AI categorisation can be private or not, depending on how it works. Rules you write yourself never need to leave the app. Learning from your own history happens inside your account. Only the last stage, sending an unfamiliar merchant to an AI model, raises real questions, and those are about what is sent, whether it is kept and whether it trains anything.
What are the different ways transactions get categorised?
- Rule-based: you or the app define that a merchant name always maps to a category. Nothing is guessed.
- History-based: the app looks at how you categorised similar items before and repeats it. Your corrections become the teacher.
- On-device: a model runs on your phone, so the text does not travel. This is possible but not universal, so ask whether it is the case.
- Server-side AI: the text goes to a model running elsewhere. This is where the privacy questions concentrate.
Most apps mix these. The useful question is not whether an app uses AI but at which step, and what leaves your account at that step.
Which questions matter most?
- What exactly is sent to the AI? A merchant name and amount band is very different from your full statement.
- Are my name, account numbers and balances included? They should not be.
- Is the data used to train a model, by the app or by the AI provider?
- How long is it retained, and can I request deletion?
- Can any person read my transactions, for example for quality checks?
- Can I switch AI off completely and still use the app?
If an app cannot answer these in its privacy page or help section, treat that as an answer.
What does a staged approach look like?
SpendixAI describes categorisation in stages. First come your own rules, then your history and similar merchants, and only if AI is switched on, an AI model for whatever is left. For that last stage it uses the normalised merchant name and an amount band, not your name, account numbers, full statement text or balances. AI can be switched off per feature, given a monthly AI budget or turned off entirely, in which case unmatched items are held for review. Your corrections become history, and you can add merchant aliases.
Does the AI learn from my data?
There are two meanings of learn. Inside your account, learning from your corrections is normal and useful: fix a label once and similar items follow. Separately, a provider might use your data to train a general model for everyone. Those are not the same thing, and you want a clear statement on the second. Look for wording about whether data is used for model training, and for switches that control it.
How do the approaches compare?
| Approach | What leaves your account | Main thing to ask |
|---|---|---|
| Your own rules | Nothing | Can I edit and delete rules? |
| History and similar merchants | Nothing, if done inside the account | Where is the history stored? |
| On-device model | Nothing, if truly on-device | Is it really local, for every feature? |
| Server-side AI | Some text, such as a merchant name and amount band | Is it retained or used for training? |
Be sceptical of vague promises such as your data is safe with us. Good wording is specific: it names what is sent, for what purpose and for how long. If you cannot find that wording, ask support in writing and keep the reply. It is reasonable to use rules and history only until you get a satisfactory answer, because those stages work without sending anything to an outside model.
A privacy checklist before you turn it on
- Read the AI and privacy page and note what is sent.
- Start with rules and history, and see how much is left for AI.
- Switch AI on for categorisation only if the remaining share is worth it.
- Set a monthly AI budget if one is offered.
- Review held or low-confidence items yourself for the first month.
- Re-read the privacy wording whenever the app announces new AI features.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

