When I served as chair of the Board of Registration for Physician Assistants in Massachusetts, every decision carried a weight that extended far beyond the immediate case in front of us. Whether we were reviewing a licensing violation, addressing unprofessional conduct, or weighing in on a practice standard, we were always, whether we named it or not, making decisions about what kind of profession we wanted PA practice to be ten or twenty years from now. The board didn’t write the regulations, but we lived inside their consequences. We saw what happened when the standards weren’t clear enough, when the boundaries of practice were ambiguous, when the profession hadn’t anticipated where medicine was heading. That experience taught me something I carry into every conversation about generative AI: The decisions that shape a profession’s future are almost never made in the future. They are made now, by people who cannot fully see what is coming.
Clinicians do a version of this every day at the bedside. We are trained not just to treat what is in front of us, but to think several steps ahead. Where is this patient likely to be medically in twenty-four hours? In the next week? What needs to be in place before they are discharged? We build care plans for futures that haven’t arrived yet. That is not prediction. That is clinical reasoning.
Health care is not applying either of these thought processes to generative AI. Instead, we are optimizing for right now.
Most of the conversation about generative AI in health care centers on efficiency. Does it reduce documentation time? Can it draft the note, triage the inbox, summarize the chart, or support prior authorization? These are legitimate questions. But in my opinion, they are not large enough for the moment we are in. If generative AI only helps us do yesterday’s work a little faster, we are thinking too small.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Health care has a long history of layering technology onto broken workflows and calling the result innovation. We digitized documentation without seriously rethinking why so much documentation was required. We opened patient portals without redesigning how asynchronous care would be staffed, escalated, or compensated. We asked new tools to revamp workflows that were misaligned long before the tools arrived. The current generative AI conversation in health care is following the same path. The tools are faster, but we are still asking them to fit inside a system that was built for a different era.
To me, the important question is not what generative AI can speed up. It is what generative AI makes newly possible. Health care’s biggest constraint is not lack of need. It is the inability to deliver continuity, navigation, personalization, and follow-through at scale. Patients have always needed more support between visits, more help understanding what to do next, more consistent follow-through on their care plans. In most cases, those services didn’t fail to exist because they lacked value. They fail to exist because the labor model can’t sustain them.
Think about what it would mean for a patient with poorly controlled diabetes to have personalized follow-through between visits, not generic reminders, but a thread of their actual care plan across gaps where most patients fall through. That capability exists now. Yet most health systems are not building it. They are using the same technology to draft prior authorization letters.
When the cost of the informational and administrative layer around care begins to fall, the question changes. It is no longer just whether we save time. It is what we can finally build.
That requires leaders to ask harder questions than clicks saved or minutes recovered. Which parts of our current workflow should disappear rather than speed up? Does this tool reduce cognitive load or simply relocate it? Does it improve continuity, follow-through, and escalation quality, or does it just help clinicians move faster inside the same broken architecture?
Some work in health care is ripe for subtraction, like duplicative documentation, manual rework born from poor workflow design, and inbox processes that exist only because no one has stopped to redesign asynchronous care. These tasks have survived not because they add value, but because they are familiar. Familiarity is not a reason to protect something. It is often a reason to question it.
And some work must never be optimized away, even if AI can perform parts of it faster. The patient conversations where trust is built, or the moment when a clinician recognizes that the patient’s story doesn’t fit the presentation. We still must keep the judgment required to know when not to follow the obvious recommendation or when to escalate symptoms due to our honed experiences. To me, that is the heart of clinical work, and we should not diminish it.
This distinction matters enormously for education. The next generation of clinicians will not just need to know how to use these tools. They will need the reasoning infrastructure to evaluate what those tools produce: to verify outputs, recognize when a plausible answer is wrong, and decide when the exception matters more than the pattern. If we are preparing clinicians for a generative AI-enabled practice environment without building the reasoning and judgment infrastructure first, we are producing something that looks like competence without requiring it.
For clinicians and health system leaders, the charge is the same one I learned on the licensing board: Make decisions today that are worthy of the world you are building, not just the one you are in. We will all, eventually, be patients in it.
Dipu Patel is a physician assistant and medical educator.
