Controlled autonomy is definitely the right approach here. It’s not about blocking AI from taking action, but about building strict permission tiers like requiring human approval for anything financial or structural, while letting low-risk automated tasks run freely.