How Much of Your Finance Team's Week is Moving Numbers?
By Gábor Szórád, Founder at AI Business Impact
Ask a finance leader what the team does all week and you will hear about judgment: closing the books, challenging the forecast, deciding what to pay and when. Watch the same team for a day and you will see something else. Someone opens the ledger, finds the overdue invoices, writes a polite reminder, adjusts the tone for the customer who always pays late, and sends it. Forty minutes each. Someone else copies figures from the bank portal into a spreadsheet, then from the spreadsheet into the planning tool, then checks that the three agree.
Over the past three years my team has worked with more than sixty small and mid-sized companies on where AI fits in their operations. The pattern in finance is similar in many of the teams we have worked with: a large share of the week is moving numbers between systems, and few have measured it.
Two numbers from one company
A client of ours, a professional services firm, received about five hundred supplier invoices a month. Each one was opened, read, coded, matched and keyed by a person. We built a reader that does the reading, the coding and the matching, and leaves the person to approve the exceptions. The arithmetic was not subtle: the task cost about 37,500 euros a year in salaried time, the build cost 7,250 euros, and it paid for itself in under three months.
In the same company a finance lead wrote every payment reminder by hand. It took her forty minutes per customer, and the reminders came to about 5,500 euros a year of her time. That one was handed over in a week.
Neither of these is a transformation. They are the kind of task that is embarrassing to explain to a new joiner, which is exactly why they are the right place to start.
Measure before you automate
The mistake I see most often is choosing the first AI project by size. Leaders look for the largest process, because the largest saving looks best in a board paper. The largest process is also the one with the most exceptions, the most people with opinions, and the longest road to a result that anyone can see.
A better first question is: which task do your people hate most? Then price it honestly. Count the people who touch it, the hours a week each of them spends, and what an hour of their time costs the company. Two people, ten hours a week each, at 35 euros an hour is roughly 36,000 euros a year. Write that number down before anyone talks about tools.
Where will the value land? There are only three honest answers: a cost you stop paying, revenue you now capture, or a hire you do not make. If the answer is "we are not sure yet", the project is a nice-to-have, and it should wait behind the ones where the answer is clear.
Keep the judgment where it belongs
Finance teams worry, reasonably, that automation means trusting a machine with money. It does not have to. The useful way to think about it is as a ladder. On the first rung the AI is an assistant: it prepares the work and a person does it. On the second it is a copilot: it does the work and a person checks every item. Only on the third, after months of clean results, does it run on its own with a person checking exceptions. Most finance tasks should stop on the second rung for a long time, and some should never leave it.
That is also where the people stay useful. The reader we built does not decide whether an invoice is wrong. It notices that this one differs from the last twelve and puts it in front of someone who knows the supplier. The judgment that was always the job is still the job. What has gone is the copying that surrounded it.
What to do this quarter
Pick one task in your own finance function that fits four conditions: someone knows it well and wants it gone, the result can be checked without a debate, it happens often enough to test on real examples, and a mistake can be recovered the usual way. Price it. Set a limit on the time and money you will spend before you review it. Then run it for a month and decide, honestly, whether to keep it, change it or stop.
You will have one number, one working automation and a team that has learned how to do the next one. That is worth more than a roadmap of twenty.
