Workflow · Data and research

Clean a spreadsheet and analyse it with AI

A first analysis, with the answer checked by a route that doesn't depend on the AI.

Try with checksIt is worth trying only with the checks on this page.

For: Office workers who know the data, though not necessarily statistics

Go to the steps Test it on your own work

What the evidence says

Studied directly. Studies look at this task or something very close to it.

Data analysis with AI: mixed results (evidence strength 2 of 5)

This is one of the tasks where AI can make work worse, so the checks below matter.

Also worth knowing (not yet graded by us):

In a pre-registered experiment at a public body, the group with AI scored 12% lower for quality on a data task, with no clear difference in time. On a document task in the same study, its quality and speed both improved. Fitzpatrick et al., 2025 (working paper)

This workflow itself is our recommended way of working, not a tested method. Studies searched to 2 November 2026.

Before you start

Keep an untouched copy of the file. Work on a second copy.
Check whether the data contains personal, client or financial details, and whether your employer has approved the tool for them. If unsure, remove those columns.
Write down the question and what you expect the answer to look like, before you look.

The steps

Replace the words in square brackets with your own. Stay in the same chat for the next prompt, unless a step says otherwise.

You
Note the row count and a few totals you can get without AI, such as the sum of a column.

Ask it to list data problems and propose fixes, without changing anything.

Prompt

Here is a spreadsheet of [what it is], [attach or paste]. Each row is [one what]. Do not change anything yet.

List: missing values, duplicates, inconsistent formats, values that look impossible, and columns whose meaning is unclear. Give the row numbers as the sheet shows them, and suggest a fix for each. For anything you would remove, tell me how many rows it affects.

You
Decide each fix yourself, and make it in your working copy. Removing data is a decision, and you record it.

Ask for the analysis with its method and formulas visible.

Prompt

Apply only the fixes I agreed: [list]. Then answer: [question].

Show the formula or steps for each number, in a form I can re-run in the spreadsheet. Say whether you ran each calculation or estimated it. State your assumptions. Report how many rows went into each number. Say if the data can't answer the question, and don't claim one thing causes another from this data alone.

You
Re-work the two most important numbers yourself in the spreadsheet, with a pivot table or a formula.
You
Compare the results with your earlier totals and expectations. Investigate any gap before you share anything.

How to catch its mistakes

Row counts and totals still reconcile with the original, apart from changes you agreed.
Your own re-calculation matches the AI's two most important numbers.
Each percentage is worked out from the right total.
A surprising finding gets a second look, not a headline. It's often a data error.

When not to use this

Published statistics, financial statements or anything regulated, unless a qualified person repeats the analysis.
You can't tell whether the answer is plausible. Get someone who knows the data to look first.
The data is personal and the tool is not approved for it.

Tools

AI built into your spreadsheet, or an assistant that can run code on a file. Examples: ChatGPT for Excel and Google Sheets, Copilot in Excel, Gemini in Google Sheets. We chose these because they are common in workplace software, not because they are the best. We leave out Claude products as examples, because we draft this site with Claude. AI in spreadsheets compared has them all, with dated facts. Use whichever your employer has approved. The steps work with any tool of this type. We take no money from AI companies.

Find out if it works for you

People often misjudge how much AI helps them. Try this on your next 10 similar tasks, letting a coin flip decide which ones use AI. What to measure: Minutes including checks, and errors found when you re-worked the numbers. Your results stay in your browser. Start the self-test →

If you manage people

Require a re-worked number before any AI-assisted analysis goes up the line. Treat a found error as a good outcome.

Written by AI Adoption's editors as our own recommendation, last checked 5 November 2026. It has not been reviewed by an outside practitioner. Before publishing, we ran its prompts once on made-up samples with a second AI model, also from Anthropic. That is a check on the wording, not a test of whether the workflow works for you. Disclosure: AI Adoption uses AI tools in its research and editorial work, including Claude from Anthropic. No company has any say in this page. Reuse under CC BY 4.0. Spotted an error? Tell us.