AI and MCP

Worked example: fixing 20 wrongly categorised transactions with AI

A step-by-step cleanup of 20 wrongly categorised transactions: group them by merchant, set rules, re-check, and see before and after category totals.

Editorial illustration for Worked example: fixing 20 wrongly categorised transactions with AI

You fix 20 wrongly categorised transactions fastest by not fixing them one at a time. List the wrong rows, group them by merchant, decide one correct category per merchant, turn each decision into a rule, and then check that the category totals moved by exactly the amount you expected. An AI assistant helps with the grouping and the arithmetic; you make the decisions.

The starting point: a month that looks off

The numbers below are made up to show the method. Say a month of spending totals ₹41,000, and it looks strange: dining seems high, "Other" is a large pile, and fuel is hiding inside Shopping. You scroll the list and flag 20 rows that are clearly in the wrong place.

CategoryBefore
Groceries₹14,200
Dining₹6,800
Transport₹3,100
Shopping₹9,500
Other₹7,400
Health₹0
Total₹41,000

Step 1: list the wrong rows and group them by merchant

Ask the assistant to take your 20 flagged rows (date, merchant, amount, current category) and group them by merchant. Do not ask it to guess categories yet. You only want a tidy grouping, so the decisions stay yours. Here is what the grouping might look like.

Merchant groupRowsAmountCurrently inShould be in
Grocery delivery orders6₹2,640DiningGroceries
Local provision shop (UPI)4₹3,150OtherGroceries
Fuel pump payments5₹4,300ShoppingTransport
Metro card recharges3₹900OtherTransport
Pharmacy2₹1,350OtherHealth

Check the row count yourself: 6 + 4 + 5 + 3 + 2 = 20, and the amounts add up to ₹12,340. If the assistant returns 19 rows or a different total, stop and ask it which row is missing before going further.

Step 2: why rules beat one-off edits

If you only edit the 20 rows, next month the same merchants will land in the same wrong places and you will repeat the work. A rule says "this merchant always goes to this category", so the fix carries forward. In SpendixAI, categorisation runs in stages: your own rules first, then your history and similar merchants, then an AI model only for whatever is left and only if AI is switched on. Correcting a category also becomes history, and you can add a merchant alias when one shop shows up under several messy names.

  1. Write one line per merchant group in the form: merchant name contains X, category Y.
  2. Create or apply those five rules in your tracker.
  3. Correct the 20 existing rows if the tool does not re-apply the rules to old entries.
  4. Leave any merchant you are unsure about unassigned rather than guessing.

Step 3: how do you confirm the fix worked?

Re-run the category totals and compare them with what you expect. Each moved amount leaves one category and enters another, so every number can be predicted before you look.

CategoryBeforeChangeAfter
Groceries₹14,200+₹2,640 +₹3,150₹19,990
Dining₹6,800−₹2,640₹4,160
Transport₹3,100+₹4,300 +₹900₹8,300
Shopping₹9,500−₹4,300₹5,200
Other₹7,400−₹3,150 −₹900 −₹1,350₹2,000
Health₹0+₹1,350₹1,350
Total₹41,000net ₹0₹41,000

The total stays at ₹41,000 because recategorising never changes how much you spent, only where it is filed. If the grand total changes, a row was deleted, duplicated or edited in amount by mistake. That single check catches most errors.

What do you do with the leftover ₹2,000 in Other?

Open those rows next. Some will be genuinely miscellaneous, and that is fine. Others need a closer look at the original statement line. If a merchant name is cryptic, check the receipt or your UPI history before naming a category. A small honest "Other" is better than a tidy but wrong one.

When to let AI handle it and when not

AI works well for grouping rows, spotting that three merchant names are really one shop, and checking sums. It is weaker at knowing what a cash payment to an unnamed person was for. If you use SpendixAI, 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 your review. If you connect Claude or ChatGPT through MCP, you can start read-only, so the assistant can look but cannot change anything until you decide otherwise.

  • Flag rows by eye first, then ask for grouping.
  • Decide the category per merchant yourself.
  • Turn each decision into a rule.
  • Match the grand total before and after.
  • Review whatever is left in Other.
SPENDIXAI JOURNAL

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

Explore your reports