Spreadsheets

QwenWorkSpreadsheets that still calculate

The failure mode of AI spreadsheets is a clean-looking file where the formulas have become numbers. This one keeps them alive — up to a size limit.

Formulas, not values

When it edits a workbook, existing formulas stay formulas. New summary columns are added as formulas referencing the original range, not as pasted results. That matters because the file will be edited again by someone who expects to drag a fill handle and see the numbers update.

Multi-file questions

The realistic use is not 「build me a dashboard」. It is 「these three exports disagree on the October total — find out why」. Point it at all three; it will match on keys, surface the rows that differ, and tell you which file has the extra entries. That is a two-hour task done in five minutes, and the output is a list you can verify.

Limits worth knowing before you start

DimensionPractical limitWhat happens past it
Rows per sheet~50,000 comfortablePreview lags; analysis still works but iteration gets slow
Files per taskA handfulMore files mean more disambiguation errors
Formula complexityStandard aggregationNested array formulas are preserved but not interpreted
MacrosNot executedxlsm is read as data; VBA is ignored
Never trust a number you did not spot-check. The model reads cells, it does not run the spreadsheet engine. When it says 「total is 1.28M」, that is a reading of the data, and a misread column can produce a confident wrong answer. Sort the source column and compare once.

Working habits that reduce rework

Analysing text-heavy exports instead? Documents handles long-form extraction better.