The five-minute close: what actually gets automated first

Every finance leader who has looked seriously at AI has heard some version of the pitch: hand the close to a model and get your week back. In practice, almost nobody does that first — and the teams who try usually end up re-doing the work by hand anyway, just later and angrier. The close isn't one task. It's a chain of small judgments wrapped around a much larger pile of repetitive lookups, and those two things automate at completely different speeds.

The teams getting real time back from AI aren't automating "the close." They're automating the parts of the close that were never really judgment calls to begin with, and leaving the parts that are exactly where they've always been — with a human, and usually a senior one.

Start with reconciliation, not review

Bank reconciliations, intercompany matching, and subledger-to-GL ties are the highest-volume, lowest-judgment work in the entire close. A model doesn't need to understand your business to flag that two entries differ by exactly a rounding error, or that an intercompany charge posted on one side and not the other. This is pattern-matching at scale, which is precisely what these systems are good at.

This is also the safest place to start because the failure mode is cheap. If the model misses a match, a human still reviews the exception queue — nothing gets posted without a person looking at it. You're not removing a control, you're removing the tedious first pass that used to happen before the control.

Then move to variance narratives, not variance decisions

The second tier is drafting, not deciding. Once actuals land, someone has to explain why marketing spend ran 14% over budget or why gross margin moved two points in a quarter. That explanation takes real time to write even when the underlying "why" is already known to the team — it's a documentation problem more than an analytical one.

This is where a model earns its keep fastest: pulling the relevant drivers, drafting the first version of the variance commentary, and leaving a controller to edit rather than originate. The controller still owns the number and the story. The model just removed the blank page.

The goal isn't a close with no humans in it. It's a close where humans only spend time on the parts that actually needed a human.

Forecasting comes third, and it should feel slower than you expect

Continuous, AI-assisted forecasting is the most talked-about use case and the one worth automating last — not because it's less valuable, but because it depends on the first two layers already being trustworthy. A forecast built on reconciliations nobody fully reviewed, or variance narratives nobody validated, just launders bad data faster.

Once reconciliation and variance drafting are running cleanly for a quarter or two, forecasting automation has something solid to sit on top of. Teams that reverse this order — chasing a flashy forecasting model before the close underneath it is clean — are usually the ones that quietly stop trusting the output within two cycles.

What to leave alone, at least for now

• Journal entry approval. Drafting is fair game; approval authority should stay human and stay auditable.
• Anything touching revenue recognition judgment calls. Pattern-matching doesn't replace technical accounting judgment on edge cases.
• The board narrative. A model can draft supporting detail. The story you tell the board about the quarter should still be yours.

The order matters more than the tooling

Most of the disappointment we hear about "AI in finance" isn't really about the models — it's about sequencing. Reconciliation and drafting first. Forecasting once the data under it is trustworthy. Approval and judgment last, if ever. Get the order right and the close gets shorter almost as a side effect, without anyone having to defend a black box to an auditor.