Clinical AI attracts attention because the stakes are legible. But in the provider systems we work with, the largest verified financial impact is coming from operational use cases: theatre scheduling, discharge prediction, clinical coding, and supply chain.
The pattern is consistent. A model that predicts discharge date within a day improves bed flow, which reduces cancelled elective procedures, which improves both margin and access. The chain is unglamorous and the benefit is real.
What separates the systems that capture this from those that do not is data plumbing across departmental boundaries. Most hospitals have excellent data inside a department and poor data between departments. The operational use cases are precisely the ones that need the boundary crossings.
We advise a deliberate sequence: pick two boundary-crossing decisions, build the data products they need, and run them with a clinical owner and an operational owner jointly accountable. Do not attempt an enterprise-wide data platform first; it will consume the budget and produce a report.
The governance question is also sharper here than in most sectors. Patients are entitled to know when a model influences their care, and clinicians are entitled to a workflow in which disagreeing with the model is easy. Both are design requirements, not afterthoughts.
