AI and MCP

What AI expense categorisation can and cannot do

A plain account of where AI expense categorisation is strong, where it guesses and why, with Indian examples like a UPI payment to a person.

Editorial illustration for What AI expense categorisation can and cannot do

AI expense categorisation is good at spotting patterns in merchant names and past choices, and weak where the statement does not hold the answer, such as a UPI payment to a named person or a merchant it has never seen. It will usually get regular shops right and will sometimes be confidently wrong elsewhere. Treat it as a fast first pass that you review, not as a final ledger.

What is it good at?

Recognisable merchants are the easy part. A debit that names a fuel station, an electricity board or a grocery chain points clearly at one category. Repeated payees help too: if you have put one payee under Groceries ten times, the next payment looks like the same. It also copes with ugly text, cutting a long narration down to the useful name, far faster than anyone can by hand.

Where does it go wrong?

SituationWhy a category is hard to know
UPI payment to a named personIt could be rent, a loan repaid, a share of dinner or a gift
A brand-new merchantThere is no history or pattern to learn from
A shop that sells many thingsThe same merchant could be groceries, gifts or household items
A cash withdrawalThe statement does not say what the cash bought
A card payment in a foreign currencyThe merchant name may be abbreviated and unfamiliar
An unusual amount at a usual placeA bulk purchase or a different purpose than the usual one

Why can it not know what only you know?

Take a UPI transfer of ₹18,000 to a person called Amit Kumar. For one reader it is rent, for another a friend repaid, for a third a family wedding contribution. The narration is the same. The assistant sees only the text and the amount, so any answer is a guess based on what is common, not on what happened. This is not a defect to be fixed; it is a limit of the information.

Is a wrong category a big deal?

It depends on what you use the category for. In a personal tracker, one transfer filed under the wrong heading skews a pie chart slightly. In business books, where categories can carry GST treatment or tax codes, a wrong category can carry through to estimates. That is why review matters more as the stakes rise, and why figures like GST estimates should always be confirmed with a CA.

How does SpendixAI approach this?

SpendixAI categorises in stages. First your own rules, then your history and similar merchants, and only when AI is switched on does an AI model look at what is left. That model works from the normalised merchant name and an amount band, not your name, account numbers, full statement text or balances. You can turn AI off for a feature, set a monthly AI budget, or turn it off entirely, in which case unmatched items are held for review. Every correction you make becomes history.

A useful way to think about it: the assistant is strong where a sensible clerk could look at the text and decide, and weak where the clerk would need to phone you. If a human with no context could not tell, the AI cannot either.

How should you use it well?

  1. Let the automatic pass run, then review the uncategorised and the large items first.
  2. Fix repeat offenders with a rule or alias instead of correcting them one at a time.
  3. Spot-check a sample every few months to see how well it is doing for your own spending.
  4. Keep a short list of payees that are people, and decide once what each one usually is.

The aim is to reduce your effort, not to remove your judgement.

SPENDIXAI JOURNAL

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

Explore your reports