What matters to me in data analysis is understanding how we get from a table to a conclusion. If AI says revenue fell, I want to see the calculation. If it offers an explanation, I want to know whether that explanation is actually in the data.

Four rows are enough to make this concrete. Below is fictional sales data for two products. The question is simple: what changed in August?

Four rows to start with

Each product has a unit count and revenue in euros excluding VAT. One row represents one product in one month. Copy this into a file named example-data.csv:

month,product,units,revenue_eur
2026-07,A,10,100
2026-07,B,5,150
2026-08,A,12,108
2026-08,B,4,120

In a real table, I would first check that every row means the same thing. Does revenue include the same items throughout? Has a missing value been turned into zero? Is a suspected duplicate actually a second transaction? I would make changes in a copy, so the original remains available.

A question to give AI

If the tool can read a file and run calculations, give it the CSV. Otherwise, ask for formulas or code and run them yourself.

Compare July and August revenue in this CSV.
Check data types, missing values and possible duplicates.
Do not change the source file.

Calculate total revenue for both months,
the difference in euros and as a percentage of July,
each product's revenue change and revenue per unit.

Show the formulas or code used and the intermediate results.
Separate observations in the data from assumptions.
If the baseline is zero, do not calculate a percentage change.

ChatGPT can run Python code during data analysis. OpenAI's documentation describes reviewing the calculations. I would keep that check even when the summary sounds convincing.

What actually changed here?

July totals 100 + 150 = EUR 250. August totals 108 + 120 = EUR 228. The difference is −EUR 22, or −8.8% compared with July.

Product A grows by EUR 8; product B falls by EUR 30. Together, they account for the same −EUR 22. That is a straightforward check that the product breakdown matches the overall total.

Then it gets more interesting. Product A sells more units: 12 rather than 10. But revenue per unit is EUR 9, down from EUR 10. Product B's revenue per unit remains EUR 30, while its unit count falls.

I can now say where the overall difference comes from. I cannot yet say why it happened. Product A might have had a discount, a price change or a return. Product B might have received fewer orders. This file cannot establish that. It cannot tell us profit either, because costs are missing.

My next question would concern product B's orders and product A's actual prices. This is where the analysis becomes useful to me: we know what to look for next.

If the same report is needed every month, we could automate the calculation. First, I would want to be sure it calculates exactly what we intended.