So what will happen if the AI has a bug? Who is liable, you?
The AI can only execute functions that the user would be already be able to execute, and you'd definitely want to have an approval step where the user can see what's going to be executed before it happens!