before it sends an email or pays for anything.⁹ Passwords and card numbers stay out of its reach, and a separate agent checks everything Muse sends to the internet.¹

Security reviewers credit these limits, more than the model, with keeping accounts protected.²

Muse runs on trust in Meta

Those limits matter because Muse asks for access to email, calendars, payments and health apps, more personal data than Meta’s apps have asked for before.¹⁰

Any service that holds such data depends on trust in the company behind it. For Meta, that trust may be the weakest part of Muse.

Meta paid a record penalty to the US Federal Trade Commission for privacy violations, and the Cambridge Analytica scandal put information on millions of Facebook users in a third party’s hands.¹⁰ What the company says about safety is also measured against two recent cases. A jury found that Meta misled the public about child safety on its platforms,¹¹ and the company agreed to pay up to $18 billion to settle state claims over addictive products for children.¹²

The doubt shows in the numbers. In a survey of US consumers, 8% said they would trust Meta with their passwords, against 30% for Google and 23% for Apple.¹³

The stakes are high because of what an agent can reach

Muse holds private data, reads outside content and can send data out. Security experts see that combination as the opening an attacker needs to turn an agent against its user.² Muse’s own short record adds to the doubt. It keeps files on people in the user’s contacts, including people who have never used it.¹⁴ Meta paused the sharing feature when shared Muse conversations turned up on the open internet,¹⁵ and turned it back on after a fix.¹⁶ In one test, human call-centre workers made calls that Muse presented as its own, until Meta stopped the practice.¹⁷ UK financial regulators are now coordinating their response to Muse.¹⁸

Meta points to Muse’s security design and promises to keep Muse conversations and data out of its advertising.¹⁰ Yet it also plans to make money from personalised ads around Muse, which makes that promise harder to believe.⁷

Conversations are used to train Meta’s models unless the user opts out, and Meta can technically access the machine Muse runs on.²⁸

A head start from Facebook

Even so, sign-ups came fast. Muse passed ten million accounts, about a quarter of them in daily use, with shopping the top activity.¹⁹ Its users come back more often than those of rival apps.²⁰

How much of that growth the product earned is hard to tell.

Coverage of the launch credits it to work on the product,²¹ while Meta’s chief executive credits the newer model, Muse Wonder, with better engagement and retention.⁷ Nearly all Muse users are already on Facebook,²² and Meta promotes Muse inside its own apps.²³

Facebook’s reach heavily skews Muse’s adoption figures.

Shops and banks decide who gets in

That reach ends at the door of every shop and bank where Muse has to act, and many now close that door by default.

Amazon blocks Muse as an unauthorised agent.²⁴

Large US banks have found unapproved agents inside customer accounts,²⁵ and several now stop outside software from logging in with a customer’s password.²⁶ One of them names Muse among the agents it blocks.²⁶

Some doors stay open on terms set by the other side. A few banks let approved providers read customer data.²⁶ One bank offers read-only connections of its own.²⁷ Meta has signed deals with retailers such as Walmart, Best Buy and Sephora.²⁷ In the EU, account-information and payment services need a licence to connect to bank accounts.²⁷

Meta’s finance chief expects businesses to set a common standard for how agents get in.²⁸

Each route depends on someone’s approval, whether a place on a bank’s list, a licence or a deal with a shop. That favours companies big enough to secure such terms or to run the whole ecosystem themselves.

The margin Meta expects from Muse therefore rests on trust in Meta and on doors that other companies open. Both lie outside the model.