This is a verification playbook for AI-assisted analysis. It focuses on what to check after AI produces SQL, transformations, metrics, or interpretations.
Use AI for acceleration, not authority
Ask AI to draft SQL, explain errors, propose tests, or suggest edge cases. Treat the output as a draft until it survives validation.
1. Restate the business question
Before reviewing the generated answer, write down the exact question and population. Many analytical errors start with an ambiguous prompt.
2. Inspect the grain
Check whether joins, GROUP BY clauses, window functions, or transformations changed the row grain unexpectedly.
3. Test edge cases
Ask for checks around duplicates, nulls, dates, impossible values, unexpected categories, and empty populations.
4. Reconcile independently
Compare totals with a trusted source or calculate the metric through a different route. Independent checks are stronger than asking the same AI to verify itself.
5. Protect sensitive data
Follow organizational rules for customer, personal, financial, confidential, and proprietary information. Use masked or synthetic examples when possible.
6. Keep an audit trail
For important analysis, preserve the question, source, transformations, validation checks, and final interpretation.
The acceptance test
Ask: What assumption could make this wrong? Can I reproduce it? Did the query change the grain? Does the conclusion actually answer the business question?