To check what an AI assistant did with your expenses, compare totals before and after, look at the list of rows that changed, and undo anything wrong. The most reliable habit is to save a CSV download before you let an assistant edit anything, so you always have something to compare against.
What should you do before an AI edits anything?
- Download your transactions as a CSV and note the date and time in the file name.
- Note the total spent for the month and the total for each category you expect to be touched.
- Start with a narrow request, such as one month or one category, not the whole history.
- Keep the AI connection read-only unless the job really needs changes.
Ten minutes of preparation turns a vague worry into a simple comparison.
How do you find out what changed?
First, see whether your tracker keeps a change or activity history and use it if it does. Not every tool shows who changed what, so do not assume it. If there is no history, use totals. The overall total for a month should not move when an assistant only recategorises, because recategorising shifts money between categories without creating or removing any. If the total moved, rows were added, edited in amount or deleted, and that needs immediate attention.
Worked example: an assistant recategorises one month
Say you ask an assistant to recategorise October. Before the change, your CSV shows a total of ₹54,000 across these categories.
| Category | Before | After | Change |
|---|---|---|---|
| Food | ₹14,000 | ₹12,500 | -₹1,500 |
| Transport | ₹6,000 | ₹7,200 | +₹1,200 |
| Shopping | ₹9,000 | ₹8,600 | -₹400 |
| Bills | ₹15,000 | ₹15,000 | ₹0 |
| Other | ₹10,000 | ₹10,700 | +₹700 |
| Total | ₹54,000 | ₹54,000 | ₹0 |
The total is unchanged, which is what you want. The changes net to zero: -1,500 + 1,200 - 400 + 700 = 0. Now look at the biggest movements. Food lost ₹1,500, so ask which food rows moved. Perhaps three taxi rides labelled as a food delivery brand were correctly moved to Transport, and that is a fix. If instead a ₹1,500 restaurant bill moved to Other, undo it.
How do you find the exact rows?
- Filter the tracker by the month and by each category that changed, and scan the rows.
- Sort by amount and look at the largest rows first, since big errors cost the most.
- Compare two CSV downloads in a spreadsheet: put the before data in one sheet, the after data in another, and use a lookup on a unique key such as date plus amount plus description to spot rows whose category differs.
- Check any row you did not expect to be touched, for example a transfer or a refund.
How do you undo a wrong change?
Correct the category by hand on the affected rows. In SpendixAI, corrections become history that the categoriser learns from, and merchant aliases can be added, so a repeated mistake can be fixed at the source. Deleted transactions go to a recycle bin for 30 days, so an accidental delete can be recovered in that window. For anything bigger, re-import from the original statement and compare again, remembering that possible duplicate imports are flagged and should be reviewed.
If you use an assistant often, make the review a calendar item, not a mood. After every session that allowed changes, spend five minutes on totals and ten on a sample of rows. Keep each before-and-after CSV pair in one folder named by month. Within a quarter you will have a clear record of what the assistant tends to get wrong, such as mixing up reimbursements and shopping, and you can write that down as a standing instruction for next time.
A short routine to keep
- Save a CSV before any bulk change.
- Check totals first, categories second, individual rows third.
- Ask the assistant to list what it changed, then verify that list yourself.
- Review a sample of ten rows, not just the summary.
- Return the connection to read-only once the job is done.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

