How to check AI-generated Excel formulas before you rely on them

AI makes spreadsheets faster, not more trustworthy

AI assistants inside Excel can now write formulas, build tables and reshape whole workbooks from a sentence. That is genuinely useful for finance teams. It also changes the risk. A wrong formula used to take someone an afternoon to write. Now it takes seconds, looks professional and arrives with a confident explanation.

Microsoft itself advises users to review and verify AI-generated content. For a close workbook, that review needs to be a routine, not a glance.

1. Read what changed

Before accepting anything, list what the assistant touched: which cells, sheets, named ranges, queries or macros. Look for changes you did not ask for, such as a hard-coded number replacing a formula, a range that stops one row short, or a filter left on.

2. Make sure you can explain the formula

If you cannot explain in a sentence what a formula does, you cannot review it. Ask the assistant to rewrite long formulas with LET and named steps, or split them across helper columns. Readability is a control.

3. Test on a case with a known answer

Run the new logic on a small set of rows where you already know the right result, including awkward cases: a blank, a negative amount, a duplicate, a date at month-end. Most AI formula errors show up at the edges, not on the typical row.

4. Reproduce the key totals a second way

For anything material, reproduce the result by an independent method, such as a pivot table, a separate SUMIFS, a Power Query group or the source system's own report. If two methods agree, you have evidence. If you only have the AI's version, you have an assertion.

5. Keep the checks running

The workbook's own checks, such as row counts, duplicates and agreement to the ledger, should run on every refresh no matter who or what wrote the formulas. See the seven close checks.

6. Log it

Record AI-assisted changes in the change log like any other change: date, what changed, that an assistant was used, who reviewed it and who approved it. When a reviewer later asks how a number was produced, the answer should not be "the AI did it".

Where AI helps most

AI is strongest as a drafting tool: suggesting a formula, writing a first version of a Power Query step, explaining an inherited workbook. It is weakest as the final word on whether two transactions match or whether a balance is right. Keep matching rules explicit and deterministic, and keep the judgment with a person.