Flatten the sheet before you ask: the layout is what it gets wrong, not the numbers

Overview

Why now. As of 25 August, the assistant built into your spreadsheet can write and run code against your workbook — ask it for an analysis and it produces code that executes, dropping the results straight into your file. That changes what a misunderstanding costs. A confused assistant used to hand you a wrong sentence about your data. It can now hand you a wrong sentence, a wrong chart and a wrong column, written into the sheet.

So one number is worth carrying. There is a benchmark built from 912 real questions people asked on Excel forums, each paired with the actual file they were wrestling with. On a fifty-question subset of it, four Excel experts scored 71.33%. When the benchmark’s authors ran Copilot in Excel by hand in 2024, on a separate sample, it scored 20.00%.

The reflex is to blame the model’s arithmetic, or your own phrasing. Often it is neither. What fails is that the assistant cannot reliably work out which cells you meant — and that lives in the shape of your sheet, not the wording of your question.

So, before you ask an assistant anything about a spreadsheet: copy the range you care about into a fresh flat table — one header row, no merged cells, no spacer rows — and write one line saying what each column holds.

The content

The reason that benchmark is hard is the reason your sheet is hard. Its files came from people asking for help with their real work, and the researchers describe them as containing “multiple tables, non-standard relational tables, and abundant non-textual elements”. That describes most workbooks in most organisations. Two tables stacked on one tab with a blank row between them. A merged cell spanning four columns that means this heading covers all of these. A second header row holding units. A note in the margin. A total sitting three rows below the data because someone left room.

Your eye resolves all of that in about a second. You do not read a merged cell as an empty cell beside a full one; you read it as a heading. The assistant gets a grid and must infer from position what you knew by looking. When it infers wrong, it does not stop. It picks cells, computes something defensible about them, and reports it in the same tone it uses when it is right.

The obvious objection is that this is a 2024 problem and 2026 models have moved past it. That has a dated answer. In June this year a successor benchmark took 321 tasks validated by domain experts, drawn from real financial reports and corporate filings — workbooks averaging 11.8 worksheets each — and ran eight frontier models under one setup that let each inspect the file step by step. The best overall score was 34.89%. On the debugging tasks it was 12.00%. And the authors’ own diagnosis is the sentence to keep: “insufficient spreadsheet inspection and incorrect target-cell selection are the dominant bottlenecks”.

Not calculation. Not fabrication. Looking, and pointing.

There is some direct evidence that structure is a cause rather than a symptom. A different team, back in April, built a system that reads a sheet and writes out a description of its layout before attempting anything. Strip that description out and accuracy drops — a small effect in their own testing, though they call the structural sketch a key part of the approach. A related benchmark on hierarchical tables points the same way: hand a model the identical content in different formats and the scores move. Same data, same question, different shape, different answer.

You cannot make your assistant sketch the structure internally. You can do the equivalent by hand, in about ninety seconds, and that is the whole habit. A flat copy is one header row, one table, no merges, no blank spacers, no stray notes — plus a sentence naming what the columns are. You are not simplifying for the model’s benefit. You are removing the guesses.

Two honest limits. First, nobody has run structure against phrasing head-to-head, so treat this as a repeatedly measured effect rather than a ranking — good phrasing still matters. Second, those setups do help: the best configuration in that April study reached 41.67% across the full benchmark. That is a real improvement, and still a long way from someone who can simply look at the file.

Try it

Take a sheet you have asked an assistant about, or would. Pick a question you already know the answer to — a specific total, or the value for one row and one period.

Run it twice.

1. Ask against the sheet as it stands. Then add: “Before you answer, list the exact cell references you used.”
2. Make a flat copy. New tab, paste only the range that matters, unmerge everything, delete blank separator rows and columns, keep one header row. Above it, write one line: Columns: A = region, B = month, C = units sold, D = revenue in AUD. Ask the identical question, with the same cell-reference request.

Compare the cell lists, not the answers. Whether it got the number right is not the finding — plenty of wrong reasoning lands on a right number. Whether it pointed at the cells you would have pointed at is.

Where this breaks. If your question genuinely spans the workbook — cross-sheet formulas, a lookup into another tab — a flat copy of one range throws away what the answer needs. And a flat copy of the wrong range is confidently wrong in a new way, with the ambiguity removed. Flattening buys you a clean question, not a correct one.

If file upload is switched off where you work, this still runs: paste the flat table in as text, and check what your workspace has enabled before assuming otherwise.

Additional reading

• SpreadsheetBench — NeurIPS 2024 Spotlight. The 912 real forum questions, and what the paired files contain.
• Towards Robust Real-World Spreadsheet Understanding — April 2026. Source of the 41.67% best-system score and the structural comparison; the 20.00% Copilot and 71.33% human figures are carried into its Table 1 from SpreadsheetBench.
• SpreadsheetBench 2 — June 2026. Eight frontier models on expert-validated tasks from real filings, and the failure taxonomy.
• RealHiTBench — ACL 2025 Findings. Merged headers, nested sub-tables, and the effect of changing format alone.
• Microsoft 365 Copilot release notes — the 25 August 2026 entry adding Python execution to editing in Excel.

Editor’s note

Everyone has indulged the temptation to adjust the prompt instead of the document when an AI assistant misreads something from a document. The prompt is your actual line of direct communication with the assistant, so it’s natural. But reality is that sometimes the document layout is what’s causing the misread, so it can be a more valuable use of your time to redirect your efforts there instead of trying to rework your prompt.