AI can produce an answer that looks analytical, uses the right terminology, and is completely wrong for your dataset.

Before accepting AI-generated SQL, Python or metric logic, run these seven checks.

1. Check the grain

Ask what one row represents before and after the transformation. A correct query at the wrong grain is still wrong.

2. Check joins for multiplication

Compare row counts and distinct keys before and after joins. Unexpected multiplication is one of the easiest ways to create plausible nonsense.

3. Check filters

Confirm that date, status, geography and other filters match the business question. AI may invent a reasonable-looking condition that was never requested.

4. Reconcile totals

Compare important totals with a trusted source. If revenue, orders or customer counts suddenly change, stop and investigate.

5. Test edge cases

Try NULLs, zero values, duplicate records, boundary dates and records with unusual categories. Happy-path examples are not enough.

6. Separate assumptions from facts

Read the explanation and mark every assumption. If the business rule was not supplied, treat the model’s guess as a hypothesis, not a requirement.

7. Check the business meaning

Even technically correct code can answer the wrong question. Confirm that the final metric actually represents the decision someone needs to make.

A useful rule

Use AI to accelerate the work. Use your analytical process to decide whether the work is correct.

If you want the broader system behind these checks, my first ebook is now live.

Related: Data Cleaning Checklist Before You Start a Data Analysis.