The useful loop is:
execution → observed failure or exception → evidence → human decision → approved change → next execution
The point is not to make AI “learn everything.” The point is to make the operating standard improve without surrendering authority to the model.