Two benchmarks, two eras
The first serious spreadsheet measurement appeared in 2024: 912 real user questions from Excel forums, with 2,729 test cases, on tables that are not tidy relational data but have colours, merged cells and several worksheets.
The best general model of the day scored 18.35%. The human Excel expert scored 71.33%. Fifty-three percentage points apart.
That benchmark is now largely "solved": the best reported result is above 70%. Which is why a harder one was needed.
The second edition, from June 2026, no longer measures individual operations but end-to-end business workflows: 321 tasks taken from real financial reports, averaging 11.8 worksheets and 593.5 modified cells per task. Eight frontier models and four commercial spreadsheet AI products were tested on it.
| Category | Best model's result |
|---|---|
| Visualisation | 62.5% |
| Template filling | 52.6% |
| Financial modelling | 34.0% |
| Debugging | 12.0% |
| Overall | 34.9% |
The overall figure is weighted across the 321 tasks, not the arithmetic mean of the four categories, which differ in size. The open-weight models finished between 7.2% and 17.1%.
The key sentence of the study tells you where the break is: today's models work with middling success at the level of an individual cell, but they do not hold the consistency between cells. They can rewrite a formula. A twenty-worksheet model where everything depends on everything else, they cannot.
The worst result is on the most common request
Let us stop at the 12% for a moment.
"Just have a look and tell me what is wrong with this spreadsheet" is probably the most common thing anyone asks a spreadsheet AI for. According to the measurement, this is the worst-performing category — worse than model building, far worse than chart making.
There is a logic to it. Debugging requires knowing what the correct result would be. The model, however, sees what is in the cell, and looks for an explanation for that. The same tendency that is useful elsewhere — making sense of the existing state — is exactly what cements the error in place here.
Where it cheats: a typology of the errors
The other important 2026 measurement ran on real, "dirty" data: 178 datasets, averaging 210 thousand rows per table (the largest 11.9 million), 36% of them multi-table, and more than half with external explanatory material attached. Here the best overall score stayed below 0.4.
But the number is less interesting than the structure of the errors:
| Error type | Share | What it means |
|---|---|---|
| Wrong filter condition | 32.4% | it fails to apply a condition the question does not state but which is obvious — for example it double-counts an already aggregated row |
| Transformation error | 23.2% | misreading a date format, badly converting a number stored as text |
| Context handling error | 8.7% | on a multi-step question it carries forward its own earlier wrong output |
| Visualisation error | 6.8% | wrong axis range, a timeline that does not cover the period asked for |
| Choosing the wrong table | 4.3% | |
| Misreading the external explanation | 3.9% |
This is the answer to "where it cheats". It does not manufacture invented rows. It quietly filters wrongly and aggregates wrongly — and the end result is a plausible number that is wrong. No error message, no warning, no suspicious formatting.
Four concrete examples from an independent test illustrate exactly this. Listing airports, it left out a city of seven hundred thousand; on another request, two major airports were missing from the list; asked to "sort by age" it actually sorted by date, and rows disappeared from the top of the list; in a list of countries it presented observer states as full members. All four errors produced a spreadsheet that looked flawless at first glance.
The formula that is going away — and what to learn from it
In August 2025 one of the large office suites introduced an AI function into its spreadsheet: you could type it into a cell like any other formula and give it a natural-language instruction.
The retirement date is 14 September 2026. After one year, and without ever leaving preview.
The vendor's reasoning is diplomatic, but its own documentation states the point: the underlying model can change, so the same formula, with the same arguments, may give a different result over time — and for any work where accuracy or reproducibility matters, they recommend the traditional formulas.
That sentence is the most important lesson of the article, and it reaches far beyond one product. A spreadsheet is a spreadsheet because it can be recalculated. If the value of a cell depends on which model version ran that day, then that cell is not a formula, it is a snapshot.
The practical rule that follows: use AI to build the spreadsheet, not to run it. Have it write the formula for you — and then let the formula stay in the spreadsheet, not the model.
The spreadsheet was faulty without AI too
There is an important counterpoint, without which all of the above would be unfair.
According to the classic summary of spreadsheet audits, the overwhelming majority of real spreadsheets in live use contain an error — 91% of those examined in the methodologically strictest audits. In a lab experiment where participants built a 25-to-50-cell spreadsheet from a written description, 51% of the work was faulty.
And the most important number in all of it: a person reviewing their own spreadsheet finds about half of the errors. That same summary notes that error detection close to 90% only occurs in the simplest processes, spell checking for instance.
So the question is not whether AI makes mistakes. It is whether your checking procedure notices them — because it does not notice a human's either.
The recipe: four checks
There is no AI-specific audit protocol, but the procedures of traditional spreadsheet checking carry over directly. Four steps, in this order.
1. ROW COUNT
How many rows in, how many out, how many dropped.
The three have to add up. If they do not, stop here.
2. CHECKSUM
The sum of the parts = the total you queried.
This catches the double-counted aggregate row — the
most common error type, a third of all cases.
3. BOUNDS
Any negative where there cannot be? A future date?
A division by zero? A percentage above one hundred?
4. SAMPLING
Take randomly chosen rows and recalculate them
by hand. Not the first one and not the last one.
For the fourth step it is worth knowing how much is enough. If you find not a single error in your sample, that still does not mean there is no error — only that the error rate is probably below a certain level:
| Rows checked | If you find no error at all, the error rate is below this with 95% confidence |
|---|---|
| 30 | 10% |
| 60 | 5% |
| 150 | 2% |
| 300 | 1% |
| 1,500 | 0.2% |
(The calculation is the "rule of three": if n trials produce zero errors, the upper bound on the error rate at 95% confidence is roughly 3/n.)
So checking twenty rows proves nothing about a ten-thousand-row spreadsheet. Three hundred already says something.
Common misconceptions
- "If it makes a mistake, it will show." A third of the measured errors are wrong filtering and a quarter are wrong conversion — both run without an error message and hand you a plausible number.
- "At least it writes the formula correctly." For an individual formula, often yes. The trouble starts with the consistency between cells: according to the measurement, that is exactly what it does not hold.
- "I will just ask it to find the error for me." That is its weakest measured capability: 12%.
- "That was just one bad model." In the harder measurement, eight frontier models and four commercial products competed. The best scored 34.9%.
- "I type it into the cell as a formula and I am done." That output is not reproducible: the model can change and the result with it. The vendor itself wrote this down before retiring the feature.
What to take away
Two sentences worth remembering.
The first: reproducibility is what decides. Whatever gives the same answer tomorrow can stay in the spreadsheet — a formula does, a model call does not.
The second: the four checks (row count, checksum, bounds, sampling) are needed not because AI is bad, but because the spreadsheet was faulty without AI as well. The machine only produces faster what always had to be checked — and it brings exactly as many more errors as there is more work being done.
Where there is any measurable gain at all in daily office work is set out task type by task type in the real time savings post; and for building the recurring, scheduled reports, the automating recurring reports post walks you through it.