Nothing an AI tool produces should leave the finance team until finance has verified the numbers. The trouble starts when the same report needs the same full check every time it runs, because then the tool is not learning, the data underneath it is not fixed, and your team has taken on a second job.
It matters because that second job is already big. Finance teams spend more than a quarter of their working week verifying or correcting AI output, according to a survey Datarails commissioned of 270 CFOs and finance leaders at large U.S. companies (Datarails sells AI finance software). That is time the tools were bought to give back. Most AI tools do not take work off the table; they add different work to the same window of time, and in finance that work lands on the people already closing the books.
Trust the direction, verify the numbers
My rule is trust but verify, and trust but learn. In finance that means two different levels of trust. You can trust that a well-built AI report is directionally right: the trend is the trend, the outlier is worth a look. You cannot send the specific numbers to a board, a lender or a customer until someone in finance has verified them.
That verification is how you learn what to trust next. Each time a report checks out, you should need to check a little less of it the following month. If the hours are flat month after month, the checking has turned into a symptom of the data underneath.
Repeated checking points to the data underneath
If the AI reads from systems that disagree, every output inherits the disagreement, and someone has to reconcile it by hand before it goes anywhere. Review cannot fix that, because the next run brings the same disagreement back. The survey bears it out: nearly two-thirds of the leaders said their most common frustration with AI is a confident answer built on the wrong data, and only 4% of their organizations have a single source of truth for finance and operational data.
Decide what gets checked before you buy more seats
Pick one workflow, the monthly flash report for example, and count the hours that go into checking its AI output. That number is part of what the tool costs, and it belongs next to the license line in the budget.
Fix the data under that one workflow before widening use. One reconciled source does more than three tools reading three versions of the truth.
Write down which outputs need a full sign-off and which can be spot-checked once they have held up, and name one person who owns those rules, the same way someone has to own the AI after a vendor leaves. Just 5% of the finance leaders surveyed trust AI to produce board-ready reports without review, and for board material that caution should stay.
Then hold off on adding licenses until the checking time comes down. More than half of the CFOs surveyed plan to expand AI licenses over the next year. More seats on top of unreconciled data buy more checking.
Small single-system teams and regulated reports are different
If your finance function is small, runs on one system and has one person who knows every number, the data problem may not exist, and the extra checking may just be a team learning a new tool. In regulated reporting, some human review is permanent by design. There the goal is to make the review faster.
Ask for one number this week
Ask your controller for the hours spent last month checking AI output, broken out by workflow, and the same figure for the month before. Whichever workflow is not coming down is where to start, and the data under it is usually the first thing to fix.
This is the kind of operating problem I work through with mid-market leaders, and there is more on how I work with companies. If the checking is eating your team's week, I am glad to talk it through.
John Wyatt · John Wyatt's page · All posts · RSS