Expense categories

Does AI Categorisation Send Your Transaction Data? A European Reader's Guide

A data-minimisation view of AI categorisation: which statement fields an expense app needs, which it does not, and how to test an app with euro examples.

Editorial illustration for Does AI Categorisation Send Your Transaction Data? A European Reader's Guide

Sometimes, and the sensible question is not "yes or no" but "which fields". Categorising a payment needs a merchant name and, at most, a rough size of the amount. It does not need your name, your account number or your balance. If an app can explain in one sentence what it sends to an AI model, and let you turn that off, you are in a much better position than with one that cannot.

What is the minimum an AI model needs?

Think of the information in layers, from least to most sensitive. Here is a worked example with a made-up payment of €86.40 to a hardware store.

LayerExample from the statementNeeded to guess "Home and DIY"?
Merchant (cleaned)Hardware storeYes
Amount band€50 to €100Helps a little
Exact amount and time€86.40 at 14:12No
Payer nameYour full nameNo
Account identifierYour IBANNo
Balance after payment€1,204.15No

A tool built on data minimisation sends the first two rows and keeps the rest to itself. You can ask any vendor to place its own product on this table.

Why do apps use AI at all?

Statement descriptions are messy, and no rule list covers every shop. AI fills the gap for the long tail of unfamiliar merchants. That is useful, but it is the last stage, not the first. In a well-ordered design, your own rules and your correction history handle most lines, and a model sees only what remains.

What does a well-ordered design look like?

SpendixAI is a useful reference here, though it is built around India and INR and not something to promise for euro users. Its order is: your own rules, then your history and similar merchants, then, only if AI is on, a model using the normalised merchant name and an amount band. AI can be off per feature, capped by a monthly budget, or turned off entirely, and unmatched items then wait for review. Optional features that read an unfamiliar statement layout do process the uploaded document, and that is stated rather than hidden.

How do I test an app before trusting it?

  1. Build a decoy file of fifteen invented lines in euros, such as €3.20 bakery, €41.00 pharmacy and €119.99 train pass.
  2. Import it with AI off and count the lines categorised without a model.
  3. Import a second copy with AI on and compare the extra lines labelled.
  4. Check that the two results agree where you would expect, and that nothing odd was guessed.
  5. Find the export and delete options and try them on the decoy data.

What should I never hand over?

Online-banking logins and one-time codes never belong in a tracker. A statement file you download yourself is enough for categorisation. If an app wants your banking credentials to "make it easier", that is a different trade, and one to think hard about.

What happens to the data after the model has answered?

A second question deserves as much attention as the first: is anything kept? A model might return a label and forget the request, or the vendor might log requests for a while. Ask whether prompts are stored, for how long, and whether they are used for anything beyond your own categorisation. A short, specific reply such as "not stored beyond the request" is a good sign; a long paragraph that never says yes or no is not.

Also ask where your own records live. A workspace that you control, with owner-level export and deletion, means a decision to leave is practical rather than theoretical.

Quick list before you commit

  • Which fields go to the model?
  • Can I switch AI off completely?
  • Do unmatched lines wait for me?
  • Can I export and delete?
  • Is there a spending cap for AI usage?

European data rules vary by country and change over time; this article does not interpret any of them. For rights and obligations, consult the official source or a professional.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports