How to audit an AI-generated financial report before you send it

Before you read the conclusions

Start with the inputs. A wrong input makes every conclusion after it wrong in a consistent, convincing way.

• Confirm the file and sheet: the right company, the right year, not the prior-year tab.
• Confirm the period: the start and end dates the report used match the period you asked for.
• Confirm the row count is plausible for that period.
• Check for totals rows in the source export — accounting exports often include them.

Reconcile the anchors

Pick two or three numbers you can get independently and make them match before reading anything else.

• Total revenue for the period against your accounting system's own report.
• Ending cash against the bank balance for the same date.
• One category total against a SUMIFS you run yourself.

Check the definitions

Most disagreements between an AI summary and a finance team are definitions, not arithmetic. Gross or net revenue. Whether refunds and chargebacks are subtracted. Cash or accrual. Whether a cost line belongs in COGS. The report should state each one; if it does not, ask it to and re-check the affected figures.

Trace every figure in the narrative

Read the written summary last. For each number in a sentence — "revenue grew 18%", "marketing is 22% of spend" — find the table row it came from. A figure you cannot trace is a figure the model wrote rather than computed, and it should be removed or recomputed before the report goes out.

Before you send it

Three last checks that prevent the most embarrassing errors:

• Percentages add up where they should (shares of a whole to 100%).
• Growth rates use the right base period.
• Charts match the tables they claim to show.

Common questions