Notes · Note 06

Let AI read your spreadsheet

The most common small-business use there is — and the checks that keep the answers honest.

The short version

  • Two minutes of cleanup decides whether the whole exercise works.
  • Ask the question you would ask a smart bookkeeper, not a formula.
  • Spot-check one number by hand before you believe any of it.

Every business has a spreadsheet nobody wants to open. Sales by month, aging receivables, job costs, a year of card statements. Asking questions of it in plain English is the fastest first win we see, and it needs no setup beyond a paid seat.

Clean up first — two minutes, and it decides everything

A model reads a spreadsheet the way a new bookkeeper would: it takes what it sees literally. Before you upload:

  • One header row, at the top, with real names in it.
  • Delete the empty rows above the data and the notes off to the side.
  • One thing per column. No cells holding both a date and a comment.
  • Numbers formatted as numbers — not text, not with stray symbols.

Skip this and you will get confident answers computed off the wrong rows, which is worse than no answer at all.

Ask like you would ask a bookkeeper

Do not ask for a formula. Ask the actual business question.

Here's our 2026 invoice export. Which customers are over 60 days late, how much in total, and which three account for most of it? Show your working.

“Show your working” matters. You want to see which rows it used, because that is how you catch a misreading in ten seconds instead of at a board meeting.

Check it before you believe it

Pick one number from the answer and verify it by hand. One, every time, for the first few weeks.

This is not paranoia. It is the difference between a tool you can rely on and one that eventually embarrasses you. Once you have watched it match your own arithmetic ten times running, you will know which questions it is safe on.

Before you paste.

Client names, account numbers and card details do not belong in a chatbot on a consumer plan. Swap names for roles and drop the account numbers — the analysis comes out the same.

Three questions that earn their keep immediately

  • “Which customers are late, and who owes the most?”
  • “What did we spend on software last year, grouped by vendor?”
  • “Which jobs came in under the quoted hours, and which blew past?”

None of those needs a formula you would have to remember. All of them are questions somebody in your business is answering by hand right now.