Often, yes: some part of your transaction data leaves your phone when an app uses an AI model to label spending. What matters is which part. A tool that sends only a cleaned-up merchant name and a rough amount range is a very different thing from one that ships your full statement, your name and your balances to a third party. Before you upload anything, you can find out which kind you are dealing with in about ten minutes.
What does "AI categorisation" actually do with a transaction?
Take one line from a statement: a $47.30 debit described as a long string of letters, a store number and a city. Categorisation means deciding whether that is Groceries, Fuel, Eating out or something else. A well-built app tries cheap, local methods first and calls an AI model only for lines it could not settle. The question for you is what the app hands to the model when it does.
Ask for the payload in plain words. "We send the transaction" is not an answer. "We send the merchant name and a band such as $25 to $50" is.
Which fields should and should not leave the app?
| Field on your statement | Needed to guess a category? | What to expect |
|---|---|---|
| Cleaned merchant name | Yes, it carries almost all the signal | Reasonable to send |
| Amount, as a range | Helps a little, for example $4 versus $400 at the same store | A band is enough; the exact cents are not needed |
| Your name and address | No | Should stay out |
| Account or card numbers | No | Should stay out |
| Running balance | No | Should stay out |
| The whole statement text | No | Should stay out unless you chose a feature that reads the document |
That last row matters. Some optional features, such as reading an unfamiliar statement layout, do have to process the document itself. That can be fine, as long as it is a feature you switch on rather than a default you never saw.
Can I switch it off and still get categories?
You should be able to. Look for an off switch, ideally one per feature, and find out what happens to lines the app cannot place: do they quietly land in a catch-all bucket, or wait in a review list for you? Waiting is better, because you stay in charge of the label.
SpendixAI works this way for readers in India: your own rules run first, then your history and similar merchants, and only then, if AI is on, a model sees what is left, using the normalised merchant name and an amount band. AI can be turned off entirely, and unmatched items are held for review. It is built around India and INR, so treat it here as an example of the answers to look for, not as something to sign up for from the US.
A ten-minute test you can run yourself
- Open the privacy page and the settings of the app. Write down every place a third party is mentioned.
- Find the AI setting. Turn it off and import a small test file of about 20 lines, such as $12.40 coffee, $63.15 groceries and $9.99 streaming.
- Count how many lines were categorised without AI. Whatever is left should wait for you.
- Turn AI on for one feature only, import a second small file, and see what changes.
- Check you can export your data and delete it, and read how long deleted items hang around.
What if we cannot find a clear answer?
Treat silence as information. A vendor that handles data carefully usually says so in specific terms: what is sent, to whom, whether it is kept, and how to stop it. Vague wording like "we may share data with partners to improve services" is a reason to ask a direct question by email and wait for a direct reply.
Rules about personal data differ by state and by company, and they change. This article does not cover any of them; for a legal question, go to the official source or a professional.
Your closing checklist
- I know exactly which fields go to the AI model.
- I can turn AI off completely, or per feature.
- Unplaced lines wait for my review.
- The app never asks for my bank login.
- I can export my data and delete it.
- I tested with a decoy file before real statements.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

