Analyse a spreadsheet with AI: ask, chart, then check

How to do it the first time

When an AI analyses a spreadsheet, it usually doesn’t just glance at it and report a number. It writes a small program or a formula to do the calculation, then explains the result. That makes it more reliable than mental arithmetic. But if it reads the wrong column, skips blank rows or treats text as numbers, every conclusion after that goes off course, and it still looks perfectly organised.

1. Tidy the sheet before uploading. Put column names in the first row (“Date”, “Store”, “Sales”), keep one kind of thing per column, and put units in the column name (“Sales (£)”). Remove personal details the analysis doesn’t need.
2. Ask it to describe the data before asking for conclusions. “How many rows and columns are there? What type is each column? Are there blanks, duplicate rows or outliers?” Compare that with what you see when you open the file yourself, and only carry on once they match.
3. Ask specific questions in plain language. “Which store grew fastest from July to September? Show it month by month” works far better than “analyse this”. Say how to handle the data: skip blanks or count them as zero, subtract returns or not.
4. For charts, say what question the chart should answer. For example: “A line chart comparing the two stores’ monthly sales, months along the bottom, amount up the side, and mark any month with missing data.”
5. Recompute a few key numbers yourself. In Excel or another spreadsheet app, use sum, count and filters to work out one or two totals and one percentage. If they match, read its conclusions. If not, ask it to show its steps or formulas and find where the difference comes from.

Prompts to copy

Replace the parts in square brackets with your own situation, then attach the spreadsheet.

I've uploaded a spreadsheet of [short description, e.g. sales records for three stores, July–September 2026].
Step one: don't analyse anything yet. Just tell me:
- the number of rows and columns, and what each column means and its data type;
- how many blanks, duplicate rows and obvious outliers there are, and where.
Step two: once I've confirmed, answer this question: [your specific question]
Requirements:
- Blanks: [skip them / count as zero / ask me first]
- Show the calculation method or formulas so I can check them myself.
- If you make a chart, say which question it answers.
- In your conclusions, separate what the data shows directly from what you're inferring.

Example

Check and adjust

• The row count and column names match what you see in the file; no sheet was skipped and the header row wasn’t read as data.
• You’ve recomputed at least two or three key numbers yourself and they match.
• You agree with how blanks, duplicates and outliers were handled.
• Chart axes, units and time ranges are right, and the vertical axis doesn’t start halfway up and exaggerate differences.
• The conclusions don’t go beyond what the data can show: a spreadsheet tells you what happened; why needs other evidence.

If the result is off, you can follow up like this:

Your figure of [number] doesn't match my own calculation of [my result].
Show me how you calculated it: which columns and rows you used (and any filters), how you treated blanks and duplicates, and the exact formula or code.
Once you've found the difference, recalculate and tell me whether any other conclusions are affected.

Next steps

Once you’re comfortable, ask it for a spreadsheet you can reuse, such as a summary sheet with formulas that updates when you paste in next month’s data. Claude and ChatGPT can both create Excel files to download; Gemini in Google Sheets and Copilot in Excel work directly in your sheet. Whichever you use, open every sheet and click into a few formulas before you hand it on.