Verify a scanned table after AI turns it into Excel
At a glance
Document testers used identical extraction instructions across three assistants, then checked the editable spreadsheets against the scans. This workflow is listed under Office Work, Document Processing, Spreadsheets.
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Have the original scanned document and an assistant session that accepts files and can produce an XLSX workbook. Keep the scan available for comparison. 2) Upload it and request editable cells rather than an image embedded in a worksheet. 3) Specify unchanged row and column relationships, grouped headers, blank cells, decimal places, negative signs, units, footnotes, and disclosure codes. Forbid guessing, rounding, and silent corrections; request locations for unreadable cells. 4) Ask for separate sheets where appropriate, correct continuation across page breaks, and a list of uncertain areas with the output filename and sheet count. 5) Download the workbook and compare its cells against the source before using it. The testers repeated this with the same documents and instructions in ChatGPT, Gemini, and Claude.
How it went
The tests found missing financial values, character substitutions, and decimal errors despite high extraction scores.
Worth knowing
Their scores measured selected cell accuracy, not formula correctness or overall workbook quality.
