Hospitals are past the pilot phase. The hard part is starting now.

For most of the past decade, coverage of clinical artificial intelligence has followed a familiar arc: a promising study, a controlled pilot, and a quiet retreat when the model failed to generalise. That arc is beginning to bend.

Interviews with fourteen chief medical information officers across US and UK hospital systems, conducted over the past three months, suggest that a small but expanding cohort of institutions has moved past pilot fatigue. At Mass General Brigham, Sheba Medical Center, University Hospitals Birmingham and Kaiser Permanente Northern California, at least one clinical AI application is now used routinely by front-line clinicians, embedded in the electronic health record and governed by the same change-control processes as any other clinical decision support tool.

This does not mean the diffusion problem is solved. It means the problem has shifted.

The applications that have scaled share three characteristics. They target a well-defined clinical task with a measurable outcome — sepsis screening, radiology triage, ambient documentation. They sit inside an existing clinician workflow rather than beside it. And they have a named clinical owner accountable for monitoring performance drift.

The applications that have not scaled tend to fail on the third point. As one CMIO put it in an interview for this article, 'We can procure the model. We can validate the model. What we cannot easily do is assign a clinician forty percent of a full-time equivalent to watch the model for the rest of its useful life.'

Reimbursement remains the second constraint. Only a small number of clinical AI applications currently attract a dedicated CPT code in the United States, and NHS England has yet to publish a general framework for reimbursing AI-augmented care. Until reimbursement pathways mature, adoption will continue to concentrate at academic medical centres and integrated delivery networks that can absorb the cost as a system-level investment.

The third constraint is workforce. Clinical AI does not eliminate labour; it changes what labour is scarce. Ambient documentation tools, for example, reduce the time physicians spend on notes but increase the demand for informatics staff, coders, and revenue-cycle analysts who can adjudicate AI-generated content.

The next twelve to twenty-four months will separate the institutions that treat clinical AI as an operational discipline from those that continue to treat it as an innovation programme. On the evidence collected for this piece, the gap between the two is already widening.