Check Which Spreadsheet Rows AI Actually Analyzed
A correct-looking total can come from the wrong rows. Use record IDs, inclusion rules, and a two-table exercise to check an AI spreadsheet answer.
AI reports a paid-order total of 100. That is the expected number. Yet the answer may have included a canceled order and omitted a paid one of the same value. Checking only the total would miss both errors.
Ask for the identity of the rows used, not just a reassuring statement that the whole workbook was analyzed. Then compare that list with the task's inclusion rule. A workbook has sheets, ranges, filters, and formulas; “all the data” does not specify which of those should count.
Define the intended records
The following two invented tables stand in for two worksheet tabs. They are fully shown here so you can check the exercise without uploading a workbook. Amounts use one currency. The table labels and row positions are part of the source.
October orders
| Sheet row | Record ID | Status | Amount |
|---|---|---|---|
| 2 | O1 | paid | 40 |
| 3 | O2 | canceled | 25 |
| 4 | O3 | paid | 25 |
| 5 | O4 | paid | 35 |
September archive
| Sheet row | Record ID | Status | Amount |
|---|---|---|---|
| 2 | S1 | paid | 30 |
| 3 | S2 | paid | 20 |
The task is: total paid orders in October only. Include O1, O3, and O4: 40 + 25 + 35 = 100. Exclude O2 because it is canceled, and exclude S1 and S2 because they belong to September. The complete source has six records; the required calculation has three.
An incorrect selection of O1, O2, and O4 also totals 40 + 25 + 35 = 100. That deliberate counterexample shows why the number alone cannot establish the analyzed scope. It is an author-constructed mistake, not a captured product failure.
Request enough evidence to inspect the scope
Paste both tables and ask:
Calculate paid-order revenue for October only.
Before the total, identify the source table and the included record IDs,
with their sheet rows and amounts. Separately list excluded records
and the reason for each exclusion. State how many source records you read
and how many you included. Show the addition.
Do not treat the September archive as October data.
If you cannot access a required table or row, say so rather than estimating.
Check four things independently: the table is October orders; the included ID set is exactly O1, O3, O4; the count is three; and the amounts add to 100. An answer that says it read six records but shows only O1 and O4 has not reconciled the source with the calculation.
For an actual uploaded workbook, also request sheet names and the ranges or table names read. Keep the source unchanged while checking. A row number can move after sorting, so use stable record IDs together with row positions. Verify cited cells in the original workbook; the assistant can invent a convincing access claim just as it can invent a conclusion.
Filters change what you see, not the question by themselves
Suppose the October sheet currently displays only O1 and O4 because of a filter. The visible paid total is 75, while the total of all paid October records is 100. Either can answer a legitimate question, but the requester must choose which one.
Excel's SUBTOTAL documentation describes how filtered-out rows and manually hidden rows affect that function. These are Excel formula rules; they do not prove that an AI upload reader inherits the same filter behavior. A screenshot of visible rows, a file parser, and an assistant operating inside Excel may receive different material. Check the specific tool's observed input instead of transferring one access assumption to another.
When you want all paid October records, say “include records hidden by the current filter if their status is paid.” When you want the visible selection, describe the filter and ask it to report that selection. If the assistant cannot establish the filter state, resolve the scope before accepting a total.
Reconcile before sharing the report
Use a short reconciliation beside the answer: six source records, three included, three excluded, no unexplained records. For larger workbooks, have the assistant produce the inclusion list or a row-level audit table rather than squeezing every record into chat. Inspect boundary rows, unexpected statuses, repeated IDs, and a sample of both included and excluded records.
The worked numbers here were independently checked from the displayed tables. This is a scope-checking exercise, not a benchmark of Excel Copilot or an upload service. In your own tool, record what it could access and whether its evidence matches the original. A correct total becomes more useful when you can explain exactly which records produced it.
