There is real time to be saved around a cashroom, and it is easier to go after it once you have drawn the line. Three things belong on the never list. Naming them plainly is what makes everything else straightforward to approve.
A lot of cashroom work is retrieval and drafting. Finding the invoice, matching the receipt, preparing the posting, writing the chaser, working out why two numbers disagree. Software is good at that sort of work, and there is a lot of it.
What makes it safe to adopt is knowing where the line sits. A short list of things the software is never allowed to do is worth more than a long policy nobody can act on. It gives everyone the same answer without a fresh argument each time.
Money leaving a client account is an act of judgement, and somebody has to be answerable for it. Accountability attaches to a name, and a name means a person who looked at it and decided.
There is a practical version of the same point. If the tool can move money, then anyone who can influence the tool can move money — and these tools are influenced by whatever text is put in front of them. A supplier email. An attached invoice. A scanned letter with new bank details. Keeping authorisation with a person means the worst a bad input can do is waste somebody’s afternoon.
Software can also be wrong quietly, in a way that reads exactly like being right. A person checking will not catch everything either, but a person is who your rules, your insurer and your regulator expect to be standing there.
So: prepare the payment, populate it, check it against the file, flag what does not match. Press the button, never.
The test is simple. If the tool were switched off tomorrow and the supplier vanished, would your audit trail still stand up on its own?
If the answer is no, that is a problem you will discover at the worst possible moment — an inspection, a complaint, or a straightforward question about something from two years ago. A conversation held inside somebody else’s product is not a record you control. Neither is a summary the tool produced that nobody saved.
The fix is unglamorous and it works. Whatever the tool drafts, the decision, the approval and the reason land in your own system, in your own format, readable by someone who was not there, kept for as long as your own rules require. The tool is a drafter. Your records are the record. There is a related piece here on building a central record that has to stand on its own.
Client money or the firm’s money. That call sits underneath everything else in a cashroom, and it is rarely mechanical. The answer lives in the engagement terms, in what was agreed about fees, in what the payment was actually for, in what was said to the person who sent it, and in your own rules. Two payments that look identical on a bank statement can belong in different places for reasons that are nowhere on the statement.
Software will give you an answer anyway. It will give you one even where the honest answer is “it depends, and someone needs to read the file”. That readiness to answer is the risk.
So let it gather. Let it say: here is the payment, here is the engagement letter, here is the last correspondence about fees, here is where similar ones went before. Then a person decides. The gathering is most of the work. The decision is the part that has to be owned.
Everything else. The list is long.
Notice what those have in common. The tool does the retrieval and the drafting; a person does the deciding and the sending. That split is not a compromise you accept to keep the compliance lead happy. It is where nearly all of the time saving lives anyway, because the minutes go on finding things, not on the moment of judgement at the end.
A clear list of what the software may never do is exactly what lets you say yes to everything else.
There is a version of this that runs the other way — use it narrowly, see how it performs, loosen the reins later. That does not hold, because nothing in the reasoning changes with time. A person authorises the payment because somebody has to be answerable, and software cannot be.
Worth being straight about the limits too. These tools work from what they are given, so an incomplete file produces a confidently incomplete answer. They are fluent when they are wrong, which is harder to spot. They do not know what nobody wrote down. None of that is a reason to keep them out. It is a reason to put them where being wrong gets caught — in front of a person who was going to look anyway.
The nevers fit on one page. Once they are written down, the conversation stops being about whether to use AI at all and becomes a much shorter one about which job to hand over first. That is a better question, and a quicker one. There is more on where the line sits in the piece on assisting the person who signs.
Working out exactly which job to hand over first, for your own cashroom, with a number against it, is what the audit does.
Nothing here is legal, regulatory or accounting advice — your own rules and your compliance lead are the authority on what your firm must do, and if you want help working out where the line sits in practice, that is what doing it safely covers.
Twenty minutes on the phone, free, no obligation. Tell us the cashroom job you had in mind and we will say plainly whether it is one to hand over, and what would have to stay with a person.