What does a good prompt look like for a small-business owner?

Most small-business owners do not lack data. They have a point-of-sale report, a spreadsheet from the accountant and a pile of supplier invoices. What they lack is an hour to sit with those numbers and ask what they mean for next month. That is where ChatGPT can help, provided the prompt gives it the right material and asks the right question.

A Beginner's Guide to ChatGPT includes the executive among its example roles, and its refined prompts for that role share a pattern: name the data and the period, say how it should be broken down, ask for the trends that matter, then ask for recommendations. For an owner who is also the bookkeeper, the buyer and the person behind the counter, one more element matters: the actual decision you need to make.

The four parts of a business prompt

• Context. What the business is, where it operates, who buys from it. "A bakery in a small Québec town, busy with cottagers in summer" changes how the model reads a July spike.
• The data. Paste the figures as a small table, with the period and units. Remove customer names and anything personal; the model does not need them.
• The decision. What are you actually choosing? Staffing for the fall, what to stock, which product to push.
• The output and the check. Ask for a fixed number of options, the reasoning behind each, and every calculation written out so you can verify it.

That last part is not optional. OpenAI has warned since launch that ChatGPT can write answers that sound plausible but are wrong, and arithmetic on a pasted table is a classic place for a quiet error.

A worked example: three months of sales

Here is a prompt for a fictional bakery. The figures are illustrative.

You are a practical business advisor for very small retail businesses.

Context: I run a bakery in a small town in Québec. Summer brings cottagers and tourists; from October, most customers are year-round residents. I have two part-time employees.

Monthly sales (CAD, before tax):
| Product line | June | July | August |
| Bread | 6,200 | 5,900 | 5,600 |
| Pastries | 4,100 | 4,600 | 4,900 |
| Custom cakes | 2,800 | 3,900 | 4,400 |
| Coffee | 3,000 | 3,400 | 3,500 |

Decision: what to change for October to December.

Please:
1. Calculate the total for each month and the percentage change from June to August for each product line. Show every calculation.
2. Name the two trends that matter most for my decision, and one that probably reflects summer traffic only.
3. Suggest three concrete decisions for the fall, each with the reason, the cost or risk, and one number I should watch to know if it is working.
Keep it under 400 words. If something in my data is unclear, ask me before answering.

The structure does the work. The model has to show its math first, separate signal from seasonal noise, and tie each suggestion to a number you can watch. The line inviting it to ask questions matters too: OpenAI notes that ChatGPT usually guesses what an ambiguous request means instead of asking.

Check the math before you believe it

With these numbers, the totals should be 16,100, 17,800 and 18,400. Bread falls by about 9.7% from June to August, and custom cakes rise by about 57%. If the answer shows anything else, the model made an error, and every recommendation built on it needs a second look.

Three habits keep you safe:

1. Recompute the key figures in your own spreadsheet. It takes two minutes for a table this size.
2. Ask the model to check itself. "Recalculate the August total line by line" often catches a slip, though it is not proof.
3. Use a tool that runs real calculations when the data grows. ChatGPT Plus subscribers can turn on the code interpreter beta, which OpenAI began rolling out in July. It lets ChatGPT write and run code on a file you upload, so the totals come from a program rather than from the model's prediction of what the numbers should be. You still read the result with a critical eye.

How do you turn your data into decisions?

Start from the decision, not from the data. "What do my sales say?" invites a long summary of things you already know. "Should I keep the second employee on Saturdays this fall?" tells the model which numbers matter and what a useful answer looks like. Write the decision at the top of the prompt, then include only the figures that bear on it.

Next, ask for options rather than a single answer, and ask what would have to be true for each one to be right. If the cake recommendation only makes sense when cake orders hold up after the tourists leave, you now know exactly what to watch in October.

Finally, make each decision small enough to test. Change one thing, pick the number that will tell you whether it worked, and bring next month's figures back to the same conversation. Over a few cycles, ChatGPT becomes less of an oracle and more of a structured thinking partner, while the judgment about your customers, your staff and your town stays with you.