The AI Financial Reporting Workflow

The reporting pack goes out on business day six. Getting there takes roughly forty hours of somebody’s week, and maybe four of those hours are the part anyone would call analysis.

That ratio is the whole problem with reporting in finance, and it is the thing worth fixing before you buy another tool.

A reporting cycle is five jobs: pull the data, check it, explain what moved, package it, and get it in front of people. AI has changed three of those and left two mostly alone. Knowing which is which is what stops you from automating the wrong half.

The five jobs inside a reporting cycle

The job	What it usually costs	What AI changed
Pull	10 to 15 hours. Exports, refreshes, the one file someone emails you.	It writes the transformation code. It does not connect your systems.
Check	6 to 10 hours. Tie-outs, reconciliations, the hunt for the variance that is a data error.	Good at spotting anomalies. Still needs your definitions.
Explain	8 to 12 hours. Variance commentary, the why behind every red cell.	The largest change of the five. First drafts in seconds.
Package	5 to 8 hours. Slides, formatting, the version control mess.	Deck generation from a structured input works now.
Distribute	2 to 4 hours, plus the follow-up questions all week.	Question answering against the pack, if the data model is clean.

Add it up and the two jobs AI barely touched, pull and check, are more than half the hours. That is why teams who start with the commentary get a fast win and a cycle that is still six days long.

Pull, and where the hours go

Nothing about a model connects your ERP to your planning tool. Somebody still has to own the pipe.

What AI does well here is write the transformation. I describe the shape of the export and the shape I need, and I get working M code back for Power Query, usually right on the second attempt. Six years ago that was a Saturday. Now it is a coffee.

The prompt that works is boring and specific:

Here are the first 20 rows of a trial balance export.
Columns: Entity, Account, Account Desc, Period, Amount, CC.
Period comes through as 2026-08 text, not a date.
I need one row per Entity + Account + Period, amounts summed,
period as a real date, and the CC column dropped.
Write the Power Query M. Explain any step that is not obvious.

Paste the sample. Every time. A model asked to write a transformation without seeing the data invents column names, and you find out four steps later.

If your pull step is the bottleneck, Power Query data cleaning is the deeper version of this, and financial data quality covers the same job when the mess is in the source rather than the shape.

Check, and why this one stays human

Tie-outs are a control. A model that reads your reconciliation and says it looks fine has not performed the control, it has commented on it, and that difference matters the first time an auditor asks who checked.

Where it earns its place is the pass before the pass. Hand it the current month and the prior three, ask what looks structurally odd, and it will surface things a threshold-based check misses. A cost center that stopped posting. An account that flipped sign. A month with 31 days of payroll in a 30-day period.

Two of those are real errors I have caught this way, and one of them was a genuine business change that I would have flagged as an error if I had not asked.

Keep your materiality thresholds. Add the anomaly pass on top. The order matters, because a model given a threshold will anchor to it and stop looking.

Explain, the eight hours you get back

Variance commentary is the best use of AI in the whole cycle, and it is the one most teams are still doing by hand at 9pm.

The trick is what you feed it. A model handed a variance table writes generic sentences about unfavorable variances. A model handed the variance table plus last month’s commentary plus the three things you already know about the business writes something you can edit in four minutes.

You write the monthly commentary for a 40-person SaaS finance team.

Here is the P&L variance table for August, actual vs budget, with prior month
for reference. Here is last month's commentary so you match the voice.

Three things I already know:
- The Boston office lease started 1 August, so facilities steps up permanently
- Two sales hires started late, which is most of the comp underspend
- The marketing overspend is the conference we moved from Q3

Write commentary for every line over $50k or 10 percent, whichever is smaller.
One paragraph per line. Say what moved, why, and whether it repeats.
Flag anything you cannot explain from what I gave you rather than guessing.

That last line is the one that matters. Without it you get confident explanations for variances the model knows nothing about, written well enough that they survive a quick read.

I still rewrite about a third of what comes back. The first draft takes the blank page away, and the blank page was most of the eight hours.

Package, and the version control problem

Deck building from a structured input works now, properly, which was not true even a year ago. Claude will take a commentary document and a set of numbers and produce a deck with the right slide per section. Copilot will do it inside PowerPoint if your organization is on the Microsoft side.

What it does not solve is the fourteen-slide pack that has grown by accretion since 2019 and survives every attempt to cut it. That is a conversation, and no tool will have it for you.

One practical rule: generate the deck from the commentary, never the other way around. Teams that build slides first end up writing commentary to fit the layout, which is how a pack ends up saying nothing at length.

Distribute, and the questions that follow

The week after the pack goes out is where the hidden hours live. Six people ask six versions of the same question, and each one costs you twenty minutes and a context switch.

Self-service works when the data model is clean and fails loudly when it is not. If gross margin means two different things in two systems, letting the business ask questions directly means they get confident, well-formatted, wrong answers faster than before.

So the sequence is: fix the definitions, publish the model, then open the door. Doing it in the other order is how self-service gets a bad name inside a company for the next three years.

For the reporting layer itself, financial dashboards covers what a good pack looks like, and Power BI commentary with Claude is the version where the commentary is generated inside Power BI.

The workflow end to end

This is the version I would hand a new analyst on day one.

1. Close lands. Export the trial balance and the prior three periods to one folder, same file naming every month.
2. Refresh the Power Query model. If the transformation broke, the model wrote it, so the model can read the error message.
3. Run the anomaly pass before the tie-outs, not after. Ask what looks structurally different from prior months.
4. Do the tie-outs yourself. Sign them yourself.
5. Generate the variance commentary with the three-things prompt above. Edit it. Assume a third is wrong until you check it.
6. Generate the deck from the edited commentary.
7. Send it, then log the questions that come back. Those questions are next month’s slides.

Step 7 is the one everyone skips. The questions people ask about your pack are free product research on your own reporting, and after two months of logging them the pack changes shape on its own.

What still needs a person

Three things, and they are the three worth being good at.

• Knowing which variance matters. A model ranks by size. You rank by whether anyone will do something about it, which is a different list.
• The number that is wrong in a way that looks right. Reporting errors that survive are the plausible ones, and plausible is exactly what a language model is built to produce.
• Saying the uncomfortable thing in the meeting. No tool ships that.

Everything else in the cycle is fair game.

Where to start this month

Pick the commentary. It is the fastest hours back, it needs no IT involvement, and you can test it against a month you have already written so you can see the gap between its draft and your final.

Do that comparison once, properly. It will either save you a night a month or it will not, and you will know inside an hour instead of arguing about it for a quarter.

If you want the wider map, AI in finance covers where reporting sits among the other workflows, and automating a month-end report is the same exercise applied to the close itself.