Using AI to read your sales: the numbers are already there, what's missing is the right questions

Nine o'clock on a Sunday night. The last trolley in the stockroom is stacked with packed boxes waiting for the morning pickup. This month, this little online homeware shop sent out over six hundred parcels, more than ever before.

Then the owner opens the banking app, and what's left over is barely different from the month they sent four hundred.

She knows the answer is in there somewhere. In the sales export she downloads from the store's back end every month, in the marketplace reports, in the spreadsheet where she logs orders from LINE. Those files have been sitting in her downloads folder for six months, never opened for longer than two minutes.

The shop isn't short of numbers. What it's short of is someone with time to sit down and ask those numbers what happened.

Your shop already has the numbers; nobody has asked them anything

Most small shops think sales analysis is something big companies with data teams do. But most of what those teams actually do is ask questions of a table and chase down the answers. Small shops have the table. They just don't have anyone doing the chasing.

That's where AI earns its keep. Several AI chat apps, including ChatGPT, Gemini and Claude, can take an Excel or CSV file and read it for you. You don't need to write formulas or know how to make a chart. You just ask in plain language.

What AI doesn't know is what you want to find out. Ask it to "analyse my sales" and you'll get a long summary of things you already knew, like the fact that this month beat last month.

Prepare your sales file so AI can make sense of it

The file doesn't have to be pretty. It just needs one row per item sold and the basic columns: date, product name, quantity, selling price and sales channel. If you've got separate files from different channels, try to give the columns matching names before you combine them.

The column most often missing, and the most important, is cost per item. Most sales systems know the selling price but not what you paid for the stock, or what the bubble wrap cost. Without that column, AI can tell you what sells a lot but not what makes money. Half an hour adding rough costs for your main products is time well spent.

Remove customer details first. Exports from a store back end often include customer names, phone numbers and addresses. Working out what sells well needs none of that. Delete those columns before you upload. The reasons are in what to share with AI and what to keep out.

The questions that get the numbers talking

A good question is one whose answer lets you decide something. Knowing a bit more isn't enough. Something like this is a good start.

This file is three months of sales from an online homeware shop. Each row is one item in an order, with columns for date, product name, quantity, selling price, cost per item and sales channel.

Please tell me: the top 5 products by total profit (not by sales), which products sell in high volume but make little profit per item, and which pairs of products are often bought in the same order. Put it in a table and show the formula behind every number.

What sells well but doesn't make money. Do the maths for an imaginary shop. Ceramic mugs sold 300 units with 25 baht profit each after protective packing: 7,500 baht. Scented candles sold only 80, but at 120 baht profit each: 9,600 baht. The product the owner packed until her hands were sore all month earned less than one nobody talks about.

What gets bought together. If people who buy mugs often buy coasters too, the shop can sell them as a set, raising the value of each order while shipping costs stay the same.

Whether customers come back. If the file has a customer ID instead of a name, AI can count what share of customers bought more than once, and what they bought first. That tells you which products bring people into the shop.

Which channel actually leaves you money. Selling on a marketplace and selling on your own store site come with different fees. Add each channel's fees to the file and you can see how differently the same order pays out.