Which table did the AI read?
Check before calculation

Open the workbook you are already authorized to review. State the question the answer is meant to address and identify the version being considered. Ask where the relevant table begins and ends. Another region on the same sheet may have its own headings and purpose; proximity does not make it part of the same table. Keep the original arrangement available to the workbook owner. A blank row might separate two regions, while a heading above several columns might qualify each of them. Do not remove those features merely to make the assistant's representation look simpler.

Reading the structure and calculating from values are different tasks. If the assistant has selected the wrong region, checking its arithmetic cannot establish that it answered the intended question. Equally, finding the right table does not establish that every heading and value was interpreted correctly. This proposed comparison concerns what the answer represents. It does not investigate spending causes, compare suppliers' commercial terms or establish whether a search index contains every record. Keep the question small enough for the owner to inspect in the source workbook.

Compare the assistant's stated table with the original. Can the owner follow the answer to the sheet, relevant range and source cells? Read the row label together with the complete column heading, including any heading shared across a group of columns. A repeated label under a different parent heading need not describe the same thing. SpreadsheetLLM gives historical research support for separating these questions. Its revised method identifies a relevant table and its boundaries before generating a response. Its larger-table procedure also discusses retaining heading context when dividing a table. The authors report improved table detection against selected historical comparisons, not universal inability to read spreadsheets. The detection evaluation required exact matches for all four table boundaries and used annotations improved by human reviewers. That measure is not a test of every heading's meaning. A separate question-answering evaluation covered search, comparison and basic arithmetic, excluded composite operations, and assessed source cell addresses or formulas containing them. A fluent explanation alone was not its evaluation target. Some research compression deliberately replaces exact numerical detail with information about data formats to help identify structure. That is not a verified basis for calculation. Nor does the paper establish that a current commercial assistant uses its technique.

Current OpenAI guidance describes useful assistance, including summaries of rows and columns and tables for row-by-row review. It recommends descriptive first-row headings and one record per row, while advising against unrelated tables on one sheet and blank rows or columns splitting data. These are documented input recommendations, not measured accuracy rates. For an existing workbook, those recommendations do not authorize deleting meaningful separators or rewriting its headings. Ask the owner what the layout means. Keep a missing heading, uncertain boundary or unclear relationship visible in the comparison rather than silently filling it in. The guidance also says a successful upload can still leave analysis incomplete because of size, complexity, images, structure or formula complexity. Available file types vary with the model, plan, workspace settings and account capabilities. Exact values may not be reliably extracted from image-based tables, scans or complex visual layouts. A displayed table therefore does not establish complete or correct reading.

We propose a short reading aid under existing company rules: the factual question, workbook version, sheet and table range, complete relevant headings, source cell references, the assistant's representation and any unresolved difference. This is editorial advice, not a workplace procedure validated by the sources. Ask the workbook owner to compare those references with the authorized original. Does the answer combine separate regions? Has a shared heading disappeared? Does a cited cell belong to the heading the answer assigns it? Record what the owner actually resolves and who will pursue any remaining question. A reference can make an ambiguity easier to locate without resolving it. If the table or heading relationship remains uncertain, leave use of that answer pending. Asking the assistant to restate it more confidently would not supply the missing interpretation. The owner may clarify the intended region or request a narrower reading; do not describe an unexamined remainder as checked. This handoff grants no new access, changes no authoritative record and authorizes no purchase, payment or other financial decision. Even a resolved reading scope leaves calculation and the answer's later use to their ordinary responsible reviewers.