To find unusual transactions, an assistant has to read transaction rows, so the privacy question is how much it reads and for how long. The careful approach is to do the rule-based checks inside the app first, share only the few rows still in doubt, use read-only access, and disconnect afterwards. You can get a useful answer without handing over a whole year.
What is different about this task?
Spotting oddities needs detail: dates, merchant names and amounts. That is more revealing than a total. A monthly total can tell nobody much, but a list of merchants tells a story about where you shop, who you pay and when. So this task deserves a more deliberate setup than a simple total.
Start inside the app, where no assistant is involved
SpendixAI has a Possible duplicate charges report that lists the same amount to the same merchant within two days, and it flags possible duplicate imports. Run them first. Say a month has 140 rows and the report lists three pairs. You now know six rows are in doubt out of 140; you can decide whether anything else needs an assistant at all.
How to limit what the assistant reads
- Pick a narrow window, one month rather than twelve.
- Connect with the narrowest scope and with read-only.
- Ask about specific patterns, such as duplicates and unusually large amounts, rather than 'review everything'.
- Do not paste statement text into the chat. A scoped connection is easier to withdraw than a pasted file.
- Disconnect when finished.
A worked comparison
| Approach | Rows the assistant reads | What you learn |
|---|---|---|
| Ask about all of the year | About 1,700 if you have 140 rows a month | More than you need |
| Ask about one month | About 140 | Enough to find a pattern |
| Ask about the three flagged pairs | 6 | A direct answer to your doubt |
Twelve months of 140 rows is 1,680 rows, so 'about 1,700' is the round figure. The point is the ratio: the third approach exposes roughly 4 percent of one month's data (6 out of 140) and answers the actual question. If it does not, widen the window, one step at a time.
What the AI categorisation stage sees
Separately from any assistant you connect, categorisation inside the app works in stages: your rules, then your history and similar merchants, and only if AI is on, an AI model for what is left, using the normalised merchant name and an amount band, not your name, account numbers, full statement text or balances. You can switch AI off for a feature, set a monthly AI budget or turn it off entirely, in which case unmatched items are held for review.
When something looks wrong, what stays offline?
- Your bank login. Never give it to an assistant or to anyone investigating a charge for you.
- One-time codes and card details.
- Screenshots of your full statement.
To dispute a charge, use your bank's official channel and quote the date, merchant and amount. SpendixAI does not read SMS and does not ask for net-banking logins, so nothing in this process requires you to expose those.
Your exits
If you want out of the whole thing, deleted transactions go to a recycle bin for 30 days, owners can export their data and choose document retention, and a whole workspace can be scheduled for deletion with a seven-day window to change your mind. For anything the assistant keeps on its side, look at its own settings.
Checklist before you start
- Run the built-in checks.
- Choose a one-month window.
- Use read-only.
- Ask for patterns, not a review.
- Disconnect and note the date.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


