Analyze a spreadsheet by asking questions in plain English

A spreadsheet full of real answers and no formula to pull them out. You don’t need to learn pivot tables for a question you’ll only ask once. You need the answer, in plain language, with enough transparency to trust it.

Handle the data safely first

If the sheet has customer names, employee details, or financial specifics you wouldn’t want exposed, strip or replace them before pasting anything into a chat tool: swap real names for “Customer A, Customer B,” round or generalize sensitive numbers, and remove anything that identifies a specific person. The structure and patterns are almost always what you need; the identifying details usually aren’t.

Do this

1. Say what you’re trying to learn, specifically. “What’s going on with this data” is too broad; “which product lines lost money in Q2” is answerable.
2. Paste the relevant columns, not the whole workbook, and describe what each column means if it’s not obvious from the header.
3. Ask for findings, anomalies, and follow-ups, using the prompt below.
4. Ask it to show its work for any specific number it states, not just the conclusion.

Here's the data (columns: [describe each column briefly]):
[paste the relevant rows/columns, with identifying details removed
or replaced]

What I'm trying to learn: [specific question]

Give me:
1. The direct answer, with the specific numbers behind it
2. Anything unusual or unexpected in the data, even if I didn't
   ask about it
3. Follow-up questions this raises that the data alone can't answer

Show your work: if you calculate a total, average, or trend, state
which rows or values it came from.

Check this before you trust it

• Every specific number against the raw data. Spot-check two or three claimed figures by hand or in the actual spreadsheet before trusting the rest.
• Whether it actually used all the rows you gave it. A large paste can get partially processed; ask directly how many rows it analyzed.
• Any anomaly it flags, before acting on it. An “unusual” pattern is sometimes a real finding and sometimes a data entry quirk you already know about.

If it goes wrong

A stated number doesn’t match the spreadsheet when you check it. Ask it to show the exact rows and math behind that number. If it can’t reconstruct its own answer, don’t trust the original figure.

It only analyzed part of what you pasted. Very large pastes sometimes get truncated silently. Split the data into smaller chunks and ask the same question of each, then combine the answers yourself.

The findings are technically correct but not actually useful. That usually means the question in step 1 was still too broad. Narrow it to the specific decision the analysis needs to support.

You’re not sure whether something counts as sensitive enough to strip out. Default to removing it. A generic label costs you nothing; a real name or account number in a chat log is not easily undone.