A document table has a heading called Distance spanning two columns: kilometres and miles. An AI extracts every number but calls the output columns A and B. Nothing looks missing until you try to use the data. The values survived; their meaning did not.
Flatten the headings deliberately
Before asking for rows, ask the assistant to describe the header structure, units and footnotes. For this example, clear column names would be distance_km and distance_miles. Keep a source page reference with the result. These names are an example, not a fixed format every table should use.
Make a small mapping first: source heading, lower-level heading, output name. If a heading spans several columns, check its scope yourself. NIST describes confident but incorrect AI output as a known risk; a neat spreadsheet is not evidence that extraction worked.
Follow a row back to the original
Take the first data row and match each value to its original cell and full heading. Repeat near a page break and at the last row. A repeated header can be mistaken for a record, while a footnote can qualify only one column.
Keep blanks, zeroes and unreadable cells distinct. Mark uncertainty rather than inviting the assistant to invent a plausible value. Compare all rows needed for your task; a few spot checks cannot establish that the whole table is correct.
Preserve the source file. If you reshape an Excel copy, do not merge cells merely to reproduce the appearance: merging can discard values outside the retained cell. Check any totals separately, and keep the original units until you have verified the mapping.