AI is useful for data analysis, but the safest workflow is not “ask a question and copy the answer.” It is a loop: generate → inspect → test → refine.
1. Give AI enough context
State the business question, data grain, column meanings, constraints and expected output. Context reduces irrelevant or plausible-looking answers.
2. Make assumptions visible
Ask AI to list assumptions before producing a recommendation. Review which assumptions are actually true in your data.
3. Validate the mechanics
Check joins, filters, aggregation, row counts, null handling and date boundaries. A query that runs is not proof that the analysis is correct.
4. Validate the result
Reconcile important totals against a trusted source or independent calculation. Sample records when practical.
5. Challenge the interpretation
Separate observed facts from explanations. Ask what alternative explanation could produce the same pattern.
6. Document corrections
Keep the final logic, assumptions and important changes. The useful artifact is the validated analysis—not the AI transcript.
For a more detailed checklist, use the 10-point AI quality-control checklist.