Healthcare organizations have no shortage of AI pilots. The harder work begins when a promising capability must operate in a clinical environment where time, trust, safety, privacy, staffing, and accountability are all real constraints.
AHRQ’s Workflow Assessment for Health IT Toolkit treats workflow as the sequence of physical and mental tasks performed across people, teams, and organizations. That is the right starting point for clinical AI: a capability is useful only when it improves a real clinical moment without creating a new delay, handoff, or ambiguity about responsibility.
Start with the clinical moment
An AI program should begin with the moment it intends to improve: a clinician reviewing a longitudinal record, a care manager closing a gap, a pharmacist reconciling medication, or an operations team balancing capacity. Map the work before choosing the model.
For each moment, identify the information that must be present, the time available, the consequence of a false positive or false negative, and the person accountable for the final action. This prevents a common failure mode: optimizing a prediction while leaving the surrounding workflow unchanged.
Evidence has to travel with the recommendation
Trust is not a training session. It is a property of the interaction. A clinician needs to understand the source and freshness of information, the boundary of the model’s role, the uncertainty or exception signal, and the path to challenge or correct the output.
ONC’s current AI interoperability work adds a practical safety layer for AI-enabled systems in care delivery. It places data, workflow, provenance, and accountability in the same design conversation. That turns change management from a launch checklist into a repeatable clinical safety practice.
Change management is a patient-safety capability
When a clinical AI workflow changes, the change system should make the impact visible. Who needs to know? What training is sufficient? Which workarounds should be monitored? What happens when the model is unavailable? Which measure shows that adoption improved care rather than simply increasing clicks?
The answers belong in the same operating rhythm as incident management, service review, clinical governance, and quality improvement. This is where long-running ITIL change management becomes valuable: not as bureaucracy, but as a disciplined way to protect service while the system learns.
What leaders can do now
Pick one workflow with a consequential decision and build a six-week evidence loop. Document the baseline, the user promise, the evidence shown at the point of action, the human review requirement, the failure mode, and the measure that will decide whether to expand. Invite frontline staff to define the exception path before the success path is celebrated.
Global Enterprise brings healthcare strategy, data foundations, AI governance, and organizational change into a single design conversation. The aim is not to make AI impressive. It is to make better care easier to deliver and safer to improve.