The Nordic hospitals running clinical AI well all hired the same job. Almost no one budgeted for it.

Ask a Nordic hospital why one clinical AI deployment stuck and another quietly stopped being used, and the answer is rarely about the model.

In conversations with clinical informatics leads at university hospitals in Sweden, Norway and Finland, the consistent difference was a person. Titles vary: clinical AI steward, model owner, informatics lead for decision support. The role does not.

What the role does

It watches performance in production against a local baseline. It fields the clinician complaint that the alert is firing too often, and it has the authority to change the threshold or turn the thing off.

It maintains the relationship with the vendor, which in practice means noticing when the vendor has changed something and asking why.

And it protects the clinical staff from the model's failure modes, which is a different job from protecting the model from the clinical staff.

Why it keeps getting cut

The role does not appear in a business case. The business case describes a licence cost, an integration cost and a projected time saving. A permanent half-time clinician plus informatics support is ongoing operational expenditure attached to a project that has already been declared delivered.

The result is a familiar pattern. Deployment succeeds. Monitoring is assigned to whoever has capacity. Drift goes unnoticed for a quarter. Clinicians stop trusting the output. The tool is still licensed and no longer used.

The number that matters

Several of the systems interviewed now apply a rough internal rule: every clinical model in routine use carries between a fifth and a half of a full-time equivalent in ongoing human oversight, depending on how directly it touches a clinical decision.

That rule turns some deployments down at the business case stage. The informatics leads who use it consider that the point.