Ask questions about a spreadsheet

This mini-book is for Level 1 User. You will ask an approved AI assistant three questions about a table, then prove each answer against the rows yourself. The examples use synthetic data and require no spreadsheet formulas.

2. Two thousand rows and one urgent question

A measurement export has two thousand rows. A manager asks for the total, unusual values, and missing entries before a meeting. You do not have time to learn pivot tables, so you upload the file and ask an AI assistant.

The response is immediate: a confident total, a neat outlier list, and a chart. There is only one problem. The assistant analysed 1,846 rows. A blank line ended the imported range, several numbers were read as text, or an unnoticed filter excluded one group. Nothing in the polished answer announces the loss.

AI-assisted analysis can remove the formula-writing barrier, but it does not remove the need to inspect the calculation. A useful answer must expose the imported row count, the method, the rows included, and the rows excluded. You then recount those rows independently. That turns “the assistant says” into a result you can defend.

3. After this you can

• Ask plain-language questions about an uploaded table without writing a formula.
• Require the method and exact row identifiers alongside every answer.
• Verify totals, threshold outliers, and missing values against the underlying rows.
• Recognise dropped rows, mixed data types, and unrequested filters.

4. Prerequisites

• T01-L02 · Trust but verify.
• An AI assistant approved for file uploads and structured-data analysis, or the illustrative outputs in this book.
• A spreadsheet viewer and a calculator or paper for an independent recount.
• A table with one header row, one stable identifier per row, and a known expected row count.

Uploading a file sends its contents into the selected service's processing environment. Storage, retention, model-improvement use, connected accounts, and administrator access depend on the service, account, settings, contract, and local policy. Use only public, synthetic, course-provided, or explicitly approved data. Do not upload personal data, patient data, employee records, customer details, unpublished research, credentials, or production exports. Apply T12-L01 · What you may paste before the upload, not after it.

5. The idea in one page

Inspect the import before the answer

An assistant cannot correctly analyse rows it did not import. Record the expected number of data rows before upload. After upload, ask for the imported row count, column names, first and last identifiers, duplicate identifiers, missing identifiers if the sequence is meaningful, and detected type of each column. If the source has 2,000 data rows and the tool reports 1,846, stop. Do not interpret the partial result.

Headers matter too. A title or note above the real header can cause columns such as amount_eur to become unnamed or mislabelled. Ask the assistant to repeat the headers it used. Correct the source or explicitly select the table before continuing.

Define the question precisely

“Find outliers” is incomplete. An outlier may mean a value over a fixed review threshold, a statistical rule, or simply something that looks surprising. At Level 1, provide the rule: Return every numeric absorbance greater than 1.00. Likewise, define whether a total excludes blanks and whether 0 is a valid value. A blank, zero, and the text N/A are not automatically the same thing.

Ask for an evidence trail

For every answer, require four things:

1. The result.
2. The method, including how blanks and non-numeric cells were handled.
3. The identifiers and values of rows included in the calculation.
4. The identifiers and reason for every excluded row, plus any active filter.

This request does not force the assistant to be correct. It makes silent choices visible. A total of 5.78 is only a claim until you can see which values produced it.

Recount independently

Do not verify an AI total by asking the same AI to “check again.” That repeats the same path. Use the displayed source rows and a calculator, or select the disclosed numeric cells in your spreadsheet and compare its count and sum. For missing values, inspect the named cells and recount them. For threshold outliers, compare each reported value with the threshold and scan the remaining values for another match.

For a large table, first reconcile the complete included and excluded row counts with the expected row count. Then manually inspect a small, preselected subset from the beginning, middle, and end to detect shifted columns or type problems. Finally, reproduce the full arithmetic with an independent spreadsheet status calculation or approved calculator. The assistant's explanation helps you choose what to check; it is not the check itself.

Watch for three silent failures

• Rows dropped at import: malformed lines, blank ranges, sheet selection, or file limits leave the assistant with fewer rows than the source.
• Mixed types treated as text: values such as 12.5, 12,5, N/A, and 12 mg share a column, so some entries may be excluded from arithmetic.
• A filter you did not ask for: only visible rows, one sheet, one date range, or one category contributes to the answer.

6. The worked example: three answers with row evidence

Use the same analysis request for both synthetic framings. Attach one fixture, replace the bracketed fields, and ask:

Inspect this table before analysing it.

Import audit
- Report the number of data rows and repeat the column headers.
- Report the first and last row IDs and any duplicate or missing IDs.
- State the detected type of each column.
- State every filter or excluded row. Do not apply a filter unless I request it.
- Stop if the imported row count is not [12].

Questions
1. What is the sum of numeric values in [value column], excluding blanks?
2. Which numeric values in [value column] are greater than [threshold]?
3. How many blank cells are in [value column]?

For each answer, show the method, the row IDs and values used, and the row IDs
excluded with reasons. Do not silently convert text to numbers. Do not create a
chart until the row checks pass.

Lab framing: a measurement export

Save this synthetic fixture as CSV or place it in a new spreadsheet. It contains 12 data rows; the header is not a data row.

reading_id,batch,absorbance
L-001,A,0.42
L-002,A,0.47
L-003,A,
L-004,A,0.51
L-005,A,1.80
L-006,A,0.45
L-007,B,0.39
L-008,B,0.41
L-009,B,0.44
L-010,B,0.46
L-011,B,
L-012,B,0.43

Set [value column] to absorbance and [threshold] to 1.00. An acceptable assistant response may use different wording, but its evidence must support these three results:

Mira verifies rather than accepting the response. She groups the numeric values into two calculator sums: batch A is 0.42 + 0.47 + 0.51 + 1.80 + 0.45 = 3.65; batch B is 0.39 + 0.41 + 0.44 + 0.46 + 0.43 = 2.13; and 3.65 + 2.13 = 5.78. She counts ten included numeric rows plus two excluded blank rows, which reconciles to all 12 source rows. She checks every numeric value against > 1.00 and finds only L-005. This arithmetic exercise does not establish scientific meaning for the sum or classify a biological observation.

Company framing: an expense export

Use the parallel synthetic Company fixture:

expense_id,team,amount_eur
E-001,Support,120.00
E-002,Sales,85.50
E-003,Support,99.50
E-004,Operations,
E-005,Sales,410.00
E-006,Operations,75.00
E-007,Support,130.00
E-008,Sales,95.00
E-009,Operations,110.00
E-010,Support,
E-011,Sales,90.00
E-012,Operations,85.00

Set [value column] to amount_eur and [threshold] to 300.00. The three supported answers are a sum of 1300.00, one threshold outlier (E-005 = 410.00), and two blank cells (E-004, E-010).

Jonas independently adds the ten numeric values in two groups: 120.00 + 85.50 + 99.50 + 410.00 + 75.00 = 790.00 and 130.00 + 95.00 + 110.00 + 90.00 + 85.00 = 510.00. The combined total is 1300.00. Ten included rows plus two excluded blanks again equals 12. He compares every numeric amount with > 300.00; only 410.00 passes. He does not call the expense fraudulent or wrong. The threshold means “review this row,” not “make a decision about it.”

In either framing, fail the run if the assistant reports fewer than 12 imported rows, includes a blank as zero without instruction, omits row IDs, or mentions a filter you did not request. Fix the import or question and restart the checks; do not repair only the final number.

7. What goes wrong

A total is accepted without a recount

Symptom: the answer looks precise, so it moves into a report with no row evidence.

Fix: require included and excluded IDs, independently reproduce the arithmetic, and reconcile their counts with the expected source rows.

Headers are not on the first row

Symptom: the assistant invents column names, treats a title as the header, or analyses the wrong range.

Fix: keep one header row directly above the data and ask the assistant to repeat the headers before calculating.

Decimal separators are mixed

Symptom: values such as 12.5 and 12,5 are not treated alike, and the numeric count is too small.

Fix: stop when the detected type or numeric count is wrong. Preserve the raw file, make an explicit approved cleanup, and document it; T10-L02 · Clean a messy export covers repeatable cleanup.

Personal or confidential data is uploaded

Symptom: the analysis works, but the file has crossed an unapproved data boundary.

Fix: decide whether the data and service are approved before upload. Use a synthetic fixture when learning and follow the incident route if disclosure already occurred.

A convincing chart uses only a subset

Symptom: the chart shape looks plausible, but its source range excludes blanks, a category, hidden rows, or part of the file.

Fix: postpone charts until row count, types, filters, and three numerical answers pass. Then require the chart's exact source rows and recount them too.

8. Do it yourself: a 30-minute checked analysis

Create one verification record and keep every step below inside it.

Minutes 0-5: choose the Lab or Company fixture. Record the filename, expected 12 data rows, expected headers, value column, blank handling, and fixed outlier threshold. Confirm that the file is synthetic and the selected tool is approved.

Minutes 5-10: upload the fixture and run the import-audit part of the prompt. Record imported row count, headers, first and last IDs, duplicates or missing IDs, detected column types, and filters. Stop if 12 source rows are not accounted for.

Minutes 10-17: ask the three questions. For each answer, paste the assistant's result, method, included row IDs and values, and excluded row IDs with reasons into the same record.

Minutes 17-25: manually verify all three answers. Re-add the ten numeric values with a calculator in two groups. Recount the included and blank rows. Compare every numeric value with the stated threshold. Write your independent result beside the assistant's result; do not ask the assistant to verify itself.

Minutes 25-30: mark each answer Pass or Fail. It passes only when the result, method, row evidence, and independent recount agree. If one fails, retain the observed mismatch, correct the import or question, rerun it, and record the new check in the same document.

Use this structure:

9. Exit check

Deliver exactly one artifact: one verification record containing the import audit and three answers, with each answer paired with the assistant's method, included and excluded row IDs, your manual recount of the underlying rows, and a Pass or Fail decision.

It passes when all expected source rows are accounted for, no unrequested filter is active, the value column has the expected type, and your independent sum, outlier check, and missing-value count reproduce the three reported answers. A screenshot or chart alone does not pass.

10. Rule to remember

Ask it which rows it used.

11. Further reading & tools

• Taught: T01-L02 · Trust but verify - applies evidence checks to AI answers before reuse.
• Taught: T12-L01 · What you may paste - decides whether a table may enter a selected service.
• Taught: T10-L02 · Clean a messy export - turns headers, types, missing values, and cleanup into a repeatable workflow.
• Catalogued: OpenAI data analysis (opens in a new tab) - primary guidance on analysing uploaded structured data and inspecting the analysis.
• Catalogued: Anthropic file uploads (opens in a new tab) - primary guidance on supported uploaded files and file handling.
• Catalogued: Gemini in Google Sheets (opens in a new tab) - primary guidance for asking Gemini to work with spreadsheet content in Google Sheets.
• Catalogued: Tools index - current product references; verify present capabilities, account limits, and organisational approval before use.