Staffing prediction has been one of the easier commercial wins for healthcare AI: census forecasting is a well-behaved time series, and the savings are legible to a CFO. The pushback now emerging from nursing leadership is not about automation in principle. It is about what the models are actually optimizing.

Census forecasts at the four systems we reviewed are accurate to within a bed or two at the unit level. Acuity forecasts are substantially worse, because acuity is documented inconsistently and lags the patient's actual condition. A unit can be correctly staffed by headcount and badly staffed by workload.

"The model says the shift was covered," a charge nurse at a community hospital told us. "The shift was not covered. Both of those are true, and only one of them is in the report the executives read."

Several bargaining units have begun asking for contract language requiring that AI-generated schedules be reviewable by a clinician with override authority and that overrides be tracked without penalty. At least two systems have agreed. The broader question — whether a model trained on historical staffing can do anything other than reproduce historical understaffing — remains open.