AI-Assisted SQL Exploration: Principles & Protocols

Generative AI is a powerful assistant for exploring schemas, drafting queries, and brainstorming analytical patterns. However, unvalidated AI output risks catastrophic data errors, fabricated column names, and incorrect aggregations.

1. The Responsible Human-in-the-Loop Protocol

Never execute AI-generated SQL blindly. Follow this 7-step engineering loop:

1. Understand Business Problem & Table Grain
                     ↓
2. Inspect Ground-Truth Schema Yourself (sys.tables / INFORMATION_SCHEMA)
                     ↓
3. Draft the Initial SQL Logic Independently
                     ↓
4. Prompt AI for Alternatives, Optimizations, or Edge Cases
                     ↓
5. Compare AI Suggestion against Ground-Truth Constraints
                     ↓
6. Inspect Execution Plan & Reconcile Aggregates
                     ↓
7. Integrate Verified Query into Version-Controlled Pipeline

2. Core Rules for AI-Assisted Querying

1. Provide Exact DDL or Metadata: Never ask AI "Write a query for customer sales" without feeding it the table names, key columns, and data types.
2. Specify Database Dialect: Explicitly instruct the AI to target Microsoft SQL Server (T-SQL) to avoid PostgreSQL or MySQL syntax (e.g. LIMIT vs TOP, DATEADD vs INTERVAL).
3. Specify the Output Grain: Instruct the AI what one row should represent in the final result set.
4. Demand Idempotence: Insist on CREATE OR ALTER VIEW syntax for maintainable objects.