Do not categorise hundreds of UPI rows one by one. Count which merchant names repeat, write a rule for the top ten or fifteen, and let those rules sort the bulk of the statement; the leftovers are the only rows you read. In a typical metro salary account the same few merchants cover most of the lines.
Why do UPI rows look so cryptic?
A UPI line on a bank statement is often a long string: the word UPI, a reference number, the payee's handle and a short note. A grocery order, a cab ride and a friend's dinner split can all look like strings of digits. The part that tells you who was paid is usually a fragment in the middle, and the rest changes with every payment.
That is why exact-text matching fails. The useful trick is to match on the stable fragment, not the whole string.
How do you find the rows that matter?
Import one month, then sort the uncategorised rows by merchant text. Count how often each stable fragment appears. Say a salaried employee in Bengaluru has 240 UPI rows in a month:
| Stable fragment | Rows | Category to assign |
|---|---|---|
| Cab app name | 34 | Transport |
| Quick-commerce app name | 28 | Groceries |
| Office canteen | 22 | Food at work |
| Food delivery app | 19 | Eating out |
| Electricity board | 1 | Utilities |
| Neighbourhood chemist | 6 | Medical |
Those six fragments cover 110 of 240 rows, a little under half. Ten more rules for the tea stall, the metro card top-up, the gym, the barber and the rest would take you well past 200.
How does the order of categorisation work?
In SpendixAI, categorisation runs in stages. Your own rules go first, then merchant defaults and your own history, and only if AI is switched on does a model handle what is still unmatched. A rule always beats a guess, so write one whenever a merchant repeats. Corrections you make become history, so a similar row finds its category the next time.
Choose the account before importing, so the rows keep the right bank or card. The import is reviewed before saving, which is when to create your rules from rows on the screen.
What does a rule look like in practice?
- Open a row from the cab app, copy the stable fragment of its text.
- Create a rule: if the text contains that fragment, assign Transport.
- Run it over the month, then check three of the rows it touched.
- Move to the next most frequent fragment.
- Stop at the point where a new rule would catch fewer than three rows.
What about person-to-person payments?
Payments to individuals, such as a plumber, a tutor or a friend, have no merchant to match. Do not write a rule on a person's name unless the payment is a regular one, like a domestic help's monthly pay. For one-offs, edit the category by hand and add a note, because six months later you will not remember which Rajesh it was.
What should you do with the last few rows?
Leave the rare ones alone. A one-off payment to a repair person does not deserve a rule; it deserves a correct category and a note. Chasing the final handful of rows costs more time than they are worth, and every rule you add is one more thing to check later. Fix the rows, write the note, and move on to the next month.
What should you do next month?
Import the next statement and note how many rows are uncategorised. If it started at 240 and the figure is now 40, the rules worked. Spend ten minutes on the remaining 40, writing rules for anything that appears twice. By the third month the review should be a coffee-length job. If a merchant appears under two spellings, add an alias so both map to the same name.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

