AI should be fast at suggesting what to do.
It should be slow and deliberate when changing production state.
That difference is the job of approval gates.
For database-backed AI workflows, I’d separate the path like this:
- read-only inspection is the default
- risky actions become proposals first
- proposals include affected rows and evidence
- dry runs produce counts and before/after samples
- approval is scoped to one tenant/action/window
- mutation tools stay unavailable until approval exists
- every step lands in the audit trail
The approval should not be “the agent can now edit the database.”
It should be:
This specific operation, against this scoped row set, based on this dry run, approved by this person, at this time.
Longer version: Approval gates for AI database actions
The point is not to make AI slow.
The point is to make suggestions fast and production changes deliberate.
Top comments (0)