How to Check AI's Answer Before You Share It

Confident is not the same as correct

If you take one thing from this entire course, take this lesson. Everything before it makes you faster. This one keeps you employed.

An AI chat tool does not calculate the way a spreadsheet calculates. A spreadsheet adds column D and gets the answer. A chat model produces the most plausible-looking next thing, and a plausible-looking total is very easy to produce and very hard to spot. It will add 300 numbers and give you a figure that is out by 4 percent, in the same calm tone it uses for everything else. There is no wobble in its voice, no asterisk, no hint that this one is a guess.

The good news is that four cheap habits catch almost all of it, and none of them takes more than two minutes.

Habit one: make the spreadsheet do the arithmetic

The best defence is to stop asking for arithmetic at all. Ask for the formula, put the formula in the sheet, and let Excel or Google Sheets do what it has been reliably doing since 1985.

Habit two: recompute a sample by hand

Whatever method the AI used, test it on a few rows you can verify yourself. Not the whole thing. Five rows, a calculator, two minutes.

Pick your five deliberately: the first row, the last row, the biggest value, one that should be excluded, and one awkward case such as a refund, a blank or a zero. If the method survives that set, it is probably sound. If the first row is right and the excluded row is included anyway, you have found a real bug at the cost of two minutes.

Before you give me any final number, show me the method:

Which rows you are including and which you are excluding, and why.
The exact rule you are applying to each row.
A worked example on the first three rows, step by step.
Then give me a spreadsheet formula that does this,
so I can run it myself and compare.

Habit three: check against a number you already trust

Every business has numbers that are already known to be right. Last month's reported revenue. The invoice total. The headcount. The figure the finance system prints. These are your control totals, and they are the fastest lie detector you own.

If your new analysis says May revenue was 412,880 and the finance system says 431,004, do not adjust the story until you can explain the gap. Usually there is a boring, findable reason: a filter that excluded refunds, a date range starting on the second of the month, or an entire region missing because its name was spelled differently in this export.

My analysis gives [412,880] for [May revenue].
The [finance system] reports [431,004]. The difference is [18,124].
Columns are [list them]. Here is my method: [describe it].
List the most likely causes of a gap this size, most likely first.
For each one, tell me exactly what to check in my sheet to confirm or rule it out.
Do not guess which cause it is. Give me the checks.

Habit four: count your rows at every step

This is the failure nobody sees coming, because nothing looks broken. You filter, you clean, you paste into a summary, and somewhere along the way 60 rows stop being included. The total goes down. The chart still draws. Every number is still a number.

Write the row count down before you start and after every step that filters, joins, removes duplicates or pastes. If it drops, find out why before you go any further. Blank cells in a lookup column, a filter still switched on from earlier, and rows below an accidental blank row are the three usual suspects.

I started with [900] rows and my summary now covers [847].
Columns are [list them]. The steps I ran were:
[filter, remove duplicates, lookup against another sheet, paste into summary].
Give me a formula or a menu step to list exactly which rows are missing.
Then list the most common reasons rows vanish at each of those steps.
Do not tell me the difference is probably fine.

The board pack rule

Here it is plainly: no number goes into a report, a board pack, a client email or a slide until you have either run it as a formula yourself or checked it against something you already trust.

Not because AI is useless at data. It is genuinely brilliant at data. But the cost of the error is not symmetrical. A formula you double-checked costs you two minutes. A wrong revenue figure in front of your board costs you the assumption that your numbers are reliable, and you do not get that back quickly.

If someone asks where a number came from and your honest answer is that the AI said so, that number is not ready to leave your desk.