How do I check the numbers in an AI-generated business report?

1. Write down what the number is supposed to measure

Start with a sentence someone might act on, such as “Lunch orders rose 40% this week.” Write down the metric, the dates, and the group it covers. Are you counting orders placed or orders completed? Which locations count? Does lunch include every order before 3 p.m., or whatever the source sheet calls lunch?

Think of it like counting meals from kitchen tickets. An open ticket, a canceled ticket, and a paid ticket each tell you something different. If your report measures completed orders, use the original export and keep only the rows that meet that definition. A draft can offer row IDs or cell references as a starting point. Open them yourself and check the row, column, and status.

The NIST generative AI risk profile describes how AI can present incorrect information confidently, a risk often called hallucination. It recommends checking output against known reference data. Here, that means using the dated spreadsheet or system export your team recognizes as the source. If you can’t get it, mark the figure as unverified and ask whoever owns the data.

2. Recalculate from the rows you can verify

Find the same dates, location, and order status in both periods. Check what one row represents. If a sheet has a line for each dish, count unique order IDs to get orders. If it already has a row for each order, check why an ID repeats before taking it out. Keep the filters and the calculation beside the corrected number so someone else can repeat your check.

For a percent change, subtract the earlier count from the later count, divide by the earlier count, and multiply by 100. Moving from eight orders to ten is two additional orders, or 25%. If the earlier count is zero, report the two counts and the change in orders. Dividing by zero cannot give you a percent increase.

For a large sheet, a spreadsheet filter and count or a pivot table can help you reproduce the total. Check that the filter includes the right rows and that your count uses the right field. Use an approved work tool for real records, and use the fictional rows below to practice.

3. Check what the report says caused the change

Suppose the report says lunch orders rose because of a coupon. Your numbers can support the increase, and the coupon may be a reasonable possibility. A before-and-after count alone leaves questions about what else changed, including regular foot traffic or opening hours. Share the count and note that the cause needs more evidence.

That final decision is part of staying human-in-the-loop. If the claim is wrong, correct the metric and keep the source and calculation with it. If the explanation is uncertain, say what you would need to check next. If a report cites an outside source, use this citation-checking guide to verify the claim separately.