What AI expense categorisation gets right and wrong

Automatic expense categorisation is very good at repetition and structurally bad at intent. It learns which account a vendor usually posts to, and it cannot know why you bought something. Almost every categorisation error traces back to that one boundary.

Which makes the useful question not “how accurate is it” but “which of my expenses are decided by the vendor and which are decided by the purpose”. The first group automates cleanly. The second never fully will.

What it gets right

Repeat vendors. Your hosting provider posts to the same account every month. After the second or third occurrence, this is a solved problem and it stays solved.

Unambiguous vendors. Some businesses only sell one category of thing. A domain registrar is not a plausible meals expense. Where the vendor determines the account, categorisation is reliable immediately, even on first sight.

Consistency. This is the underrated win. A human categorising a hundred receipts across three months will not make identical decisions each time — they’ll post the same vendor to two different accounts depending on mood and memory. Software is mechanically consistent, and consistency is what makes reports comparable month to month. Being consistently slightly wrong is easier to fix in bulk than being randomly right.

Volume. The marginal cost of categorising the thousandth receipt is the same as the first. For anyone whose expense volume grew past what they’ll actually sit and sort, this is the difference between having categorised books and not.

What it gets wrong

Intent. The same purchase from the same shop is a different expense depending on why you made it. A laptop is equipment; the same laptop bought for resale is inventory. Lunch alone is not the same as lunch with a client. Nothing on the receipt distinguishes these, so nothing reading the receipt can.

General retailers. A big-box store or a large online marketplace sells across a dozen categories. Vendor tells you almost nothing, and only line items — the least reliably extracted part of a receipt — carry the answer.

Splits. One receipt, two accounts: the weekly shop that includes both office coffee and your own groceries. Categorisation assigns a receipt to a category; it does not naturally express “60% this, 40% that”, and where it can, it has no basis for choosing the ratio.

Your specific chart of accounts. Categories are conventions, and yours are not universal. If your books distinguish “Software — infrastructure” from “Software — tools” for reasons that live in your head, the distinction is not inferable from any receipt.

New vendors. With no history, a first-time vendor gets a guess from vendor-name patterns. Sometimes right, sometimes confidently wrong, and either way unremarkable-looking in a list of a hundred entries.

Rules beat learning where rules exist

The important design point: an explicit rule is a decision you already made, and it should override anything learned.

If you know every charge from a given vendor posts to Software, write that down as a rule. It then applies deterministically, forever, without drifting, without needing examples, and without a confidence score. You can read it, audit it, and change it in one place.

Learned categorisation is for the long tail — the vendors that are too varied or too infrequent to be worth a rule. Use it there and it’s genuinely useful. Rely on it where a rule would do and you’ve replaced something auditable with something probabilistic for no gain.

A practical split that works: rules for your top twenty vendors by transaction count, which is usually most of your volume, and learned categorisation for everything else with a review pass on anything new.

Reviewing by exception

You will not review every categorised expense, so decide deliberately what you do review:

First occurrence of any vendor. This is where a wrong pattern gets established, and where correcting it once prevents it recurring.
Anything above an amount that matters to you. Small miscategorisations are noise. Large ones distort reports and, at year end, tax positions.
The general-retailer accounts. Whatever bucket your marketplace and big-box purchases land in, read it monthly. This is where intent-dependent expenses accumulate.
Anything with a category the software chose but flagged. Low confidence on a category usually means the vendor genuinely is ambiguous — which is exactly when a human decision is worth recording as a rule.

The realistic outcome

Well set up, automatic categorisation handles the large majority of your expense lines with no attention, gets your recurring costs exactly right, and hands you a short list of genuinely ambiguous items to decide.

Badly set up — no rules, no review, no vendor normalisation — it produces books that look complete and are quietly wrong in the accounts you care most about. The technology is the same in both cases. The difference is entirely in whether the decisions that only you can make were written down anywhere.