Using AI to Write SQL: What's Real, What's Hype

 "Show me top customers by sales for last quarter."

AI writes the query, formats it, and even adds a comment.  Look at the time you saved!!!  Magic, right?  enh.... Let’s talk about what’s real and what’s hype when it comes to AI writing SQL.

The Promise

AI-Assisted query generation has exploded.  Between Copilot in SSMS, ChatGPT and the new Azure AI Integrations, we're seeing something incredible.  People who don't even speak SQL can now build queries that actually run!

Here’s what AI already does well:

• Generates boilerplate queries and joins in seconds.
• Converts English into working T-SQL (ie., 'show me top customers by sales').
• Suggests filters, aggregations, and window functions.

It’s faster, smarter, and, WHEN GUIDED CORRECTLY, can be remarkably accurate for common requests.

The Reality

We really mustn't forget the most important thing. The part that makes the SQL correct, efficient, usable, safe, even just applicable -- this part still belongs to you. 😉

AI doesn’t:

• Know your schema or naming conventions.
• Understand your business rules or data quality quirks.
• Predict query plan costs, indexes, or blocking risk.
• Catch logic traps like date overlaps, cardinality mismatches or cartesian joins.

Think of AI like a gifted intern.  Quick with code but you still need to review.

What’s Actually Working

Capability                    Tools                              Strong Use Cases

Natural language → SQL     ChatGPT, Azure Copilot, OpenAI         Quick analysis, ad-hoc queries

AI code completion                GitHub Copilot, Tabnine                  Auto-complete joins, templates

Query explanation                Plan Explorer AI, ChatGPT        Simplifies query plans, error logic

Schema-aware assistants      Databricks, Synapse                  Context-aware SQL generation

The Hype to Ignore (for Now)

“AI can write your entire ETL.”

It can mock it, but it cannot run it. Error handling, dependencies, and logging still need a DBA.

“AI tunes your queries automatically.”

Nope. Not yet. It can explain why something’s slow, but it cannot fix it safely in context.

“Natural language replaces SQL.”

NEVER. SQL remains the universal data language. AI just makes it less painful.

Make AI Work For You

Here’s a few tips for how to make AI provide real, production-grade value:

Feed your schema.

Input your table and column names into the prompts first, you'll see accuracy jump by 50-70%.

Ask for reasoning.                                                                                  

This is very clever because it forces AI to show its logic.

“Explain why you chose that join and filter.”

Refactor, don’t originate.

Use AI to format, document, and parameterize your query.

Compare against actual plans.

AI doesn’t know your indexes -- this is your job.

What’s Coming Next

Two research efforts to watch:

Execution-Guided SQL Generation — AI that runs partial queries to self-correct

SQL-Factory Frameworks — multi-agent AI systems that refine SQL collaboratively

These promise not just fluent SQL, but accurate SQL.  Not sure that I agree, but many speculate that Copilot may soon be more colleague than autocomplete.  

Bottom Line

AI won’t replace the DBA.  It will ease the long hours of writing SQL, endless joins, syntax errors, formatting, testing and debugging.

But what about the creative, diagnostic, performance-driven parts?  The part that understands the business problem and develops the solution?  This piece is still 100% human.  Still very much you.

AI may be able to write SQL faster than you, but AI doesn’t know your indexes.  It doesn't know your business rules, who is looking at your data, or any other conditions therein.  At least not yet, I should say.