It can, which is why the best answer is a trial you run yourself rather than a promise you read. Upload a trimmed, sample version of a statement, compare results with AI off and on, and look at what the app says it sends. After that, switching AI on becomes a decision you made, not a default you drifted into.
Step 1: what should I trim from the sample?
Make a copy of one statement and remove what a categoriser does not need. Delete your name and address from the header, and any account identifiers. Keep the lines. Better still, type a sample of fifteen made-up lines: A$5.80 coffee, A$132.40 groceries, A$68.00 fuel, A$24.99 streaming. The app does not care whether they are real.
Step 2: how do I import with AI off?
Find the setting and switch AI categorisation off entirely, or per feature if that is offered. Import the sample. Note how many lines have a category and how many are waiting. In a sound design, unmatched lines wait for review rather than being filled in silently.
Step 3: what changes when AI is on?
Import the same sample again with AI on, into a separate test workspace if you can. Compare the lines that were previously waiting. Are the guesses sensible? Would you have chosen the same label?
| Result | What it suggests | Next move |
|---|---|---|
| Few lines wait, guesses fine | Rules and history do most of the work | AI adds little; consider leaving it off |
| Many lines wait, guesses fine | AI saves you time | Set a monthly cap and turn it on |
| Many lines wait, guesses poor | Your merchants are unusual | Write rules and keep AI off |
| App cannot say what it sends | Missing transparency | Ask in writing before going further |
Step 4: what should I ask the vendor?
Ask which fields reach the model. For comparison, SpendixAI states that it sends a normalised merchant name and an amount band, not your name, account numbers, full statement text or balances, and that AI can be turned off entirely. It is built around India and INR, so it is not something to promise to a reader in Australia; it is just a yardstick.
Does it matter which device I use?
Ask whether the app stores your records in a workspace you can reach from several devices, and who can see them. Role-based access and the ability to export and delete are the features that matter more than the device.
What do I do if the answers are unclear?
- Keep AI off and categorise with rules.
- Import less, for example only the lines you need.
- Email one specific question and keep the reply.
Australian privacy obligations are outside this article. For rights and legal questions, use the official source or a professional adviser.
Should I use real data in the trial at all?
Not at first. A trial with invented lines tells you how the app behaves, and costs you nothing if the behaviour is poor. Move to a single real month only when the vendor's answers about what is sent have satisfied you, and even then consider leaving out the months you would least like exposed. You can always import more later; you cannot easily take an upload back.
It also helps to repeat the check after any large update to the app. Settings sometimes change, and a toggle you turned off last year deserves a glance today. Put a reminder in your calendar twice a year: open the AI setting, confirm it is as you left it, and read the privacy page for changes.
Your test-plan checklist
- Sample file prepared, with no personal header.
- Imported with AI off and results noted.
- Imported with AI on and results compared.
- Vendor's answer about sent fields recorded.
- Export and delete tried on the sample.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

