Ask Questions of a Spreadsheet with AI (and Check the Math)

Ask a spreadsheet one question and get a short finding with a chart and numbers you've rechecked.

Level 2 · Assistant Free · 30 minutes · Updated Oct 31, 2026

“Upload your spreadsheet and ask AI what’s going on” sounds like magic, and sometimes it is. In two minutes you can get trends, a chart and a neat explanation. But a confident paragraph about your numbers is only as good as the math behind it, and AI can pick the wrong column, skip a row, or explain a pattern that isn’t really there. In the next 30 minutes you’ll ask a real question of a real dataset, make the AI show its work, and check two numbers yourself in Excel.

Remember the intern: smart, fast, first day on the job. They’re happy to crunch numbers, but they don’t know what your columns mean, which rows are odd, or what your manager will ask. And like any new analyst, they need to show their working before anyone trusts the result.

What you’ll get

What you need

• Claude, any plan. Claude analyzes data by writing and running code. This is on by default; if it isn’t working, check that Code execution and file creation (called Create and edit files in newer versions) is switched on in Settings > Capabilities. Other tools work too; see the end.
• Excel or Google Sheets, for checking two numbers yourself.
• A dataset. We’ll use a real, public one: E-Commerce Retail Sales from the US Census Bureau, quarterly, in millions of dollars, seasonally adjusted. Download it as a CSV from FRED, the St. Louis Fed’s data site (series ECOMSA). Swap in your own spreadsheet when you’re ready.
• A question. Ours: is US online retail still growing as fast as it used to, and which quarters stand out?

With your own data, the usual boundary applies: only upload what your company allows in this tool. Customer-level or salary data usually doesn’t belong here.

The three rules

Part 1: One question, answered with proof (30 minutes)

Six steps. Copy the prompts as they are, then swap in your own data and question.

Step 1. Look at the data yourself (3 minutes)

Open the CSV in Excel or Sheets before the AI sees it. Write down:

• What each column is, and its units. (Here: a date, and e-commerce sales in millions of dollars.)
• The first and last date, and roughly how many rows.
• Anything odd: blank cells, totals mixed in with data, notes at the top.

Step 2. Upload it and make it describe the data back (5 minutes)

In Claude, click +, add the CSV, and send:

This file has quarterly US e-commerce retail sales from the Census Bureau, in millions of dollars, seasonally adjusted.
Before any analysis, describe the data back to me: column names, units, first and last date, number of rows, and any missing or odd values.

Compare its answer with your notes from Step 1. If the date range, row count or units don’t match, stop and correct it. Everything later depends on this.

Step 3. Ask your question, and ask for the working (6 minutes)

My question: is US online retail still growing as fast as it used to, and which quarters stand out?
Please:
1. Calculate year-over-year growth for each quarter (this quarter vs. the same quarter a year earlier).
2. Compare average growth in 2015-2019 with 2022 to the latest quarter.
3. List the 3 biggest quarter-over-quarter increases and decreases.
Use code for all calculations. For each result, tell me the formula and the exact rows (dates) you used.

The last line is the rule “show the working.” It turns “trust me” into something you can check in Step 5.

Step 4. Ask for a chart and the caveats (4 minutes)

Make one line chart of year-over-year growth by quarter, with the 2015-2019 average as a horizontal line.
Then list the caveats: what this data doesn't measure (for example, is it adjusted for inflation?), and anything unusual in the data that could distort the comparison.

The caveats question matters. In this dataset, a likely one is that the pandemic years will look extreme for reasons that have little to do with long-term trends. Check anything the AI claims about how the data is built (like inflation adjustment) against the source’s own notes on FRED or the Census website.

Step 5. Recalculate two numbers yourself (7 minutes)

Pick the two numbers that matter most for your answer. For example:

• The latest year-over-year growth. In Excel, if the latest quarter’s sales are in cell B108, the same quarter a year earlier is four rows up, in B104, and the formula is =B108/B104-1. (Your row numbers may differ; what matters is “same quarter, one year earlier”.)
• One of the “biggest changes”. Find those two dates in your sheet and do the same calculation.

If your number matches, good. If it doesn’t, ask: “For [quarter], I get X. Which rows did you use?” The difference is usually a wrong row, a wrong column, or a different definition of “growth.”

Step 6. Write the finding in three sentences (5 minutes)

Write the finding for my manager in 3 sentences: the answer to my question, the one number that matters most, and the main caveat. Only use numbers I've checked or that you calculated with code above.

Then read it once more against Steps 4 and 5. A finding without its caveat is a claim, not an analysis.

Part 2: Repeat it every quarter (10 minutes to set up, then about 10 minutes per update)

This data is updated every quarter, like many work spreadsheets (sales reports, budget trackers, survey results). Instead of rewriting the prompts each time, put them in a Project with standing instructions.

1. Start from a run that worked. Only set this up after Part 1 gave you numbers you checked.
2. Create a project. In Claude’s sidebar, choose Projects, then + New Project. Name it “Quarterly e-commerce check.”
3. Set the instructions. Click Set project instructions and paste:

In this project I upload an updated CSV of quarterly US e-commerce retail sales (Census Bureau, millions of dollars, seasonally adjusted).
Every time:
1. Describe the data back to me first: columns, units, first and last date, row count, anything odd. Wait for my OK.
2. Calculate year-over-year growth for each quarter, the 2015-2019 average, and the average since 2022.
3. Say what changed since the previous quarter's result, if I paste it in.
4. Use code for all calculations and give the formula and rows for each number.
5. List caveats.
6. Write a 3-sentence finding.

4. Each quarter: open the project, start a new chat, upload the new CSV, and paste last quarter’s three-sentence finding.
5. Still redo two numbers (Part 1, Step 5). The instructions make the analysis consistent; they don’t make it correct.

Why you can’t skip the checking

AI is good at finding a value in a table, and noticeably weaker at everything that comes after. Researchers have measured that:

• Simple lookups are fine; harder math is not. A 2026 benchmark tested ten leading models on questions about financial spreadsheets. Simple lookups were answered correctly about 89% of the time. But tasks like medians and percentiles were right only about 20% of the time across all models (33% for the best three), and accuracy fell further on the largest spreadsheet. The authors concluded that no model was accurate enough to use unsupervised for professional finance work (Ravnik et al., FinSheet-Bench, 2026). The data was synthetic and the models read the sheets as text.
• Messy tables throw it off. In another 2026 study, small distortions in a table (the kind real spreadsheets have) cut accuracy by at least 22% even for a top model, and models usually didn’t notice the problem on their own (Dutta et al., 2026). That’s what the “tidy the sheet” box is for.

Both tests mostly had models read the table as text. Asking Claude to calculate with code (Step 3) helps, because the arithmetic is then done by a program. But code can still use the wrong column, the wrong rows, or a different definition of growth. That’s why Step 2 checks how it read the data, and Step 5 has you redo two numbers by hand.

Checklist

Before you share the finding:

• The AI’s description of the data (columns, units, date range, row count) matched what I saw
• Every number came from a calculation it showed: formula and rows
• I recalculated two key numbers myself, and they match
• Units are stated (millions of dollars, percent)
• Caveats are included, and claims about how the data is built are checked against the source
• The chart shows what the sentence says, with labeled axes
• Nothing in the upload was data I’m not allowed to share with this tool

What you just learned

On the surface you analyzed a spreadsheet. Underneath, you practiced checking in the form that matters most with numbers: tracing every result back to rows and a formula, and redoing a sample yourself. You also practiced giving context: units and definitions the AI can’t guess.

The same method works for: finding which region or product is behind a sales dip, comparing this year’s budget to last year’s, summarizing survey results, spotting unusual expenses, and checking a colleague’s report before it goes out. Change the question in Step 3; keep “describe the data back” and “redo two numbers.”

Using a different tool?

The prompts work anywhere that can run calculations. Only the buttons change.

Features and plan names change often. This table was checked in October 2026; the ChatGPT, Gemini and Copilot rows still need checking against their help pages.

Sources

• E-Commerce Retail Sales, series ECOMSA (FRED, US Census Bureau data, updated Aug 2026)
• Create and edit files with Claude (Claude Help Center)
• FinSheet-Bench: where LLMs break on financial spreadsheets (Ravnik et al., arXiv, Mar 2026)
• Robustness of LLMs under tabular distortions (Dutta et al., arXiv, Jan 2026)