Begin with a copy
Keep the original workbook unchanged. Create a small anonymized table with clear column names, units, and dates. Explain what one row represents. A column called “amount” means little without a currency, period, and definition.
Ask for questions before conclusions
Have AI suggest three useful questions and identify missing or ambiguous data. Choose one. For a fictional event budget, you might ask which category exceeds its allocation—not whether the event is “successful.”
Keep the calculation visible
Ask for a spreadsheet formula, an explanation of its inputs, and a small worked example. Paste the formula into your copy and calculate a few rows by hand. Check empty cells, refunds, mixed currencies, and duplicate records.
Explain the limits
A chart can look convincing while using the wrong denominator. Keep units, filters, exclusions, and sample size beside your result. Treat AI as a way to propose an analysis you can inspect, rather than as the authority over your numbers.
Make it practical, step by step
Make a safe sample
Use synthetic or anonymized data. Describe each column, its units, and what a row means.
Choose one question
Ask for possible analyses and data-quality concerns before requesting a result.
Verify the calculation
Use visible spreadsheet formulas and check a few rows manually, including blanks and negatives.
Save the assumptions
Record filters, exclusions, and what the dataset cannot establish.
A prompt to take with you
Here is a small anonymized table: [data]. One row means [definition]. Units: [units]. Suggest three questions this data can answer and flag data-quality issues. For the question I choose, show a spreadsheet formula, explain each input, and provide a hand-checkable example. Do not infer missing values.
A few things to keep in mind
- Do not upload payroll, customer records, or confidential financial data without authorization.
- Generated formulas and conclusions require independent checks.
Sources & further reading
- Microsoft: Analyze Data in Excel (23 Oct 2026)
This is an idea to try; a recorded hands-on test isn’t available yet.
How we check and correct our work