Who Does Your AI Work For?

Last week I argued that a money app should not be built to keep you, and that the Duty of Loyalty in the new Consumer Reports framework names the right expectation. What that post did not do is hand anyone a way to check whether a company is meeting it.

That gap bothers me, because the claim itself is free. Every company will tell you it puts your interests first, and there is no way to audit the sentence from outside the building. You cannot read the internal metrics. You cannot see which number the team is graded on. The roadmap you are shown is the one somebody decided to show you.

Structure is the part you can see. How a product makes money, what it hands you on the way out, and whether the AI reading your finances is one you chose are all visible from where you are standing. Intentions stay private. Incentives leave marks in the design. So the rest of this post runs those checks against my own product, starting with the one underneath everything else.

The revenue model shapes every other answer

The standard carries no legal force. Consumer Reports is an advocacy organization, and I flagged that last week, so I will not lean on it as if a regulator wrote it. What the document has is precise language, and the second of its nine principles is Privacy and Data Minimization: "Consumer data is collected, used, and shared responsibly and proportionately, only to the extent necessary for service delivery and compliance requirements (as opposed to generalized commercial purposes), and consumers retain meaningful and frictionless control over how their data is handled." The clause in parentheses is the whole test. Generalized commercial purposes covers nearly every way a finance app turns your transaction history into revenue that does not come from you.

A free tier deserves the hardest look here. Running this kind of product costs real money every month, so if you are not the one paying, something else is, and the transaction record sitting in the database is the most valuable asset in the building. That pressure exists on any app in this category, mine included, and a free plan does not make it go away.

The Trupocket answer is published rather than promised. The privacy policy states that your information is used only to provide Trupocket services and that "we NEVER sell your data". Subscriptions are the revenue model today, which means there is no data buyer anywhere in the loop for the product to be designed around. A line written into a policy document can be held against me later. An intention cannot.

The exit has to exist before you need it

I spent a full post on this in who owns your financial data, so the short version fits here. What matters is what your app actually hands you when you decide to go.

For Trupocket the real exit is the API. 60+ endpoints with full read and write access, the same ones the web app itself runs on, so anything you can see on a screen can be pulled out by a script. Your financial data should work for you covers why the product was built in that order. That post predates the AI connection shipping in June and still describes it as upcoming, so I would take the portability argument from it and nothing about the AI layer.

If you do not write code, the privacy policy commits to a copy of your transactions, accounts, budgets, categories, hashtags, payees, and settings in machine-readable JSON, sent on an email request, within 45 days and usually within seven, once every 12 months at no charge. The self-serve version, listed as Data Export, is still In Progress on the roadmap with no launch date attached. Both of those things are true at once, and I would rather say so than round either one off.

The assistant should be replaceable without the data moving

This is the check I care most about, and most apps cannot answer it at all. If the AI reading your finances is welded into the product, every question about whose side it is on collapses back into whatever the company that built it says today. You are trusting a stated intention again, which is where this post started.

Model Context Protocol changes the shape of that problem. It is a published standard that lets any AI assistant read from a service somebody else built. MCP turns the assistant into a component you picked rather than a fixture you inherited. Trupocket has run an MCP server since June 11, and the mechanics of connecting one are covered in the post from launch day, so I will skip them here. The part that matters for loyalty is the arrangement underneath. Your transactions, budgets, and accounts stay in your Trupocket account. The assistant looking at them is a client you point at that account and can unpoint. If you pick one today and want a different one in six months, your financial history stays where it is and the setup does not get rebuilt.

I want to be exact about the boundary. An assistant connected over MCP on Premium can pull your transactions and run your reports, and that is where it stops. On Pro and above it can write too, using scopes you pick when you connect it. The API I described above is the surface you drive yourself with code you wrote. Neither one moves money, because Trupocket holds no money to move.

The honest footnote is price. MCP access starts on Premium at $2.99 a month, so the swap I am describing sits behind a paid tier rather than being something every account has. The Free plan includes API access capped at 20 calls a day, which is enough to try the API out and nowhere near enough to pull a transaction history out in bulk. An argument that you can leave and you can swap gets weaker if the tier it lives on goes unmentioned, so there it is.

Why I would rather be checked than believed

None of this requires anyone in this business to be a villain. A team whose revenue depends on holding your data can believe every word of its own mission statement and still ship something that quietly points the other way, because that pressure shows up in defaults and small design choices rather than in anyone's stated intent. So instead of a paragraph about values, I put the answers in a privacy policy, a pricing page, and a roadmap that separates what shipped from what is still being built. Those can be pulled up in a year and compared against what the product actually does by then. If you want to run these questions against Trupocket yourself, the app is right here.