Write budget variance commentary with AI and source evidence

This workflow turns a reconciled variance table into draft management commentary. It is useful when the arithmetic is complete but the explanation still requires repeated drafting. The result is a short narrative, a list of evidence gaps and a review record. It is not an automated close or a substitute for an accounting judgment.

1. Freeze the input and calculate outside the language model

Use one reporting period and one currency. Export a read-only snapshot with stable row identifiers. Calculate actual minus budget and percentage variance in the spreadsheet or reporting system. For a zero budget, mark percentage variance as not applicable rather than dividing by zero. The following fictional costs are in thousands of dollars; positive variance means overspending.

For live work, evidence IDs should resolve to a controlled source or a named owner's approved explanation. Exclude information the drafting service is not permitted to process. A familiar product name does not establish that the organization's configuration is approved for financial data.

2. Ask for an explanation with explicit gaps

Draft management commentary using only the supplied table and driver notes.
Treat the calculated figures as fixed; do not recalculate or change them.
Return: (1) a paragraph of at most 100 words, (2) unresolved questions,
(3) a claim-to-source list using row and driver IDs.
Distinguish timing differences from permanent savings.
Do not invent a cause, forecast, materiality judgment or corrective action.
If a cause is missing, say it remains unconfirmed.

3. Compare with the expected result

Total costs were $142,000 against a $140,000 budget, an overspend of $2,000 (1.43%). Software was $12,000 over budget: approved additional seats explain $8,000; the remaining $4,000 is unconfirmed. Travel was $10,000 below budget because a planned trip moved to October. That variance reflects timing rather than established permanent savings.

This is a manually prepared answer key, not a captured product result. Its claims map to R1/D1, R2/D2 and the total row. A draft that describes “software inflation” or “sustained travel efficiencies” fails the example even if every number is correct. The unresolved question is what explains the remaining software variance.

4. Review, then circulate

Have the analyst run the claim-checking task. The controller then reviews the drivers, reporting perimeter and language. Store the input version and corrections with the approved paragraph. Move it into the board pack only through the normal approval process.

Tool choice and success measure

An approved assistant that can work with the fixed table is sufficient. Copilot in Excel may reduce copying for an existing Microsoft user, but the spreadsheet calculations and reviewer remain the source of truth. No additional purchase is needed to try the fictional example.

Compare total preparation plus review time against the normal process. Count unsupported explanations and corrections separately. Repeat with a zero-budget row, a reclassification and an incomplete driver note before extending the pilot. Return to AI for CFOs to choose the next finance task.

If the exercise reveals a broader training need, the CFO guide also discusses an optional finance-workflow course. Completing this example does not require a course purchase or a new software subscription.