Can I trust the numbers AI calculates?
Ask AI to use code instead.
You paste in a column of 60 expense amounts and ask for the total.
AI replies straight away with a tidy number, to the cent.
You check it in a spreadsheet and it's off by a few hundred. It doesn't look wrong, and that's the problem.
A chatbot writes answers by predicting likely next pieces of text (see "AI isn't looking anything up"). Doing that, it can produce a number that looks like a total without ever adding anything up.
AI also doesn't see numbers digit by digit. A long number gets chopped into chunks, so "1234567" might come in as "123", "456", "7".
Token: a chunk of text the model reads as one piece. Words, parts of words and groups of digits can each be a token.
Short, common sums usually come out fine. Long ones, and lots of numbers at once, are where it slips.
In August 2025, Microsoft added a COPILOT function to Excel that let you put AI directly into a cell.
Microsoft's own guidance said not to use it for tasks that need accuracy or reproducibility. That's a strong caveat for a spreadsheet feature.
A year later it was gone. Microsoft retired the function in September 2026, and cells that used it now show an error.
Most AI tools can now write a small program and actually run it. The program does the arithmetic, and computers don't get arithmetic wrong.
Code tool: a feature (called "analysis", "code interpreter" or similar) that lets the AI write a short program and run it, then read the result. It's the difference between someone guessing a total and someone typing it into a calculator.
If it used code, you'll usually see it: a code block, or a little "Analysing…" or "Running code" step, followed by the output.
If the answer just appears in the text with no code anywhere, it did the maths in its head.
When it's helping with a spreadsheet, ask for formulas, not values.
"Give me the formula for column F" gets you something the spreadsheet calculates and you can inspect. "What are the values for column F" gets you numbers it predicted.
A formula you can check once. Predicted values you'd have to check one by one.
Say it out loud: "Use code to calculate this."
Look for the code or the output. If there isn't any, assume it guessed.
Sanity-check the result. Does the total look about right? Is it bigger than the largest item, and less than the count times the largest?