Expense categories

Does AI Categorisation Send Your Transaction Data? A Canadian Rules-First Approach

Cut what an AI model ever sees: build rules and corrections so most lines are labelled locally. Worked example in C$ with a ten-minute privacy test.

Editorial illustration for Does AI Categorisation Send Your Transaction Data? A Canadian Rules-First Approach

In many apps, yes, some data goes to an AI model, but you decide how much by how you set the app up. The fewer lines that need AI, the less data ever leaves. A few rules and a habit of correcting mistakes can handle most of a month's statement before any model is asked anything.

How much of my statement actually needs AI?

Say a month has 200 lines. Suppose your own rules label 120 of them, such as rent, the phone bill and the regular grocery chain. Your correction history handles another 50, because you labelled similar merchants earlier. That leaves 30 lines, or 15 percent, for which an AI model might be asked. The numbers are invented, but the arithmetic is the point: 200 minus 120 minus 50 is 30.

Next month, after you correct the 30, perhaps 18 remain. The pile shrinks every month you keep up the habit.

Which rules are worth writing first?

  • Fixed bills: rent, internet, phone. These repeat exactly, so one rule per merchant.
  • Transit and fuel: one rule each covers many lines.
  • Your two or three usual supermarkets.
  • Subscriptions you recognise, such as C$16.99 music or C$9.50 cloud storage.
  • Anything you have mis-categorised twice; stop correcting it by hand and write the rule.

What does the AI see for the leftover lines?

A careful design sends a cleaned merchant name and an amount band, not your name, account numbers, balances or the whole statement. That is how SpendixAI describes its own approach. It is built around India and INR, so mention it here only as a benchmark for what a specific answer sounds like. Ask any app you are considering for the same level of detail.

StageWho decidesLines in the example
Your own rulesYou120
History and similar merchantsThe app, from your past corrections50
AI model (if switched on)A model, using merchant name and amount bandup to 30
Held for review (AI off)You30

What if we do not want AI at all?

Switch it off. A fair app then holds the unmatched lines for you instead of guessing. In the example above you would review 30 lines in a few minutes, and your corrections would shrink the next month's list. You lose nothing except a little speed.

What else should I check?

  • A monthly cap on AI usage, if the app offers one.
  • Whether reading an unfamiliar statement layout is a separate, optional feature that processes the document.
  • How to export and how to delete.
  • Whether the app asks for banking logins. It should not.

Canadian privacy rules and your bank's policies are outside this article. For those, read the official source or ask a professional.

What does a rule actually look like?

A rule is a plain statement: if the merchant contains this text, use that category. For a C$1,850 rent payment to a landlord, the rule might say "contains landlord name, Category Rent". For a C$62.40 weekly grocery shop, "contains the supermarket's name, Category Groceries". Once saved, those lines never reach AI again, because the first stage settles them.

Spend twenty minutes writing rules in your first month and the benefit repeats every month after. Corrections work the same way: each time you fix a wrong label, the app remembers, so a similar merchant next month is placed without help. The routine, not the technology, is what protects your data.

  • Write one rule per recurring payee.
  • Fix mislabelled lines straight away, while you remember the reason.
  • Review the leftover list weekly so it never grows into a chore.

A simple routine

  1. Import the month.
  2. Correct unplaced lines and add a rule for any repeat.
  3. Glance at the AI setting once a quarter and confirm it is where you left it.
SPENDIXAI JOURNAL

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

Explore your reports