# Analyze Excel and CSV with AI: validate data before interpreting numbers

> Define the row grain, inspect missingness, duplicates, and units, then request an analysis that can be recalculated.

**Author:** Awamer  
**Published:** 2026-11-11  
**Category:** Practical guides

---

A polished analysis can be wrong because the model misunderstood what a row represents. Five sales rows might mean five orders or five line items belonging to three orders. Define the data and measurement before asking for a leading product or an explanation of declining revenue. Confidence comes from validated inputs and reproducible calculations.

## Prepare a small data dictionary

Record the source, reporting period, extraction date, row grain, column meanings, and units. Explain statuses such as cancelled, returned, or completed, and whether amounts include tax, shipping, or discounts. Do not leave those definitions to guesswork.

In a fictional store export, repeated order identifiers may reflect legitimate line items. Removing them as duplicates loses sales, while counting them as separate orders inflates order counts. Use the appropriate identifier and grain for each metric, and check a small example manually.

## Inspect quality before cleaning

Review missing values, duplication, dates, categories, and unexpected amounts. Zero is not necessarily missing, and negative values may represent returns. Do not automatically remove outliers. Record each cleaning decision and its effect on row counts and totals.

Before joining files, validate keys and relationship cardinality. Compare counts and amounts before and after the join and inspect unmatched records. A sample should not be presented as a complete-file result.

Decision question: [question]. Row, column, unit, and period definitions: [dictionary]. Sanitized data: [source]. Inspect quality and ambiguity first. Explain calculations, metric formulas, and exclusions before conclusions. Return reproducible numbers, distinguish description from causal explanation, and state coverage limits. Suggest one chart with labeled axes and units that addresses the question.

## Verify interpretation as well as arithmetic

Use the [data analysis builder](/en/ai-prompt-builder/data-analysis), then inspect formulas or executable steps. Retain code and the analyzed dataset version when relevant. A narrative answer alone is not proof of a correct calculation.

A decline does not establish its cause. Check period lengths, stock availability, price changes, and order definitions. Treat explanations unsupported by the available data as hypotheses.

Report conclusions with metric definitions, periods, and limitations. The [claim review template](/en/ai-prompts/evidence-claim-review) helps inspect overreach, while the [decision matrix](/en/ai-prompts/weighted-decision-matrix) clarifies tradeoffs between actions. Share only necessary business data in accordance with the organization’s policies.
