AI vs Rules-Based Transaction Categorisation: A Treasury Guide

Categorisation is one of the oldest jobs in treasury. Tag every transaction so the data downstream actually means something: forecasts, Variance Analysis, Cash Positioning, board reporting. For years, treasurers have done it one of two ways: by hand, or with rules. AI is replacing both.

The interesting argument isn't AI vs manual. Most teams gave up on tagging thousands of transactions a week by hand long ago. The real argument is AI vs rules, because rules are what most TMSs and Excel-based processes still run on. That's the layer AI is taking out.

Why rules-based transaction categorisation breaks down

Rules look elegant on paper. If the counterparty contains "HSBC" and the amount is negative, tag it as a bank charge. If the description starts with "PAYROLL", tag it as payroll. Easy.

Reality is messier. Counterparty names change when a vendor rebrands. Payment descriptions get truncated by the bank. A new entity is acquired and starts paying into the same account with a different reference format. A subsidiary in another country sends a tax payment no rule has ever seen. Each of those breaks the rule silently. The transaction either falls into "Uncategorised," or gets miscategorised and quietly distorts the forecast.

Then there is the maintenance burden. Every new counterparty needs a rule. Every edge case needs an exception. The person who built the engine left two years ago, and nobody is sure which rules are still in use. The library grows, contradictions multiply, and the rules never quite catch up with how the business actually moves money.

Rules work on the world you've already seen. They collapse on the one you haven't.

How AI transaction categorisation works

AI categorisation is pattern recognition with feedback.

The model trains on your historical transactions and the categories a human assigned to each one. It learns the signals that distinguish a customer payment from an intercompany transfer, a one-off bonus from a payroll run, a tax settlement from a supplier payment. Not just the counterparty name, but the amount, the description, the account pair, the currency, the timing, all weighed together.

When a new transaction lands, the model doesn't ask "does this match any rule?" It asks "which pattern does this look most like?" and outputs a category with a confidence score. Above the threshold, it tags automatically. Below it, the transaction goes to a treasurer for review. Their decision feeds back into the model.

That feedback loop is the key difference. Rules don't learn. AI does. Every correction makes the next prediction sharper. New counterparties get classified the moment they appear. Description changes don't break anything. The model handles ambiguity because it doesn't need an exact match.

What AI categorisation can do that rules can't

Rules need explicit logic. AI captures patterns you'd never think to encode: that a particular customer always pays on the 27th, that a supplier's invoices come in two amount bands, that a certain intercompany flow only happens at quarter-end. Those signals improve accuracy in ways no rule library could.

And rules don't scale across entities. AI does. One model categorises transactions across every subsidiary, in every currency, with consistent logic.

How treasurers stay in control with AI categorisation

The fear with AI is the black box. A good system isn't one. Every prediction comes with a confidence score and a visible reason. The treasurer can override any tag, and the model learns from that override. Transactions can be marked as System Outliers so they don't pollute future training. The taxonomy can be set at any level of hierarchy: entity, account, subsidiary.

You keep the control rules gave you. You lose the maintenance burden.

AI vs rules-based categorisation: the bottom line

Rules-based categorisation was a step forward thirty years ago. It is now the bottleneck: the library nobody owns, the edge cases nobody covered, the silent miscategorisations nobody catches until the forecast misses.

AI categorisation handles the messy reality of real treasury data, learns from every correction, and stays accurate as the business changes. It isn't a faster rules engine. It's what rules engines were trying to become.