Work against a read-only or test copy first

Even a careful review misses things, so the environment gives you a second layer of safety. Where you run an AI-assisted query decides how bad a mistake can get.

Set up a safer target before you need it:

• Use a read-only connection for exploration. If the account cannot write, no generated statement can damage data, no matter what it contains. This is ideal for the drafting and explaining work where AI shines.
• Keep a test or staging copy for anything that writes. A migration or a bulk update should prove itself here first, against data shaped like production, before it goes anywhere near the real thing.
• Guard production behind human approval. Destructive changes run there only after a person has reviewed them and a recent backup exists.

Volume is the reason a test copy matters more than it seems. A query that returns instantly on a thousand test rows can lock a table for minutes on fifty million. Running it first on a realistic copy surfaces the performance surprise while it is still cheap.

This also frees you to move fast. When the connection is read-only or the database is a disposable copy, you can let AI draft freely, run its suggestions, and learn from what breaks, because nothing you care about is at risk. Save the careful, backed-up, reviewed path for the moment you touch real data.