HomeBlog › AI Guides
AI Guides 8 min read 30 July 2026

Using AI for spreadsheet work without getting burned

AI assistants are genuinely useful for analytical work and confidently wrong on a regular basis. The verification routine matters more than the prompt.

AI assistants have become a normal part of analytical work, and for good reason. Explaining a formula you inherited, drafting documentation for a workbook nobody documented, converting a vague request into a written specification, suggesting an approach to a transformation — these are all things they do well and quickly.

They are also confidently wrong often enough that using them without a verification habit is a genuine professional risk. Both things are true at once, and the useful skill is knowing which situation you are in.

Where they are reliably good

Explaining existing work. Paste a 200-character nested formula and ask what it does. The explanation is usually accurate, because the logic is fully present in the text you supplied — the model is reading, not inventing.

Drafting documentation. Describe a workbook's structure and ask for a documentation tab. You will edit it, but starting from a draft beats starting from an empty sheet.

Turning vagueness into specification. "The regional managers want to see how their teams are doing" is not a brief. An assistant is good at asking the clarifying questions and producing a structured specification you can then confirm with the stakeholder.

Naming things and structuring. Suggesting consistent naming conventions, a sensible tab order, or a way to break a monolithic formula into readable steps.

Where they fail, and how

Fabricated functions. Assistants occasionally produce functions that do not exist in Excel or DAX, with entirely plausible names and syntax. You discover this when the formula errors, which is the good case.

Plausible but wrong logic. The dangerous case. The formula runs, returns numbers, and computes something subtly different from what you asked — the wrong filter context in DAX, an off-by-one in a date range, an inclusive boundary where you needed exclusive. Nothing errors. The number is simply wrong.

Outdated syntax. Answers drawn from older material: VLOOKUP where XLOOKUP is available, or a workaround for a limitation that no longer exists.

Confident invention about your data. Ask about a file the model cannot see, and it may describe it anyway. Assume nothing about your own data comes from anywhere except your own data.

The verification routine

Apply this to every generated calculation, without exception:

  1. Read it and explain it back. If you cannot say what each part does in your own words, do not use it. Not understanding a formula in your model is a bigger problem than not having it.
  2. Test on data where you know the answer. Ten rows you can total by hand. If it disagrees with your manual figure, the formula is wrong, not your arithmetic.
  3. Test the edges. Empty values, zero, negative numbers, the first and last day of the period, a duplicate key. Edge cases are where plausible-but-wrong logic reveals itself.
  4. Cross-check the total. Compute the same figure a completely different way — a pivot table, a filtered sum — and compare. Two independent methods agreeing is real evidence.
  5. Only then put it in the live model. Never paste an unverified calculation into a file someone else relies on.

The confidentiality question

Before pasting anything into a public assistant, ask whether you would email that content to an external supplier. Customer records, salary data, unpublished financials and anything under NDA generally fail that test. Most organisations now have a written policy; if yours does not, the safe default is to paste structure and formulas rather than real data — an assistant can help with a formula perfectly well when the values are anonymised.

Want the full sequence rather than a single article? Our courses cover these subjects with realistic datasets, exercise files and solution files — browse the catalogue or see pricing. Written by BREVORA LTD, company number 17370266.

More articles

Excel Tips

XLOOKUP vs VLOOKUP: what actually changes in your work

VLOOKUP has three structural weaknesses that cause most lookup errors. XLOOKUP removes all three — here is exactly how, and what to do when your Excel version does not have it.

Power BI Tutorials

Your first Power Query pipeline: from monthly clean-up to one click

The single highest-return change most Excel users can make. A worked example, the mistakes that make queries fragile, and how to build one that survives next month.

Business Intelligence

Why your two teams report different numbers (and how to fix it)

Sales says 4,120 active customers. Finance says 3,880. Both are right. The problem is not the data — it is that nobody wrote down what "active" means.