How to Categorise Transactions with AI

Finally, a job it is actually built for

Here is the good news you have been waiting two lessons for. Categorising bank descriptions is a language problem wearing a finance costume, and language problems are what these tools are for.

A bank export is a wall of shouty abbreviations. Somebody has to read "SQ *THE COFFEE HO LONDON GB CARD 4471" and decide whether that is staff welfare, client entertaining or the director buying a flat white on the way to the office. That decision needs pattern recognition across thousands of half-remembered merchant names, which is precisely the thing a language model does well and precisely the thing that turns your Tuesday into a slow grey smear.

So use it. Carefully, with four rules, and never on autopilot.

Rule one: give it your actual chart of accounts

Left to itself, the tool will invent categories. It will produce Office Supplies, Miscellaneous Expenses and Business Costs, none of which exist in your client's ledger, and you will spend longer mapping its inventions onto reality than you would have spent doing the job by hand.

Paste your real category list. The exact names, and the codes if you use them. Then say plainly that nothing outside the list is allowed.

You are helping me categorise bank transaction descriptions.
These are the ONLY categories you may use. Use the exact names as written:
[paste your chart of accounts, one per line, with codes if you use them]
Rules:

Never invent a category. If nothing fits, use UNSURE.
If a description is ambiguous between two categories, use UNSURE
and name both candidates. Do not pick the more likely one.
Do not change, reformat or reorder the descriptions.
Return a table with four columns: original description,
category, UNSURE yes or no, and a short reason.
Here are the transactions:
[paste them]

Rule two: demand UNSURE instead of a guess

This is the change that makes the whole technique safe, and it is one word.

By default these tools guess. They are built to produce an answer, and an answer is what you will get, delivered in exactly the same confident tone whether the merchant is obvious or completely opaque. A guess dressed as a classification is worse than no classification at all, because a blank cell gets reviewed and a filled cell does not.

So make refusal an option and make it explicit. "If you are not confident, write UNSURE" turns the tool from something that hides its uncertainty into something that hands you a review queue.

Rule three: review every UNSURE, spot check the confident ones

The UNSURE rows are not the risk. You are already looking at those. The risk lives in the rows it was sure about, because those are the ones you are tempted to wave through.

So take a sample of the confident rows and actually check them. Ten rows out of a hundred, chosen deliberately rather than randomly: the largest amounts, anything that might be capital rather than revenue, anything with VAT implications, any category that suddenly has far more entries than usual, and anything from a merchant name you do not recognise.

If your sample is clean, the batch is probably fine. If two of your ten are wrong, stop, fix the prompt or the category list, and run it again. Do not patch individual rows and carry on, because whatever caused those two errors is still sitting in the other ninety.

From the table you just produced, list for me:

The 10 largest amounts and the category you gave each.
Every row where the amount is over [1000], whatever the category.
Any category that has more than [15] rows in it.
Any merchant or description that appears with two different categories.
Do not re-categorise anything. I am reviewing your work, not asking you to redo it.

Item four is the quiet winner. The same merchant landing in two different categories in the same batch is either a genuine judgment call or an inconsistency, and either way you want to see it.

Rule four: it never posts

No output from a chat tool goes straight into a ledger. Not through an import, not through a copy and paste into a bulk update screen, not through an automation somebody set up cleverly six months ago.

Every categorisation is a proposal. You approve it. That approval is the control, and it is the only thing standing between a plausible-looking classification and a set of accounts that quietly stopped being right in March.

One more prompt: teach it your quirks

Every client has habits. The director who uses the business card for the school run. The supplier whose name changed but whose bank reference did not. Write those rules down once and paste them in every time.

Before categorising, apply these client-specific rules,
which override your normal reasoning:

Anything from [merchant] is always [category].
Anything containing [reference text] is always [category].
Any purchase over [amount] is UNSURE, even if the category looks obvious,
because I need to consider whether it should be capitalised.
Anything on a weekend from [card] is UNSURE.
[add your own]
Now categorise the transactions below under those rules
plus the chart of accounts I gave you.

That third rule is a small thing that saves a large embarrassment. A laptop bought on a card looks exactly like stationery bought on a card, right up until the audit.