Do you get the same answer?
You'll typically find:
→ 3 ways of asking "total collection this year". 3 different tables picked.
→ 2 ways of asking "top products". 2 different #1 products.
The problem is that it won't crash to warn you. Just inaccurate answers.
Why? Because the meaning isn't in the database. It's in the application.
The AI sees 500 tables. It doesn't see that "a sale" excludes three status codes, or that the app quietly filters by company.
A bigger model doesn't fix this. Grounding does.
You need to teach it the business definitions, and check every answer against the system's own reports.
How are you checking your AI's numbers, friends?