Let it read your spreadsheet.

The most common thing a small business actually needs, and the one most people never try. No formulas required.

The thing to understand first

There are two completely different ways AI handles a spreadsheet, and knowing which one you're getting is the difference between a reliable answer and a confident wrong one.

It reads the numbersThe model looks at the values as text and reasons about them. Fine for "what stands out here", unreliable for arithmetic across thousands of rows.

It writes code that calculatesThe model writes a small program, runs it on your file, and reports the result. This is real arithmetic and it's the one you want for anything you'll act on.

ChatGPT, Claude, and Gemini can all do the second one when your file is attached. If a total matters, say so explicitly: "calculate this, don't estimate it."

Clean up takes two minutes and saves the whole exercise

You don't need a tidy database, but you do need these:

One header row at the very top. Not a logo, not a merged title bar, not three rows of notes.
One thing per column. "Smith, John — paid 3/4" should be three columns.
Delete the decorative rows — blank spacers, subtotals in the middle, the row of colour coding only you understand.

Export as CSV or XLSX and attach it. Screenshots of spreadsheets work far worse than the file.

Start with the question you'd ask a smart bookkeeper

Not a formula. A question.

Attached is our invoice history for the last 18 months. Columns are client, date sent, date paid, and amount. Calculate — don't estimate — the following: 1. Average days to payment, overall and per client 2. Which clients are consistently slower than average 3. How much is outstanding right now, past 30 days Show me the numbers in a small table. If any rows are malformed or ambiguous, list them separately instead of guessing.

That last sentence is what separates a useful answer from a fictional one. Bad rows exist in every real spreadsheet, and you want them surfaced rather than quietly averaged in.

Check it before you believe it

Do this every time, at least at first. It takes thirty seconds.

Ask for the row count. "How many rows did you use?" If it says 412 and your file has 1,300, it truncated and the answer is wrong.
Spot-check one number you can verify by hand. One is enough to catch a systematic error.
Ask it to show its work. "What did you exclude and why?" Real exclusions are fine. Unexplained ones aren't.

The failure that costs people money.

The dangerous output isn't a wrong answer — it's a wrong answer formatted beautifully with a confident summary paragraph. Presentation quality tells you nothing about arithmetic. Check the row count.

Three that earn their keep immediately

Who's about to leave. Give it 18 months of orders and ask which regulars have quietly stopped buying. Almost every business has a handful and nobody's watching.
What the messy column says. The free-text "notes" field nobody can analyse — ask it to group those into themes and count them. This is the single best use of AI on business data.
Explain the inherited file. Paste a formula you didn't write and ask what it does. Faster than reverse-engineering it.

Do this now

Export one real spreadsheet — sales, invoices, or customers.
Delete the decorative rows so there's a single header row.
Ask the three questions above, with "calculate, don't estimate."
Ask how many rows it used, and check one number by hand.

Before you paste

Customer spreadsheets usually contain names, emails, and sometimes more. Worth two minutes on what you should never paste in first — for most analysis you can delete the name column entirely and lose nothing.

Want someone to just do this with you?

The AI Once-Over maps where AI actually saves time or makes money in your business, then builds the first one with you.

See the AI Once-Over →