Where AI helps in the month-end close, and where it doesn't

Most finance teams ask the same question about AI: where do we start? The answer is less exciting than the sales decks suggest, and more useful.

The month-end close is a good place to look. It is repetitive, time-pressured and well documented. It is also where a mistake ends up in the board pack, so the test for any automated step is whether you can prove it was done correctly.

Where AI earns its place

• Matching and reconciliations. Bank, intercompany and supplier statement matching is pattern work. Rules handle the exact matches. AI is useful on the near misses, such as a payment reference typed slightly wrong or one receipt covering three invoices.

• Reading documents. Invoices, contracts and remittance advices arrive as PDFs and emails. Extracting the supplier, amounts, dates and purchase order number is now reliable enough to replace keying, provided a person reviews the exceptions.

• First drafts of commentary. Given the numbers and last month's narrative, AI will draft variance commentary in seconds. The analyst's time moves from writing to checking, and to explaining what the numbers cannot.

• Spotting what looks wrong. A journal posted to an unusual account, a duplicate invoice, an accrual that has not moved in six months. AI is good at raising the flag. A person decides what it means.

Where it doesn't

• Judgement. Provisions, impairments, revenue recognition on an unusual contract. AI can gather the evidence and summarise the standard. The decision and the accountability stay with a qualified person.

• Anything you cannot evidence. If an auditor asks why an entry was made and the answer is that the model suggested it, the control has failed. Every automated step needs an input, a rule or prompt, an output and a named reviewer you can show.

• Messy master data. AI does not fix a chart of accounts with three codes for the same thing, or a supplier list full of duplicates. It reproduces the mess faster.

• Final sign-off. The close is a statement that someone stands behind the numbers. That cannot be delegated to software.

Check what your ERP already does

Before buying anything, look at what you already own. Most modern ERP systems include bank matching rules, recurring journals, approval workflows and scheduled reports, and most finance teams use a fraction of them.

Native tools are usually easier to control and audit because they sit inside the system of record. I use AI for the gaps: unstructured documents, near matches and narrative.

A simple test for any close task

1. Is it repetitive, with clear inputs and a known right answer?

2. Can a person review the exceptions without redoing the work?

3. Can you show an auditor what happened and who approved it?

Three yeses, and the task is a candidate for automation. If the third answer is no, fix the control design first.

Start with one process

Pick one process, usually a reconciliation. Measure the hours before and after, and keep the evidence. That gives you a result you can defend and a template for the next one.