Meta Launches Muse: What a Personal AI Agent That Shops and Books Means for Your Business

Meta picked a busy week to make its biggest product move of the year. On 8 September the company released Muse, a personal AI agent that does not just answer questions but actually gets things done: it books travel, fills in forms, negotiates bills and buys products, and it keeps working after you close the app. The launch of Meta Muse is the clearest sign yet that the big platforms see agents, not chatbots, as the next thing ordinary people will pay for.

For now the rollout is limited. Muse is available to adults in the United States, through a dedicated app on iOS, Android and the web, or simply through WhatsApp. Support for Meta's AI glasses is promised later. There is a free tier, but every account, free ones included, requires a payment card at signup. That detail says a lot about what this product is really for.

If you run a business, this launch is worth twenty minutes of your attention. A platform with billions of users has just handed ordinary consumers an assistant that carries a payment method and a browser. Sooner or later, some of the visitors on your website will not be people anymore. Let's sink in!

What Meta Actually Launched

Muse runs on Muse Spark 1.3, which Meta describes as its most capable model to date, built for real-world agentic work. Each user gets their own isolated virtual machine in Meta's cloud, with browser access, and the agent works there on its own schedule. You can hand it a task from the app or from a WhatsApp chat, close your phone, and get a notification when something needs your approval or when the job is done.

The examples Meta leads with are deliberately everyday ones. Muse can compare and book flights, fill in long forms, ring around to negotiate a bill down, or turn a saved recipe video into a grocery list and order the ingredients. None of this is science fiction on its own. The new part is that one consumer product does all of it in the background, connected to your email, calendar, payments, shopping and even smart home apps.

It is also worth noting what the launch took. The product was originally expected in April and Meta held it back for months to work on safety. Chief AI Officer Alexandr Wang framed Muse as the first practical step towards the "personal superintelligence" Mark Zuckerberg wrote about last summer, when he predicted that billions of people will have a personal agent within five years.

Pricing, and the Card Behind Every Account

The pricing has three tiers. The free plan includes up to 100 million tokens of processing per week, which is generous, and Meta's AI chief has said most users will manage fine on it. The Power plan costs $20 a month and Maximum sits at $100 a month for heavy use. Nothing unusual there, the numbers mirror what OpenAI and Anthropic charge for their paid tiers.

The more interesting decision is that a payment card is required to sign up at all. Muse handles purchases through an integration with Stripe Link, with purchase protections included, and support for Shop Pay and 1Password is promised. Meta says the agent itself never sees your card details or passwords, and every action it takes is written to an audit trail you can review.

Read those choices together and the strategy is plain. Muse is not built to be a better chat window. It is built to transact. Meta wants the agent to be the place where buying decisions get made, and it has wired in the payment rails from day one.

The Safety Layer, and Its Limits

Because Muse can spend money and write emails in your name, Meta built a gatekeeper system it calls Sentinel. It distinguishes actions that only read information from actions that change something, grants time-limited permissions, and requires explicit user approval for consequential steps like payments. A second monitoring agent reviews what Muse plans to do before it does it.

Meta is unusually frank about the gaps. The company admits the agent will make mistakes and that prompt injection, where a malicious website plants instructions the agent then follows, remains an unsolved problem across the whole industry. Reporting on the launch described rough internal testing too, including cases where the agent got around its own guardrails and one where private photos were exposed.

Privacy commitments are similar in shape. Meta pledges that Muse conversations are not shared with its advertising systems and that users can opt out of model training, but for now these are policy promises rather than technical guarantees. A Confidential mode, in which Meta itself cannot observe the agent's workspace, is planned for later in 2026. Until then, businesses handling sensitive information through an employee's Muse account should treat it like any other third-party cloud service.

When the Customer Is an Agent

The part of this launch that matters most for business owners has little to do with Meta's app itself. It is the traffic. If even a small share of Muse's users delegate their shopping and booking, your website starts receiving visitors that read pages in milliseconds, never see your banner, and compare your offer against three competitors before a human ever looks at the screen.

Some companies are already pushing back. The restaurant booking platform Resy discourages automated agent access, and plenty of sites block anything that looks like a bot. That is an understandable reflex, but it carries a cost: if an agent cannot complete a booking or a purchase on your site, it quietly completes it somewhere else. The user never knows your business was even an option.

The alternative is to treat agents as a new customer segment. They reward the boring virtues: clear product data, honest pricing on the page, standard checkout flows, structured markup that machines can parse. Sites that already invested in accessibility and clean SEO are, somewhat by accident, the best prepared for agentic visitors.

Practical Steps for Businesses Building Digital Products

First, test your own site the way an agent would use it. Try completing a purchase or an enquiry with keyboard only, check that prices, availability and delivery terms are stated in plain text rather than hidden in images, and make sure your structured data is valid. These are cheap fixes, and they serve human customers too.

Second, look at your checkout rails. Muse pays through Stripe Link today, with Shop Pay coming. If your store already supports these, an agent can pay you without friction. If your checkout depends on CAPTCHAs at every step or on custom payment widgets, you have effectively put up a wall, so decide consciously whether you want that wall.

Third, make a deliberate bot policy instead of an accidental one. Decide which automated visitors you welcome, which you rate-limit and which you block, and write it down. And keep an eye on your analytics: agent traffic will distort conversion rates and session times, so the numbers you report to management need a footnote sooner than you think.

Finally, think about presence. Muse lives inside WhatsApp, a channel many businesses already use for support and sales. A business that keeps its WhatsApp catalogue, opening hours and replies in good order is easier for an agent to deal with, and easier to recommend.

The Bigger Agent Race

Muse does not arrive in an empty field. OpenAI, Google and Anthropic have all shipped agentic features this year, and the industry has spent the last week discussing unsupervised agent swarms after researchers found OpenAI agents editing a public wiki for a month unnoticed. Meta's bet is different in one respect: distribution. It does not need anyone to download a new habit, because the agent sits inside WhatsApp, an app a third of the planet already opens daily.

The absence on the map is Europe. The US-only launch is no accident, since the EU AI Act's obligations for this kind of product are exactly the sort of thing a cautious legal team wants to study before shipping. European businesses get a grace period to prepare, and they should use it, because these launches have a habit of crossing the Atlantic within a year.

Final Notes

It is easy to dismiss consumer agent launches, and the first weeks of Muse will certainly produce funny screenshots of booked flights nobody wanted. That happened to every early product in this category. The direction, though, is not really in doubt anymore: the biggest consumer platforms are all converging on assistants that act, spend and negotiate on their users' behalf.

For businesses, the sensible response is neither panic nor a shrug. Make your site and checkout legible to machines, decide your policy on automated customers, and watch how the first wave of Muse users actually behaves. The companies that treated mobile traffic seriously in 2010 did not regret it. Agentic traffic in 2026 looks a lot like that moment.