Why you need AI transaction categorisation: for cash flow forecasts you can trust
by Rodel van RooijenAsk a treasurer what makes a cash forecast trustworthy and they will talk about data quality. Ask an engineer, and the answer is more specific: categorisation. Every transaction that flows into Palm has to be assigned to the right category, operational payments, payroll, vendor payments, interest income, and so on. Those categories are the building blocks of the 13-week forecast and every report layered on top of it. Get them right and the whole system sharpens. Get them wrong and the error carries straight through to the numbers people make decisions on.
This is why we treat categorisation as core infrastructure rather than a nice-to-have. A forecast is only ever as good as the classification beneath it. Clean categories give you accurate forecasts, reporting that reflects reality, and a clear view of where cash is actually moving. Messy ones give you variance nobody can explain at quarter-end.
Why you should switch to AI categorisation
For most treasury teams, categorisation has always been manual. People read transaction descriptions, apply judgement, and assign a category by hand, thousands of times a month. It is slow, it is inconsistent between people, and it does not scale with transaction volume. Worse, it pulls skilled treasury professionals into data entry when they should be managing liquidity and risk.
Rules-based automation was the first attempt at a fix, but rules are brittle. Bank descriptions are messy free text, full of reference numbers, counterparties, and noise, so a keyword rule that works for one bank breaks for the next. And rules do not just break on day one, they decay. A rule written for how the business looked last year gradually turns redundant as the business changes, as a company moves banks, or as it launches in a new region with new formats, new counterparties, and new patterns of activity. You end up maintaining hundreds of rules that made sense once and no longer do.
Yet bank statements are a treasure trove of meaningful data that can really shed light on what is happening with the business, live, every day. The signal is all there. The challenge is reading it before it gets lost.
This is exactly the kind of problem modern AI is built for. Language models read a transaction the way a person would, weighing the description, the amount, and the direction of the cash flow together, and infer the right category from context rather than a rigid rulebook. They generalise across banks and formats, they improve as they see more data, and they handle the long tail of unusual transactions that rules never anticipate.
What we found at Palm was that having machine learning and LLMs working together gives the best and most reliable results. LLMs provide a breadth of information based on the billions of data points they are trained on, while the machine learning algorithms we train in Palm are based on your data and corrections to make your future categorisation more deterministic.
Where categorisation most often needs a second look
No system, human or machine, gets every transaction right the first time, and in treasury some flows are notoriously easy to misread. These are the ones that surface again and again:
Intercompany sweeps and cash pooling. The classic offender. Zero-balancing sweeps, notional and physical pooling movements, header-account concentration, and in-house bank settlements all look like genuine inflows and outflows, but they are cash moving between your own entities. Misclassify them as operating activity and you inflate gross flows on both sides, double-count liquidity that never left the group, and hand yourself a pooling balance that overstates true available cash. For a group running daily sweeps, this is the single biggest source of phantom variance.
Treasury and FX settlements. Spot and forward settlements, swap legs, and cross-currency funding transfers are easily miscategorised, especially where a single economic event lands as two or more line items across accounts. Split the legs into the wrong buckets and both your FX exposure view and your currency-level cash position drift from reality.
Debt and financing flows. RCF drawdowns and repayments, term-loan amortisation, interest and coupon payments, and commercial paper roll-overs are routinely confused with operating cash. When a revolver drawdown reads as an operating inflow, your forecast quietly masks a funding need, exactly the signal a treasurer most needs to see.
Interest income versus investment principal. For money market fund subscriptions and redemptions, deposit placements and maturities, the split between principal and yield is subtle and high-value. A large redemption booked as income, or interest booked as investment, skews both the return picture and the maturity ladder.
Payroll, tax and statutory payments. Payroll, PAYE, VAT, and other period-end obligations are sensitive to direction and timing. A stray positive in a payroll line, or a run that lands a week early or late, produces a variance treasurers spot instantly and lose trust over.
Catch-all categories. Broad buckets like Revenue and Supplier Payments absorb any transaction with a blank or thin remittance reference. Convenient, but they are exactly where genuine misfits, chargebacks, reversals, and one-off items hide from view.
Worth saying: not all of these are errors. Many reflect how a particular treasury team chooses to structure its categories, so the goal is to give treasurers agency over their own definitions, not to impose a single notion of correctness.
Introducing recategorisation
This is where Palm's recategorisation comes in. Instead of asking users to hunt through tables for the transactions above, Palm runs categorisation health reviews and surfaces the ones worth a second look. Errors rarely happen in isolation, so rather than fixing transactions one at a time, we group similar transactions into clusters and let users review and correct an entire pattern in a single action.
Every suggestion comes with a recommended category and a confidence level, so the safe changes can be accepted fast and the borderline ones given a closer look. And every fix is ranked by financial impact, so users start with the changes that move their numbers most. The whole experience lives inside the Forecast Overview, next to the chart it affects, so applying a change and seeing the forecast respond happens in the same place.
The result
Together, AI categorisation and recategorisation change the economics of forecast accuracy. The AI does the heavy lifting automatically. Recategorisation gives users a fast, high-leverage way to refine the edges, turning what used to be days of manual correction into minutes. Early testing with our customers has been encouraging, with users calling the flow simple and effective and several saying they would open it almost daily.
Categorisation may not be the flashiest part of a treasury platform. But it is the part everything else depends on. What may seem like a small step is proving to be a major driver in how we forecast cash.
Frequently asked questions
See your cash future, clearly
Book a Demo
Our Treasury Kit
Go to Learning Hub
