How do you handle AI output that strays from your UI component guidelines?

We tried feeding our entire component setup into AI to speed up wireframing. It definitely output clean on brand code but everything started feeling a bit dry because it relied entirely on past decisions.

Sometimes a fresh user flow needs a completely new pattern rather than forcing old buttons and cards into a spot where they don't quite belong.

For product teams using design tokens with AI: do you force total adherence, or do you let the prompt stray outside your system to test new concepts?