Most health systems hired data scientists for clinical AI. The modeling was never the bottleneck. Getting a model from a notebook to the bedside is. That journey runs through clinical context, local validation, EHR integration, and governance.
When clinical AI moved from research to the boardroom, most health systems hired the role the rest of the economy was hiring: the data scientist. AI is built by people who build models, so hire those people. Models got built and demonstrated well in validation, and very few changed anything a clinician did on a Tuesday morning.
Clinical AI isn’t primarily a modeling problem. A model is useful only once it’s validated on the system’s own patients, surfaced to the right clinician at the right moment inside the EHR, trusted enough to act on, and governed so it stays safe as the population drifts. Each is a distinct capability a general-purpose data scientist supplies none of. The modeling was the part that was never really the bottleneck.
One per gate on the journey above, most of which a conventional data-science org doesn’t contain.
What they doA clinician (physician, nurse, or pharmacist with informatics training) who can read a model, judge its clinical validity, understand the workflow it lives in, and write or evaluate code. The translator between data science and the bedside, and the single most decisive, most scarce hire.
Why not a data scientistThey’re clinicians with genuine technical depth, not data scientists with a healthcare interest. A combination almost never surfaced through a data-science recruiting funnel.
What they doValidates models on the local population before deployment and monitors them after for drift, degradation, and bias: the discipline some call algorithmovigilance. A model trained on another system’s patients can fail quietly on yours.
Why not a data scientistIt blends data science, epidemiology, and regulatory awareness into a function that keeps a deployed model honest, not one that builds it.
What they doGets model output into the workflow, inside Epic or Oracle Health via FHIR, SMART on FHIR, and CDS Hooks, so the insight reaches the clinician at the decision point without adding clicks or alert fatigue. AI outside the EHR goes unused.
Why not a data scientistA specialized integration skillset few generalist engineers have, routinely underestimated because leaders assume “the vendor will handle it.”
What they doOwns which models deploy, the safety case behind each, equity and bias review, and alignment with quality, compliance, and regulators’ transparency expectations. “Should we deploy this, on whom, and how will we know if it harms someone” is a patient-safety question.
Why not a data scientistIt blends clinical quality, ethics, and compliance: what separates a defensible program from an adverse event waiting to be written up.
Now the system is deploying capabilities it didn’t build and doesn’t fully understand, on its own patients, which raises the need for local validation, integration, and governance. The four roles live in a thin, contested market: clinician-informaticists are scarce by definition, EHR-integration engineers concentrated and expensive, validation and governance leads at an intersection of skills that rarely co-occur. Recruiting them means competing with academic medical centers, health-tech, and the EHR vendors themselves.
The kind of writing on workforce, AI, and enterprise hiring you'd actually want to read on a Sunday morning. No vendor pitches, no ad copy, no fluff.