Nearly every hand went up when attendees were asked who was exploring or piloting AI.
Then the question changed: Who was already scaling AI across teams, departments, or the broader organization?
The room looked very different. The ActiumHealth team's estimate was that roughly one in five hands remained up.
That gap shaped two days of discussion at the HIMSS AI in Healthcare Forum in San Diego. Health systems are not short on AI ideas. The harder work is deciding which use cases deserve to move forward, what evidence they need, and how to put governance around them before they expand.
Four themes stood out.
1. Broad interest has not yet become broad scale
The opening show-of-hands captured a pattern the ActiumHealth team heard throughout the Forum. Many healthcare organizations have AI pilots in isolated parts of the system. Far fewer have moved those projects across departments or into an enterprise operating model.
The organizations making progress described a deliberate sequence. They establish governance, define criteria for prioritizing use cases, decide whether to build or buy, and set expectations for evaluation and proof of concept.
That sequence gives a pilot somewhere to go. It also gives leaders a reason to stop work that is not producing useful evidence.
2. Fixed capacity is pushing operational use cases forward
Daniel Yang, MD, vice president of Artificial Intelligence and Emerging Technologies at Kaiser Permanente, opened with the pressure behind the AI conversation: healthcare demand continues to rise while the supply of highly trained people remains constrained.
The practical response is not to automate everything. It is to find work where added capacity can improve a defined patient or staff outcome.
Deepti Pandita, MD, chief medical information officer and vice president of clinical informatics at UCI Health, drew a useful distinction during an enterprise transformation panel. Many clinical AI applications still have a difficult path to hard ROI. Administrative and operational workflows can offer more measurable starting points.
Patient access fit that pattern. Scheduling services, collecting payments, routing requests, and helping patients complete routine work came up repeatedly in sessions and conversations. Each workflow has a recognizable request, an expected next step, and an outcome a team can inspect.
3. Governance has to travel with the technology
Dr. Yang's keynote centered on trust. Patients may trust a health system and may use AI on their own, yet still question how a healthcare organization applies AI to their care or information.
Governance is how a health system makes that use inspectable. Leaders need clear ownership, privacy and safety requirements, defined boundaries, transparent review, and a way to monitor performance over time.
The Forum discussions treated governance as operating infrastructure, not a final approval meeting. The same controls that determine whether a use case can launch also need to shape how teams evaluate changes, investigate failures, and decide whether the system is ready to scale.
4. A pilot needs both an exit path and an evidence plan
Speed came up in practical terms. Harun Rashid of Akron Children's described a six-month expectation for moving an initiative from start to go-live. The point was not speed for its own sake. A defined window forces a team to resolve ownership, workflow, integration, measurement, and deployment questions while there is still time to act on the answers.
The evidence bar is also rising. David Newman, MD, chief medical officer of virtual care at Sanford Health, pressed vendors to support claims with peer-reviewed evidence. Health systems also need their own operational evidence: what happened before deployment, what changed after launch, where the workflow stopped, and whether patients completed the job they came to do.
That evidence should be defined before a pilot begins. Otherwise, a successful demo can become a long-running experiment with no agreed path to expansion.
What this means for patient access
For patient-access teams, the Forum reinforced a practical place to begin: repeated patient need.
Call volume alone does not show which workflow should be automated first. Intent, resolution, non-resolution reasons, transfers, complaints, and voicemail patterns can reveal where demand and operational friction meet.
Those patterns may point to an AI agent. They may also point to training, routing, process change, staffing, or a policy that prevents the team from completing the request.
Start with one patient job. Define what completion and appropriate escalation look like. Name the systems and teams required. Then decide what evidence will prove the workflow is safe, useful, and ready to scale.
The question to carry into the next AI review is simple:
Which patient job should this use case complete, and what evidence will prove that it did?