Ask better questions of a spreadsheet before trusting a chart
Define rows, dates, units and missing values before asking AI for spreadsheet insights, with a small worked example you can calculate yourself.
A useful task, worked through.
A chart can be beautifully wrong when the assistant misunderstands what one row represents. Begin with a definition of the data and a question small enough to check by hand. The first useful result is often an exceptions list, because it reveals which records should not quietly enter the calculation.
Make the row meaning explicit
Use this fictional table: request A, Monday, 30 minutes, completed; request B, Monday, blank minutes, open; request C, Tuesday, 50 minutes, completed. Each row is one service request. The question is average recorded minutes for completed requests in these two days. This is a teaching dataset, not a claim about an actual team.
Write down the units and what a blank means. A blank duration is unknown, not zero. An open request may be unfinished rather than forgotten. If the export can contain several rows per request, say so before asking for a count. Otherwise the assistant may count updates as separate pieces of work.
Demand the admitted rows and the arithmetic
For the stated question, A and C are admitted. Their total is 80 minutes across two completed requests, producing an average of 40 minutes. B is excluded because it is open and has no recorded duration. A result of about 26.7 minutes would suggest dividing by all three rows or treating the blank as zero.
Ask for a small supporting table before a visualization. It should show the included identifiers, excluded identifiers with reasons, total and denominator. Recalculate the example independently. Agreement on a three-row sample is a starting check, not evidence that every column and filter in a large workbook has been interpreted correctly.
A request you can adapt
Analyze this approved sample only. First state what one row means, the units, date window and inclusion rules. For average recorded minutes on completed requests, show admitted and excluded IDs, reasons, total and denominator. Treat blanks as unknown. Do not infer missing values or create a chart until the calculation is checked.
Try the awkward record on purpose
Add request D, Tuesday, zero minutes, completed. If zero is a valid recorded value, the completed-request average becomes 80 divided by three, about 26.7 minutes. If zero means “not recorded” in this system, it should instead be flagged or excluded under an explicit rule. The correct treatment depends on the data contract, not the model’s preference.
Now add a second row with ID A and a different duration. Do not simply delete it as a duplicate. It could be an update, a second service or an error. Ask the data owner how repeated identifiers work. Keep the unresolved rows visible until that rule is known.
Let the chart say only what the data supports
A suitable chart title might be “Recorded minutes for completed requests in the supplied sample.” It should not say “team productivity improved” without comparable periods and a defensible measure. Use the same admission rules for each group, and show when one group has too few records for a useful comparison.
For practice, remove all completed requests. The output should say there are no admitted records and no average, rather than showing zero. Keep that test with the workbook instructions. When you move to real data, protect the original, use your organization’s approved environment and record which file version produced the result.
Take it further with a changed example
When two calculations disagree, first compare which rows they include. A formula may be correct for a different denominator. Write the question in a full sentence, identify admitted rows and explain how blanks, duplicates and open records are handled.
Keep a tiny hand-calculated example beside the larger analysis. If completed requests took 30 and 50 minutes, their mean is 40 minutes. Adding an open request with no recorded duration should not silently lower that completed-request average. The source state and the numerical value are separate facts.
For a further exercise, make every record open. The result for completed requests is “no completed records,” not zero minutes. Then duplicate one completed row and decide whether it represents another event or a repeated record before changing the calculation.
Take this into your next task
Check the included rows and denominator before inspecting the chart.
Questions people ask
Should blanks become zero?
Only if the data owner explicitly defines them that way. Unknown and zero usually need different treatment.
Are repeated IDs always duplicates?
No. They may represent updates or multiple events. Establish the row definition first.
Can AI edit the original workbook?
Use a working copy and an approved workflow. Review the changes and calculated results before replacing a source file.
Can I trust a chart if the formula looks right?
Check the admitted data, units, labels and denominator too. A correct formula can still answer the wrong question or be presented misleadingly.
Sources and editorial notes
Reviewed 2026-09-07. Product capabilities depend on the app, plan, region and workspace settings. Workflows and prompts are authored teaching material, not recorded model results or measured time-saving claims.
Keep going
- AI at Work: Practical Skills, Examples and a Learning Roadmap
- Turn a project update into a presentation with a point
- AI Mission Control: From Idea to Verified Result
Related Power of AI pages
Keep reading with Start here, everyday uses, the tool guide, the writing workshop, and sources and standards.