A bank statement may be the most honest diary you will ever keep, because you never intended for it to be one. Transactions can reveal a detailed biography: a repeated pharmacy charge, a weekly payment to a counsellor, a recurring donation.

Meta launched Muse on September 8, and within two weeks, it became the top free app in the US. One of the first ideas it suggests to users is: “I can find and cancel those random subscriptions you forgot about.”

As I just so happen to have built a subscription tracking app, I took notice. One of the biggest companies on Earth just offered that service for free.

To do this, Muse asks you to upload a bank statement, or grant it direct access to your accounts. In other words, Meta wants to read that private unfiltered diary.

Carrying is not reading

Historically, the concern with third-party finance apps was the “middleman.” You trust your bank, but you often don’t know the third-party aggregator acting as a courier between your bank and an unaffiliated financial app. These aggregators typically keep a de-identified copy of your data for marketing segmentation.

An AI agent is fundamentally different. Traditional apps take in data, do some math, and return meaningful results. An agent reads your data, remembers it, and acts on it. Meta states plainly that Muse “remembers what matters to a person, so it can make suggestions unprompted and act on details that person only mentioned once.”

An aggregator carries your diary to a third party where it becomes a data point. An agent reads every page of that diary, enriches it with what it already knows about you, and keeps notes for later use.

The first job is the most revealing one

To identify “random subscriptions,” an agent needs to parse your entire financial status. There’s no way to find streaming charges without also reading the pharmacy ones. When associated with your inbox, calendar, and messages, this data reveals even more: potential health conditions, your location on specific evenings, or your personal beliefs.

Meta already knows how sensitive this information is. In 2022, they stopped allowing advertisers to target users based on health, religion, or politics because of these exact inferences. They have been capturing and managing those inferences for years. An agent with your statements would have far richer material than any ad system ever did.

Something else I noticed when I installed Muse is that it asks you to approve a default authorization of a recurring monthly scan to catch new charges and price changes. Muse doesn’t just read your statements once; by default, it keeps reading your financial activity in perpetuity until you tell it to stop.

”Meta says”

Meta attests to substantial guardrails: each agent runs on its own virtual machine, Muse cannot see passwords, and conversations aren’t shared with advertising systems. They even promise a private mode later this year where even Meta cannot read your data.

But these are attestations, not architectures. They require you to trust Meta to do what it claims.

In 2019, the FTC imposed a $5 billion penalty on Facebook. One of the charges was that they collected phone numbers for security but used them for advertising without disclosure.

I was working in identity management at the time and that settlement put the fear of God into the entire industry.

A privacy protection you cannot verify is a marketing claim, even if the founder and CEO is the one making it. Currently, Meta is offering a “trust me bro” description of its security.

Follow the money

Muse is free for most people, but Mark Zuckerberg has stated Meta expects to profit over time by taking a small fee from transactions. A company taking “a piece of the action” performs better the more it knows about your past and future spending habits.

A model that earns a fee per transaction is rewarded for every purchase you make. That’s an incentive, whatever anyone intends.

Furthermore, your conversations train its models by default. Most people never change default settings, and product designers know this.

This isn’t illegal, but it is unprecedented. We are watching the trade of privacy for convenience happen out in the open, one task at a time.

What this means for sistnt

I’ll be honest: for most people, a free agent that cancels subscriptions in thirty seconds beats an app where the user has to manually type them in. For a few days, it looked like the end for my subscription app, Steward.

But I believe Muse makes Steward more valuable.

I am not against AI agents - the good ones are amazing. What I am against is the assumption that an agent should have access to all of your data and the authority to act on your behalf, with a business model tied directly to your bank account.

sistnt’s bet has always been that software can run on the device you own, keep nothing it doesn’t need, and be paid for once rather than by subscription. Muse makes that bet sharper.

The question for the next decade isn’t whether you’ll use an agent - you will. The question is where the agent lives and who else is in the room with you. Software that runs on your own phone and keeps its notes there can pass that test.

That is the kind of software I want to exist, and the kind I build, at cottage scale - not a planetary one.

Own your tools. Mind your data. Be persistnt.

Sources