How to analyze your sales spreadsheet with Claude and find where money leaks out

You have a monthly sales export: rows of dates, products, amounts and reps. You open it, look at the total at the bottom and close it, because it is not clear what to do with that number next. The spreadsheet honestly shows how much you earned, but says nothing about where you could have earned more. Claude (an AI assistant by Anthropic, claude.ai) reads spreadsheets like this and pulls out what the eye misses: which items hold your revenue, which rep is lagging, on which days orders quietly leak away. Below is a step-by-step way to do it so the result goes straight to work.

Why does a raw sales spreadsheet tell you nothing?

A raw export is hundreds of rows with everything piled together. The grand total answers one question: how much money came in. But a business runs on other questions. Which five products bring half the revenue, and which just sit in the warehouse. How much you gave away in discounts. Which weekdays consistently bring fewer orders. To answer, you need to group, compare and spot outliers, and that means pivot tables, formulas and an hour you usually do not have. Claude does that grunt work for you, and you read the conclusions right away.

How do you prepare the spreadsheet so Claude understands it?

One rule: cleaner data means a sharper breakdown. Three quick steps.

1. Export sales to an Excel or CSV file (a plain text table format where values are separated by commas). Almost any accounting system and any marketplace can produce that export.
2. Check the column headers. They should be clear: date, product, quantity, amount, rep, channel. If the top row reads like gibberish such as col1 and col2, Claude will have to guess what they mean.
3. Attach the file to your message instead of pasting a thousand rows into the chat box. With a file, the model works more carefully and sees the whole table.

Strip out customer personal data, phone numbers and anything that must not leave your systems before you upload. To find patterns, products, amounts and dates are enough.

What prompt gives a breakdown instead of a retelling?

If you just write "analyze this spreadsheet," you get a polite retelling of what is already obvious. The real breakdown starts when you ask business questions. Take this template and plug in your own columns:

`` Here is a monthly sales export (file attached). You are a financial analyst. Analyze it and answer point by point: 1. Top 5 products by revenue and their share of the total. 2. Products that barely sold, candidates to drop from the range. 3. Average order value and how it shifts by weekday. 4. If there is a discount column: how much revenue was lost to discounts. 5. Compare reps by total sales and by average order value. At the end, give three conclusions on what to change next month. If some data is not in the table, say so honestly instead of inventing it. ``

That last line matters more than it looks. It stops the model from filling gaps with made-up numbers and asks it to flag what is missing instead.

How do you find where the money actually leaks out?

Once the general breakdown is ready, dig into the weak spots with pointed follow-ups in the same chat. Claude remembers the spreadsheet and answers from it.

• "Show products with high returns relative to sales." Returns often eat profit more quietly than you think.
• "Count orders on weekends versus weekdays." This shows whether you lose sales when nobody is answering customers.
• "Split revenue by channel and tell me which channel is cheapest per unit of turnover." It helps you decide where to add budget and where to pull back.
• "Find days with sharp drops and spikes." A spike is a promo worth repeating, a drop is a glitch worth fixing.

How do you avoid trusting invented figures?

Claude is strong at spotting patterns, but arithmetic over a big table still deserves a check, especially totals and percentages. Ask it to show the math: "show how you calculated the top-5 share and from which rows." Verify a couple of numbers against the source by hand. If a result clashes with what you know about your business, do not rush to trust the table, more likely there is a format error in the data, point it out and ask for a recount. We wrote more about why a neural network sometimes invents things and how to catch it in a separate article.

What do you do with the conclusions next?

A breakdown is useless if it stays in the chat. Ask at the end for a short plan: three changes for next month with concrete actions. Drop three dead products, push the channel that gives cheap turnover, coach the rep with the low average order value. If you run these breakdowns every month, do not resend the context into a fresh chat, set up a Claude Project: put the prompt template and a business description in once, then just upload each new export.

Where do you start right now?

Export last month's sales, attach the file to claude.ai and run it through the template above. Five minutes, and you have a breakdown that used to take half a day, or never got made at all. Then tune the questions to your niche.

If you want to learn to work with Claude step by step instead of by guesswork, look at our free materials: the walkthrough at /guides and the first free lesson at /try/b0-01-unit, where in one sitting you set up your assistant and try it in action.

AGINE Academy is an independent product, not affiliated with Anthropic. Claude is owned by Anthropic.