Like most physicians, I have earned a great many continuing medical education credits in rooms where I learned very little. A slide deck, a speaker reading the slides aloud, a ten-question quiz at the end that anyone could pass by recognizing the longest answer. The credits counted. Whether my practice changed is another question.
I am an endocrinologist and a geneticist. Both fields move quickly. Genetic variant classifications get revised, new therapies reshape metabolic care, and guidelines are updated faster than any of us can reread them. Keeping up is not optional, yet the tools we use to keep up have barely changed in decades. We still mostly learn by being broadcast to.
I have come to believe that the most useful thing artificial intelligence can do for clinicians right now is not to answer our questions faster. It is to teach us better. And the version of AI that can do that is not a chatbot. It is what people in the field call an agentic system: software that can plan, take several steps toward a goal, check its own work, and adapt to the person in front of it.
Why the lecture model fails
Static education fails clinicians for reasons we all recognize.
It is one size for everyone. A first-year resident and a physician with twenty years in practice sit through the same hour, and both lose. The resident is lost; the veteran is bored.
Real physician voices, twice a week
Free, and one click to unsubscribe.
It tests recognition, not reasoning. Multiple-choice questions written for compliance tell us whether someone can spot a fact. They rarely tell us whether that person can work through an ambiguous case at 2 a.m. with incomplete information, which is the actual job.
It ends when the session ends. Learning science has long shown that we retain what we retrieve and practice over time, not what we hear once. A single lecture with no follow-up is close to the least effective format we could have designed.
It is disconnected from evidence as it changes. A recorded module is frozen on the day it was filmed. The literature is not.
What changes with agentic learning
Picture a different experience. Instead of watching a module on adrenal incidentalomas, a clinician works through a case with an AI tutor. The tutor presents a patient, asks what she would order next, and waits. If she orders the right test for the wrong reason, it asks her to explain her thinking and gently corrects the reasoning, not just the answer. If she is clearly already expert, it skips ahead to the edge cases. If she struggles, it slows down and returns to the concept she missed a week later, framed differently.
That is the core of agentic learning. The system adapts to the learner. It simulates cases rather than reciting them. It assesses reasoning, not recall. And it can cite the evidence behind every teaching point, so the learner can see exactly which guideline or study a claim rests on and judge it for herself.
For nurses, dentists, pharmacists, and scientists, the same idea applies. Education stops being a room everyone has to sit in and becomes a conversation tailored to what each person needs to learn next.
There is also a quieter benefit. Many clinicians are uncomfortable admitting what they do not know in front of colleagues. A patient, private tutor that never judges creates room to ask the basic question one has been too embarrassed to ask for years. I suspect that matters more than we admit.
The guardrails are not optional
I would be doing readers a disservice if I described this future without the risks, because they are real.
AI systems can be confidently wrong. A tutor that teaches an error with conviction is worse than no tutor at all. So the first guardrail is grounding: Every clinical claim should be tied to a cited source, and claims that cannot be supported should not be made. Citation should be checked at the level of the individual statement, not sprinkled on afterward.
The second guardrail is physician oversight. The curriculum, the cases, and the assessment standards should be authored or reviewed by clinicians who are accountable for them. AI can generate a hundred case variations; a physician should decide which ones are accurate and worth teaching. In my view, any program that awards credit should be able to name the humans who signed off on the content.
The third is transparency with learners. People should know when they are talking to an AI, what it was trained or grounded on, and where its limits are. An AI tutor should be willing to say “I don’t know” and point to a human expert.
The fourth is privacy. Case-based learning is most powerful when it feels real, but real patient data does not belong in a training tool without appropriate safeguards and consent. Synthetic or properly de-identified cases should be the default.
The fifth is equity of access. Many clinicians who would benefit most work where educational resources are scarce, often in languages other than English. If agentic education only reaches well-funded academic centers, we will have widened a gap we should be closing.
Finally, we need to measure whether it works. The honest answer today is that the field is early. Institutions adopting these tools should study them the way we would study any intervention: Did knowledge improve? Did it last? Did practice change? Enthusiasm is not evidence, and our profession should hold education technology to the same standard we hold a new drug.
What I hope clinicians do next
I work with colleagues building agentic learning tools for health professionals, so I am not a neutral observer. But that work has made me more convinced, not less, that the most important decisions here belong to clinicians and educators rather than to engineers.
If your institution, specialty society, or residency program is considering AI-driven education, get involved early. Ask who authored the content. Ask how claims are cited and checked. Ask how learners are assessed and how outcomes will be measured. Ask what happens when the AI is wrong.
Lifelong learning is part of the oath we take, even if the oath never mentions slide decks. We finally have tools that could make that learning personal, practical, and grounded in evidence. Our job is to make sure they are built with the care we would expect at the bedside.
Fady Hannah-Shmouni is a physician-scientist, AI founder, investor, and educator. He is a board-certified endocrinologist, internist, and biochemical geneticist with research funded by the National Institutes of Health, more than 100 peer-reviewed publications, and an h-index of 28. As an assistant clinical professor at the University of British Columbia, he supervises trainees at the intersection of clinical medicine and emerging science.
In 2025, he founded Hormonaly.ai, a clinical AI platform purpose-built for peptide therapy and longevity medicine. Hormonaly’s multi-agent, agentic clinical decision support system has opened its API to enterprise buyers, enabling health systems, digital health companies, and compounding pharmacies to integrate evidence-based AI into their workflows. The platform is live in Canada and the U.S., with bilingual (English and Arabic) infrastructure and active academic research partnerships in precision medicine.
He also founded DrVibe.ai and Healthspan Digital and serves as director of the GCLS AI Academy in Geneva. Through FHS Capital, he has invested in and advised more than 40 early-stage companies, including Eli Health, Science&Humans, Exerkine Biotech, Supernatural, and Neurable, at the intersection of health tech, AI, and biotech.
His mission is to build the infrastructure that empowers health care providers to become AI innovators and to ensure clinical intelligence reaches patients at scale. More about his work is on his website.
