AI and MCP

Why you should not give AI delete access to your expenses at first

Why deleting is the riskiest thing to hand an AI assistant in an expense tracker, how it goes wrong, and safer routes: flag, review and a recycle bin.

Editorial illustration for Why you should not give AI delete access to your expenses at first

Do not give an AI assistant delete access to your expenses at the start, because a deletion removes the record you would use to check the mistake. A wrong category or a wrong amount can be spotted and corrected; a wrong delete leaves a gap that is easy to miss. Begin with read access, flag things for review, and keep deletion as your own job.

What actually goes wrong with bulk deletes?

The failures are rarely dramatic. They come from an instruction that sounded clear to you and was interpreted more widely by the assistant.

  • The wrong filter. You say 'delete the test entries'. The assistant matches every transaction with a small amount or a repeated merchant name, and some of those are real.
  • The duplicate clean-up that goes too far. Two payments of 1,200 rupees to the same merchant on the same day may be a double import, or may be two genuine purchases. A rule like 'same amount, same merchant' cannot tell them apart.
  • The date range slip. 'Remove last month's cash entries' is read as the calendar month, or as the last thirty days, or as the financial year so far.
  • The cascade. Deleting a merchant's transactions also removes the receipts and notes attached to them, which you may need later.

A worked example of one bad afternoon

Say your tracker holds 420 transactions for October and you ask the assistant to remove duplicates. It finds 38 pairs matching on amount and merchant and deletes one from each pair. On review, 9 of those pairs were real, for example two chai orders of 60 rupees each at the same stall.

You now have 38 fewer rows, of which 29 were true duplicates and 9 were genuine spends. The month looks 9 purchases lighter than it was. If the 9 averaged 450 rupees, your October total is understated by 9 x 450 = 4,050 rupees, and you have no list of which rows vanished.

What is a safer way to clean up duplicates?

  1. Ask the assistant to list suspected duplicates, not remove them.
  2. Look at the list yourself, with dates and accounts side by side. Two rows on the same day in different accounts are often a transfer, not a duplicate.
  3. Delete the confirmed ones by hand, or in small batches you have checked.
  4. Recount the month and compare it with your statement total.

Some trackers also flag likely duplicates themselves. SpendixAI flags possible duplicate imports and has a report that lists the same amount paid to the same merchant within two days, which gives you a list to judge instead of a purge to trust.

There is also a quieter cost. After a mistaken clean-up, you stop trusting your own numbers, and a tracker you do not trust is a tracker you stop opening. Preventing that is worth more than the few minutes a manual review takes.

What protects you if a delete does happen?

  • A recycle bin. Where deleted rows are held for a period (SpendixAI keeps them for 30 days), a mistake is a restore, not a rebuild.
  • A fresh export before any bulk change. A CSV saved on your device is a plain backup you can read in a sheet.
  • A note of the monthly total before and after. If the totals differ, something moved.
  • The original statement. For imported rows, the bank statement is the truth you can re-import from.

How does this compare with a human mistake?

When you delete the wrong row yourself, you usually know it the moment you click: you saw the row. An assistant works through a filter, so you see a count, not the rows. That gap is why the same error is harder to catch. Ask for the count first, and for a sample of five rows it would remove, before agreeing to anything.

When might delete access make sense later?

If you have used read, add and edit access for a couple of months, understood what the assistant does with vague instructions, and have a working recycle bin and recent exports, you can consider allowing deletes for narrow, well-defined jobs. Even then, ask for a count first: 'How many rows would that remove?' If the number surprises you, stop and look.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports