How Does AI Categorize Bank Transactions? → ML vs Rules

Short answer

AI categorizes a transaction by predicting its category from signals such as the cleaned-up merchant name, the merchant category code, the amount, how often it repeats and how you labelled similar ones before. Rules-based systems match codes and keywords instead. The most dependable setups combine both and let your own rule win.

Every finance app has to turn a line like “SQ *BLUE BOTTLE 0412” into “Coffee”. There are three ways to do it: hand-written rules, a machine-learning model, or a large language model reading the text. Each has a different failure mode, and knowing which one your app uses explains most of the miscategorized transactions you see.

The signals a categorizer reads

Merchant text, cleaned up: “SQ *BLUE BOTTLE 0412” becomes “Blue Bottle”. Usually the strongest single signal.
Merchant category code (MCC), a four-digit code card networks attach to the merchant. Some bank exports include it.
Amount and direction: a $4 debit and a $1,400 debit at the same store are often different things; money in is usually income or a refund.
Recurrence: the same merchant, a similar amount, a regular interval points to a bill or subscription.
Your history: how you labelled similar transactions before.

Rules, machine learning or a language model?

Method	How it decides	Strength	Weak spot
Rules	Matches codes and keywords to categories; your own rules on top	Same result every time; easy to see why	A new merchant needs a keyword or a rule
Machine-learning model	Predicts from patterns learned on many labelled transactions	Copes with messy merchant text	Harder to predict; a correction may or may not stick, depending on the app
Language model	Reads the description and picks from your category list	Handles unusual descriptions	Sends transaction text to an AI provider; suggestions need review

Many apps mix them: rules for the clear cases, a model for the rest, and your own rules always on top. The question to ask an app is simple: if you correct the same merchant once, does the next import come in right?

How Finman categorizes transactions

Finman’s categorization, as described on the features page. Rules are applied in the order you set; the first match wins.

Finman sorts with rules. Each imported transaction gets a category from the merchant category code (monobank statements carry one) or from keywords in the description, in English, Ukrainian and Russian. A rule you set sorts the expenses that guess leaves in Other when you click Apply rules. Because the result is the same every time, your budgets, subscription detection and the AI CFO’s answers rest on numbers that do not shift between imports.

The AI comes in after sorting: the AI CFO reads your categorized spending to answer “where did I overspend?”, and with the AI on, an OpenAI GPT model reads receipt photos and suggests a category for you to check. See how other apps describe their sorting in our comparison of each AI finance app.

Get clean categories in one sitting

Import a month, fix the merchants that landed wrong with a rule each, and next month comes in sorted. Free during the open beta.

Try Finman Free

Where categorization still goes wrong

Payment processors such as PayPal or Square can hide the real merchant behind their own name. A rule on the text after the processor prefix fixes most of these.
Stores that sell everything: a superstore receipt can be groceries, clothes and electronics. Pick the category you budget against, or photograph the receipt to see the line items.
Transfers between your own accounts look like spending in one file and income in the other. Record them as a transfer between accounts so they stay out of your budgets.
Refunds arrive as money in. Keep them in the original category so the month nets out.

A 15-minute setup that pays off every month

Import one full month of transactions.
Look through the uncategorized and obviously wrong ones.
Add a rule for each of your ten most frequent merchants.
Next month, import again and check only what is new.

Clean categories make everything downstream more useful: subscription tracking, spending anomaly detection and budget alerts all read from them.

Frequently asked questions

How does AI categorize bank transactions?

It predicts a category from signals such as the cleaned-up merchant name, the merchant category code, the amount, how often the charge repeats and how you labelled similar transactions before. Rules-based systems match codes and keywords instead, and many apps combine both with your own rules on top.

What is transaction categorization with machine learning?

A model trained on many labelled transactions learns which merchant text, codes and amounts usually belong to which category, then predicts a category for new ones. It copes well with messy merchant names, while its decisions are harder to predict than a rule's.

Why does my app keep miscategorizing the same merchant?

Usually because no rule covers that merchant, or because the app's model does not learn from your correction. Add a rule for the merchant if the app supports it; in Finman one Apply rules click sorts every matching expense left in Other.

How accurate is automatic transaction categorization?

It is strongest on clear merchant names and codes and weakest on payment processors that hide the merchant, stores that sell many kinds of goods, and transfers between your own accounts. Check the first month closely, add rules for the misses, and later months need little cleanup.

Does Finman use AI to categorize transactions?

Finman categorizes with rules: the merchant category code or keywords in the description, in English, Ukrainian and Russian, and your own rules sort what the guess leaves in Other with one Apply rules click. The AI CFO then works with those categories, and with the AI on, a GPT model suggests a category when it reads a receipt photo.

Can I create my own categorization rules in Finman?

Yes. Create custom categories and rules; one Apply rules click sorts every matching expense left in Other, so a merchant you set a rule for once is handled the same way each month.

Sort a month in Finman

Import a statement, add rules for the merchants you know, and ask the AI CFO what the month looked like.

Get Started Free