Jeremy is the finance manager at a 50-person F&B group with three outlets. The month's numbers are closed by day five. Then the commentary takes another two days. Which variances matter? What caused them? What does management need to decide?

AI can help with financial reporting by drafting that commentary from a checked reporting pack. But there's one line it mustn't cross: describing a movement is not the same as explaining its cause. AI can produce a convincing explanation with no evidence behind it.

Here's a practical workflow that keeps the numbers, and the story, honest.

I. Prepare one agreed reporting pack

Confirm the period, entity, currency, units and comparison. Actual versus budget answers a different question from actual versus last month, so don't mix them.

Keep all calculations in your accounting system or a controlled spreadsheet. AI writes the words, not the maths. For each reporting line, supply the actual, the comparison figure, the variance and a reference the reviewer can trace. Add business notes separately, with their owner and date.

Use only an AI tool your company has approved for financial data, and remove personal or customer details the commentary doesn't need.

II. Separate facts from explanations

Here's an illustrative revenue line from Jeremy's pack:

  • Budget: S$200,000

  • Actual: S$180,000

  • Variance: S$20,000, or 10%, below budget

  • Verified note: a S$12,000 catering order moved to next month

A good draft says: "Revenue was S$20,000 below budget. The confirmed catering order deferral accounts for S$12,000. The remaining S$8,000 needs investigation."

It shouldn't fill that gap with "weaker demand" unless there's evidence. The same applies to margins, staffing and suppliers. And if the comparison figure is zero, show the actual movement and mark the percentage as not applicable.

III. Use a prompt that keeps commentary traceable

"Draft internal management commentary using only the checked figures and approved business notes provided. Keep the period, currency, units and comparison basis. In reviewer notes, reference the reporting line and source note for each statement. Separate confirmed explanations from questions to investigate. Don't invent causes, forecasts or recommendations the evidence doesn't support."

Ask for commentary only on the lines management needs to discuss. A paragraph for every small movement makes the pack harder to read. Jeremy decides which items matter and the thresholds for them.

IV. Check the commentary before it's shared

Review five things before approval:

  • Figures: Do the amounts and percentages match the controlled calculations?

  • Direction: Are increases, decreases and favourable or unfavourable movements described correctly?

  • Context: Are the period, entity, currency and units consistent throughout?

  • Causes: Does each explanation have evidence, with any unexplained amount left visible?

  • Actions: Are suggested next steps kept separate from decisions already made?

Watch costs in particular. A lower expense isn't automatically good news if necessary work was simply delayed.

V. Keep a record and improve each month

Save the inputs, draft, corrections and approved commentary with the reporting pack. If the numbers change after review, check the commentary again.

Track the time spent and the corrections made each cycle. If reviewers keep fixing the same issue, change the input layout or the prompt before using AI for more of the report. This workflow is for internal management reporting, not statutory filings.

VI. Common questions

Can AI do variance analysis on its own? It can describe movements and draft commentary. Keep calculations in your accounting system, and get causes from people who know the business.

Is it safe to use financial data with AI? Only use tools your company has approved for that data, and share only what the task needs.

For more on building workflows like this, see our beginner's guide to AI automation. To build one around your own reporting, explore our AI for Finance Teams programme or book a discovery call.

Jeremy is an illustrative example. All figures are hypothetical.