What does your review process catch before AI-generated UI reaches production?

AI can now generate a convincing interface faster than most teams can properly review it. The obvious problems are easy to catch: broken layout, inconsistent spacing, missing responsive states. The harder problems sit underneath the screen: the generated flow assumes a product decision nobody made empty, loading, failure and permission states are missing the same component behaves differently across pages an action has no clear system response * the UI looks complete but cannot be handed to development without explanation I have started reviewing AI-generated UI in three passes: Visual consistency - does it follow the existing component, spacing and motion rules? Product behaviour - are triggers, actions, system responses and next states clear? Failure handling - what happens outside the happy path? This catches much more than asking whether the generated screen looks good. For teams already shipping AI-generated UI, what is in your review checklist before it reaches production? And which failures still slip through most often?