Diagnose the sheet before asking for insight

Back to: AI Data Analysis Lab

Inspect a spreadsheet’s grain, headers, types, gaps, and hidden assumptions before any AI-generated calculation.

Reviewed September 2026. Copilot, Python, and analysis features vary by license, account, platform, and update channel. Check the linked official documentation in the environment you use.

The situation

A sales sheet can look tidy and still be analytically broken. One row may represent an order, another a customer subtotal, and a final row the month total. Dates may be text in one region’s format and true dates in another. If you ask for “trends,” a tool may faithfully calculate nonsense.

This lab uses a fictional workshop-registration table. The decision is whether to repeat the event and which acquisition channel deserves attention.

Do the work

1. Make a working copy. Record the file owner, extraction date, source system, and whether the sheet is complete for the stated period.
2. Identify the grain: what exactly does one row represent? Mark subtotal, note, separator, and repeated-header rows that violate it.
3. Inspect headers and types. Each column should have one meaning; each cell should contain one value; dates, currency, percentages, and categories should use consistent formats.
4. Profile missing, duplicate, and impossible values. Separate zero from blank and “not applicable.” Look for dates outside the period, negative quantities, and category spelling variants.
5. Check hidden rows, filters, formulas, merged cells, and extra sheets. These can silently change what an upload or Copilot session sees.
6. Write a diagnosis before cleaning: five issues, their decision impact, and the proposed treatment. Get agreement on any rule that changes totals.

DATA INTAKE
File/owner:
Extracted:
Period covered:
Grain — one row is:
Row count before cleaning:
Key fields:
Missing means:
Known exclusions:
Hidden filters/rows checked:
Issue | decision impact | proposed treatment | approved by

Quality check

• The original is preserved.
• One-row meaning is explicit.
• Zero, blank, and not-applicable are distinct.
• Hidden filters and formulas were inspected.
• Cleaning rules are approved before totals change.

Keep current

Use these first-party references for current product behavior and established risk guidance. For consequential or regulated analysis, follow your organization’s approved methods and obtain qualified review.

• OpenAI: Data analysis with ChatGPT
• Microsoft: Copilot in Excel tips