Most conversations about AI in a month-end close start with a list of tasks and sort them into "automate" and "don't". That sorting usually goes by difficulty: the repetitive work goes to the machine, the hard work stays with people. It is a reasonable instinct and the wrong line.
The better line runs between two kinds of step. Some steps produce evidence: they gather, match, compare and flag. Others produce a judgement that someone will later have to stand behind. A close needs both, and AI is useful for the first kind in a way it is not for the second.
What belongs on the evidence side?
Pulling balances from the systems that hold them. Matching transactions between a ledger and a statement. Comparing this month's figures with last month's and with the budget, and listing the movements a commentary will need to explain. Checking that every journal has its supporting document attached.
All of these share a property: their output can be checked against something outside the model. A match either ties to the statement or it does not. A flagged movement either exists in the data or it does not. That is what makes them fair candidates to hand over, provided the check against the outside source actually runs. A match that ties can still be misclassified, so the check has to test more than whether it balances.
What belongs on the judgement side?
Deciding what an unmatched item actually is. Choosing an estimate where the business has to pick one, such as an accrual or a provision. Deciding whether an unusual movement needs explaining to someone outside the team. Signing the close off as complete.
Here the output cannot be checked against the data alone, because the data is the thing in question. A model can propose an answer, but the proposal reads as settled when it is a guess with good grammar. These steps need a named person who has seen the working, can say no, and will be the one asked about it later. In a close, those conditions are not optional.
Where does it go wrong?
Rarely at either end. It goes wrong at the seam, when evidence quietly becomes judgement: the point where the output changes kind, which is where a check earns its place. The variance commentary is the usual place. It starts as a draft built from real movements, and somewhere between draft and pack it gains explanations that nobody checked: a reason for a cost increase that sounds right, a timing difference that was assumed. By the time it reaches a reader, the evidence and the guesswork look the same.
The fix is a hand-off you can see. Mark where the machine's work stops and a person's starts, and keep that mark in the file, not in someone's memory. If a sentence in the pack came from the model, someone should be able to tell, and someone should have decided to keep it. That is disclose and verify applied to a single document.
A test for your own close
Take the steps in your close and ask of each one: if this output were wrong, how would we find out? If the answer is "the next check would catch it", the step is a fair candidate for automation. If the answer is "someone would have to know", it stays with a person, however routine it feels.
This is about workflow design, not accounting advice. If you want a second look at where the hand-off between your tools and your people sits: [email protected]
Where to put the checkpoint: where a check earns its place in a long workflow.
AI projects fail at evaluation, not at the demo: why "it looked right" is not a test.
What an auditor will ask about your AI system: the questions any AI step in your close will eventually face.
TL;DR: In a month-end close, automate the steps whose output can be checked against something outside the model: gathering, matching, comparing, flagging. Keep with a named person the steps where someone must answer for the result: judging exceptions, choosing estimates, signing off. Most failures happen at the seam, when a model's draft explanation becomes accepted fact, so mark the hand-off in the file.
Not legal, security or financial advice. Governed AI performance.