Spreadsheets and AI: real help, one sharp edge
AI is excellent at writing formulas and explaining spreadsheets, and unreliable at doing arithmetic on your data. Where each mode is safe to use.
Spreadsheets are where many small businesses actually live, so it matters that AI helps with them in two very different ways. One is safe and excellent. The other looks identical and quietly produces wrong numbers. The whole trick is knowing which mode you are in.
The safe mode: AI writes, the spreadsheet computes
Ask a model to write a formula and you are using it for what it is: a text predictor that has read effectively every spreadsheet tutorial ever written. “Give me a formula that sums column C where column A says paid and the date in B is this month” produces a working SUMIFS faster than you would find the syntax yourself, with an explanation attached.
The same applies to lookups and joins, conditional formatting rules, pivot table setups, data cleaning steps, and untangling an inherited formula nobody understands. In this mode the spreadsheet still does the arithmetic. The model only wrote the instructions, you can read them, and the numbers come from your actual data through deterministic rules.
This mode also makes the model a patient tutor. Paste a formula you do not understand and ask what it does, step by step. That explanation is checkable against the formula itself, which makes it one of the lower-risk uses in the entire toolbox.
The sharp edge: AI as the calculator
Paste fifty rows into a chat and ask “what is the average margin here” and you have changed modes entirely. Now the model is predicting number-shaped text. It is much better at this than it used to be, and it remains a text predictor, not a calculator: totals can drift, a row can be silently skipped or duplicated, and the confident summary will not flag any of it. What an LLM actually is explains why fluency and correctness come apart; nowhere do they come apart more expensively than in arithmetic.
The rule that keeps you safe: numbers that matter are computed by the spreadsheet, never by the chat window. If the model analyzed data for you, the analysis is a draft of leads to verify, and every figure that survives into a decision gets recomputed in the sheet, per the verification habit.
Some AI tools blur this line helpfully: they write and execute real code against your data rather than predicting the answer text. Execution beats prediction. If your tool shows the code or formula it ran, you are back in the safe mode and the result inherits the spreadsheet’s reliability, not the chat’s.
A test before you trust
Before using AI on a spreadsheet that matters, run it once on a sheet where you already know the answers. Take last quarter’s numbers, ask the questions you would actually ask, and compare. Five minutes of this tells you more about your specific tool than any review, and it calibrates your trust to reality rather than to the demo.
While you are at it, check what happens to the data you paste. A spreadsheet of customer names and balances is pile-two material under the privacy baseline: anonymize it, use a workspace with training off, or keep the analysis local to your spreadsheet’s built-in AI where the vendor agreement covers it.
The honest division of labor
The model is the explainer, the formula writer, and the first-draft analyst. The spreadsheet is the calculator and the record. You are the one who decides which questions matter and whether the answer smells right against your sense of the business.
That last filter is worth keeping sharp. The owner who notices “that margin cannot be right, March was terrible” catches what no tool catches. AI extends your reach in a spreadsheet. It does not replace knowing your own numbers, and it works best for the people who would notice when something is off.