You have a dentist appointment next Tuesday. Last month's credit card statement shows an unpaid balance from that same dental office. Your inbox holds a reimbursement notice from your insurer, due this week. A human connects these three facts and draws a conclusion: file the insurance claim before Tuesday. Taken separately, they're just an appointment, an unpaid bill, and an email. It's the connection that produces the judgment.
Muse, the personal AI agent Meta unveiled in September, claims it can automate that connection. It asks, in a single request, for access to your email, calendar, payments, and health services. To understand why it asks for all of them at once, you first have to look at how an agent like this actually has to work.
One Data Source Doesn't Make Context
An agent that only sees your calendar can tell you when your next meeting is. One that only sees your email can count your unread messages. One that only sees your payment history can total up last month's spending. None of that goes much beyond what a smart notifications app already does.
Connect the three and things change. If an agent knows who's on your calendar for tomorrow's morning meeting, has the email thread you've exchanged with that person, and sees the related charges from the past two months, it can pull all of that into one screen right before the meeting — a summary of the relevant conversation plus anything still unresolved. That's work a person would otherwise have to do by checking three separate places and connecting the dots themselves.
Health data belongs in the bundle for the same reason. An agent can flag whether a medication schedule overlaps with a calendar meeting, or check that a new checkup booking doesn't clash with an outside appointment already on your calendar. The more domains an agent can cross, the more it can do — and crossing domains requires data from each one.
An AI agent's usefulness comes not from how many types of data it holds, but from how many connections it can draw between them. Every new data type multiplies the possible connections to the ones already there. The bundled request simply follows that math.
Combination Doesn't Just Create Features
Combining data creates features. It also creates something else at the same time.
Calendar data alone reveals where you go and how often. Email data alone reveals who you keep in touch with and how. Payment data alone reveals what you spend money on. Health data alone reveals your habits and physical condition. Bring all four together in one place, and you get a behavioral profile. Combine when, where, and with whom you meet, what you buy, and what ails you, and ad targeting reaches a whole new level of precision.
Advertising accounted for more than 97% of Meta's total revenue in 2024. That level of dependence on ad income, on the part of the company running Muse, hints at where the data this agent collects could end up being used. Granting permissions with a clear-eyed view of a platform's incentives is a different decision than granting them without one.
Trust issues arise from the gap between what a terms-of-service document says and what actually happens in practice. In the 2018 Cambridge Analytica scandal, data from roughly 87 million Meta users was funneled to a third party through channels the terms of service never disclosed. The investigation that followed led the US Federal Trade Commission to fine Meta $5 billion in 2019. The central issue was the mismatch between what the terms said and how the data was actually handled.
As an agent combines more domains, both its usefulness and the precision of the profile it builds rise together.
What to Check When You Get a Permissions Request
Muse isn't the first to do this. Apple Intelligence runs on iCloud Photos, Notes, and Calendar. Google Gemini connects to Gmail, Drive, and Maps. Microsoft Copilot draws on a company's entire email and document store. Future agents will keep asking for permissions the same way.
The first thing to check on any permissions request is where the data is stored. Whether processing stays on your device or gets uploaded to an outside server determines the exposure path. Some Apple Intelligence features process data entirely on-device and never touch a server. Meta hasn't published details on where Muse stores its data.
Next is whether you can grant permissions item by item. Check whether it's all-or-nothing — accept the entire bundle or reject it entirely — or whether you can, say, allow the email-calendar link while excluding health data. Handing over extra data for a feature you'll never use is a trade that only benefits the other side.
Last, check the platform's revenue model. Subscription-based and ad-based platforms have different incentives for how they handle data. A subscription business uses data to improve service quality. An ad business uses it to improve service quality and ad-targeting precision at the same time. Neither model is categorically better. Knowingly deciding which kind of platform you're handing your data to produces a different outcome than simply going along with the convenience.
What to Delegate, What to Hold Onto
For a one-person business, the real value of an AI agent lies in automating repetitive tasks: sending the same quote template every month, pulling together and summarizing relevant materials before a meeting, logging deadlines onto a calendar. Hand these off to an agent and you get your time back. But making that delegation work requires giving the agent access to your data. You can't block access and still expect the convenience.
The practical question, then, is where to draw the line. Data whose exposure to a third party costs you nothing — the kind involved in repetitive tasks, scheduling, and information lookup — is fine to connect. Data whose exposure carries real downside — contract terms with clients, personal health records, financial information — should only go in after you've confirmed the feature genuinely requires it.
An agent speeds up your hands, not your judgment. Muse can draft an email, but a person still decides what goes in it. It can propose a schedule change, but a person still decides whether to accept it. Drawing a clear line between what an agent can handle quickly and what requires you to hold the reasoning yourself is the practical instinct you actually need when using a personal AI agent.
How Muse actually performs once it ships is something we'll only know by using it. But it's already safe to guess that bundling requests for email, payment, and health data isn't a Muse-specific design choice — it's going to be the common grammar of every agent that follows. Build the habit of checking three things whenever a bundled request lands in front of you — where the data is stored, whether you can choose item by item, and what the platform's revenue model is — and you won't be caught off guard when it does.
