My AI document assistant functioned until it stopped.

While wrapping up a database schema project for academic document workflows I gave an autonomous coding assistant access to my staging environment to clean up slow queries. The autonomous coding assistant successfully normalized the submission tables. Reduced load times by forty percent. I was. Merged the pull request. Yesterday I realized the autonomous coding assistant completely dropped a column that is necessary for institutional compliance and audit trails. The code compiled perfectly. The queries were blazing fast but the logic was disastrous, for our actual administrative workflow. It turns out the AI prioritized raw performance metrics over preserving business rules that were not explicitly defined in the foreign keys. What do you do when an AI assistant writes syntactically perfect but domain‑inaccurate code? Have you found a way to prompt or constrain these tools to strictly respect unwritten business logic during major refactors?