From Raw Table to Report: An AI Data Pipeline

By the end you can turn a raw spreadsheet into one page of conclusions you can present: what to hand to AI at each stage, what to ask, and where you must decide yourself.

1. Why "analyze this table" produces filler

Drop a spreadsheet in with "analyze this" and you get: an upward trend, some categories performed well, unusual fluctuations deserve attention. Every sentence is true and none survives a meeting.

The request has no constraints: the model does not know your business, which metric matters, or whose question you are answering. The ceiling on analysis quality is set by the question, not the model.

2. Five stages, one problem each

1. Raw table: work on a dated copy, never the original.
2. Cleaning: duplicates, gaps, formats, outliers. Skip it and every later number is wasted.
3. Aggregation: group by dimension. Definitions freeze here.
4. Charts: one chart, one question.
5. Conclusion: numbers become judgments and actions.

Most people jump from stage one to five, so the conclusion has nothing under it. Details are in cleaning data with AI and spreadsheet analysis with AI.

3. The three-part prompt

• Context: "A training company; course orders January to October 2026; the audience is the sales lead."
• Definitions: "Valid order means completed with zero refund. Revenue is cash received, tax excluded. Year-over-year compares to the same month last year."
• Questions: one to three. "Why did October revenue fall, which tier drove it, which accounts stopped repurchasing?"

Context is ... Fields are ... Definitions are ... The three questions are ... Before answering, tell me which fields you need and in what order, do not conclude yet. That clause catches a wrong plan early.

4. The conclusion formula

October revenue fell 14 percent month over month, driven mainly by 37 fewer enterprise orders in the East, so schedule renewal calls with the top 20 East accounts in November.

All three parts are required. A claim alone is an opinion, a number alone is a dump, an action alone is a guess. No specific number means you have not finished computing; no action means no business value.

5. Charts show shape, text delivers judgment

The chart lets someone grasp the shape in three seconds: rising or falling, concentrated or broken. The text says what the chart cannot: what it means, where it departs from expectation, what to do next.

A common waste is a chart labeled "East is highest" plus a paragraph repeating it. Write instead: "East is highest but growing slowest, so the base has plateaued and growth must come from new products." Never start a value axis at 80: it turns a 5 percent move into a collapse.

6. Report pages: the headline is the conclusion

Write each page title as a complete judgment, not a noun phrase. "October business performance" teaches nothing. "October revenue fell 14 percent, East enterprise churn is the main cause" delivers the point to anyone scanning titles.

Keep the body to three lines: the key number, the cause, the action. More than three and the page should split. Layout detail is in building decks with AI.

7. Three bad reports

• The number dump: eight metrics per page with no sentence connecting them. Readers remember zero.
• The undefined metric: "orders grew" with no definition of which orders count, so the conclusion flips when the definition changes.
• The missing recommendation: careful analysis ending in "keep monitoring this", which produces no action.

All three come from writing to show what you know instead of enabling a decision. The test: after the report went out, did anyone change one specific thing?

8. Checklist before sending

1. Can every number be traced to a cell in the summary table?
2. Are definitions stated up front or in a footnote?
3. Are comparison bases correct, with new regions and categories labeled separately?
4. Does every conclusion carry an action?

Weekly, monthly and retrospective reports need different skeletons, covered in writing reports with AI. Tooling keeps changing after November 2026, but three rules hold: people set definitions, numbers get checked, conclusions carry actions.