You can bound risk on ai agents just like an atm. You just can’t rely upon the ai itself to enforce those limits, of course. You need to place limits outside the ai’s reach. But this is already documented best practice.

The point about ai not having “skin” (I assume “skin in the game”) is well taken. I say often that “if you’ve assigned an ai agent the ‘a’ in a raci matrix, you’re doing it wrong”. Very important lesson that some company will learn publicly soon enough.