The right question about AI autonomy is not 'how much can we automate' but 'where does accountability live'. In operational systems, accountability is the feature that makes speed sustainable.
Well-designed agents prepare work rather than complete it. They gather evidence, evaluate alternatives, draft the recommendation, and then stop at the boundary where a business commitment would be made. A human with the authority to decide reviews the packet and approves, modifies, or rejects it.
This is not a limitation; it is what makes the system deployable. Legal, audit, and the people who carry the risk all need the same thing: a record showing what was recommended, on what evidence, and who signed off.
Design the approval path first and the agent second. The teams that skip this step end up with impressive tools that nobody is allowed to use.