AI and the month-end close: 5 tasks to hand over, 3 to keep
Where AI already does better than a junior accountant in the month-end close, where it must never act alone, and the control framework that makes the difference. A former auditor's view.
I started my career as an auditor at Arthur Andersen. Since then, I have seen the month-end close from three seats: the auditor who checks the numbers, the controller who produces them, and the finance leader who has to defend them in front of management and the board. Today I sit in a fourth one: I implement the ERP systems that generate those numbers.
From those four seats, one thing is clear about AI in finance. The enthusiasts are wrong about what it will replace. The skeptics are wrong about how fast it is coming.
AI will not replace finance professionals. It will replace the parts of the close we should never have been doing by hand: matching, chasing, re-keying, formatting. What it will not replace is judgment, professional skepticism and accountability. Exactly what an auditor is trained to protect.
Here is how I would split the work today.
Five tasks where AI already beats a junior accountant
1. Bank reconciliation
Hundreds of bank lines, truncated references, partial payments, one customer paying three invoices in a single transfer. This is pattern matching at volume, and it is precisely what AI does well. It proposes matches in seconds, with a confidence level. The accountant’s job shifts from matching everything to reviewing the exceptions.
The control: only matches above a defined confidence threshold are proposed for bulk validation; everything else goes to a human queue. The person who validates is named in the system.
2. First drafts of variance commentary
“Why did operating expenses rise this month?” Every controller knows the drill: open the general ledger, filter by account and cost center, find the three entries that explain the gap, write two sentences. AI can do the digging and draft the commentary from the ledger detail.
The control: the draft is a draft. The controller corrects it, because AI explains what moved, not always why. “Rent increased” is a fact; “rent increased because we signed the new warehouse lease early” is knowledge.
3. Identifying accrual candidates
Goods received but not invoiced, recurring contracts with no invoice this month, utility bills that usually arrive on the 15th and have not. AI can scan purchase orders, contracts and history to list what is probably missing.
The control: AI lists candidates; the accountant decides what is booked and at what amount. An accrual is an estimate, and estimates belong to people.
4. Supplier invoice coding
Reading a PDF invoice, recognizing the supplier, suggesting the account, the cost center and the tax code: this is now mature technology. On recurring suppliers, the suggestion is right most of the time, and it learns from corrections.
The control: new suppliers, unusual amounts and changes to bank details are always routed to human review. That last point is not about AI; it is about fraud.
5. Anomaly detection
Duplicate invoices, suspiciously round amounts, entries posted on a Sunday night, a first payment to an unknown supplier, an account that suddenly moves after months of silence. A human reviewer looks at a sample. AI can look at every single entry.
The control: each anomaly flagged is either explained or corrected, and the explanation is kept. An unexplained alert that someone simply closed is worse than no alert at all.
Three things AI must never do alone
1. Judgment-based estimates
Provisions for bad debts, inventory write-downs, impairments, revenue cut-off on a contract that spans two periods. These require context that is not in the ledger: a customer’s situation known through a phone call, a commercial negotiation in progress, a legal opinion. They also require someone accountable for the figure.
2. Final posting without review
AI prepares, a named person approves. Segregation of duties does not disappear because the preparer is a machine. If anything, it becomes more important: a model that makes a systematic error repeats it a thousand times, without fatigue and without doubt.
3. Anything it cannot explain
If an output has no audit trail, it does not go into the books. Your auditor will ask how a figure was produced; “the AI suggested it” is not an answer. Every AI-assisted entry should keep its source documents, the suggestion, who validated it and when.
The rule that ties it together
AI prepares. Humans decide. Every decision leaves a trace.
This is not a technology principle; it is the internal control principle finance has always applied, extended to a new kind of preparer.
Where to start
You do not need a large project to begin. In practice, I recommend three steps:
- Map your close calendar and mark, for each task, the time spent and whether it is repetitive or judgment-based. The repetitive, high-volume tasks are your first candidates.
- Start with bank reconciliation or invoice coding, where results are easy to measure and errors are visible quickly.
- Write the control before you switch on the tool: who reviews, above which threshold, and where the evidence is kept. Then show it to your auditor before year-end, not after.
Modern ERPs, Odoo among them, already include part of these capabilities in their accounting modules. The tool matters less than the framework around it.
The finance teams that will benefit most from AI are not the ones that automate the most. They are the ones that can still explain every number they sign.
What is the most time-consuming step in your month-end close today? I would like to hear it: reply on LinkedIn or write to us.