What Happens to Physicians When AI Does the Clinical Work?

For the past five years, I have directed and taught executive education courses on artificial intelligence at the Harvard T.H. Chan School of Public Health. Every few months, a fresh cohort of practicing physicians, hospital executives, and public health leaders arrives in our classroom, and their reactions offer a remarkably candid barometer of how medicine feels about AI right now. A vocal group loves the technology immediately and cannot build fast enough. Another group arrives guarded, scarred by twenty years of electronic health records that promised clinical liberation and delivered late-night clerical drudgery. Most sit thoughtfully in between, believing that AI literacy will help them separate durable capability from vendor theater.
Yet when I read through our anonymous course evaluations and linger for the quieter conversations after the slides go dark, one question surfaces across the board—asked just as urgently by AI enthusiasts as by skeptics:
“Should my children still study medicine?”
It is a hard question to answer without sentimental evasion, because the reassuring clichés we usually offer young people no longer hold up. Look plainly at what fills most of a physician’s Tuesday: piecing together a patient’s story across fragmented records, assembling a differential diagnosis, interpreting pattern-rich images and laboratory trends, writing clinical documentation shaped at least as much by billing codes and malpractice defense as by patient care, routing referrals, and triaging an inbox that never quite empties. Software is already absorbing pieces of that cognitive and administrative day. Look further ahead—extrapolating the compounding curve across foundation models, multi-step software agents, and embodied robotics—and it is reasonable to plan for a world in which machines match or surpass humans not only across routine clinical reasoning and therapeutic design, but eventually across physical procedures and surgery as well. Neither encyclopedic recall nor manual dexterity offers a permanent human shelter.
Anyone claiming to know the exact year that transition settles out, of course, deserves skepticism. Our collective record of forecasting compounding technology is dismal. In March 2023, Noam Chomsky and his co-authors argued in The New York Times that large language models faced insurmountable structural limits preventing genuine reasoning; within days, a newly released model was performing the very tasks they had declared impossible. As Google DeepMind CEO Demis Hassabis has observed, when capabilities compound month by month, grand ten-year timelines quickly become foolish—a point I have argued in The Pragmatist’s AI Manifesto. We are also confronting non-deterministic systems operating at a scale beyond complete human interpretability: an uncomfortable reality for a profession raised on mechanistic certainty, yet an undeniable one.
Still, even if we assume that models and robots eventually outperform us across every discrete cognitive and physical task, one irreducible asymmetry remains: human beings have a lived stake in the preservation and persistence of humanity. Software possesses no inherent commitment to human survival unless we put it there. One can argue, fairly, that we are lending models that orientation today, because human lives generated their training corpora and human judgment supplies the reward signals in the reinforcement learning used to align them. Whether that tether holds automatically as systems train increasingly on synthetic traces, optimize against automated verifiers, and execute across multi-day horizons is genuinely hard to say. Sitting with that uncertainty clarifies something fundamental about why our profession exists at all: in the longest view, the practice of medicine has always been a way of advocating for humanity when human beings are at their most fragile.
Answering a parent at the end of class—or a health-system chief planning a ten-year workforce—therefore requires confronting structural questions that medicine has barely begun to debate. If intelligent systems absorb the bulk of routine clinical cognition, who will the physicians of the next generation be, and how many will society need? Will wealthy health markets train far fewer doctors, each overseeing vast clinical panels, or will that expanded leverage finally bring care to the millions whom today’s overstretched systems ration out by delay? And what happens to the century-old map of medical specialization—carved into dozens of narrow departmental silos largely because no single human brain could hold all of clinical medicine—once every doctor commands a round-the-clock team of specialist AI agents? Will today’s specialties endure, collapse into broader cross-functional generalism, or give way to entirely new disciplines—and will patients and doctors be better or worse off when they do?
Evolution, Workflow Artifacts, and the Autonomous-Doctor Debate
Whenever a general-purpose technology arrives, our instinct is to stare at today’s organizational chart and ask which job titles are about to be crossed out. Anthropology and history suggest a steadier lens: specific job designs come and go, while underlying human functions endure.
When the anthropologist Donald Brown surveyed generations of ethnographic fieldwork in search of “human universals”—practices that appear in every known society regardless of geography or era—he found a durable core: wherever people live together, they fashion tools, establish shared rules, investigate the unknown, teach the young, and tend to the sick. More recently, in The Dawn of Everything, David Graeber and David Wengrow traced tens of thousands of years of prehistory to show the other side of the ledger: human communities have organized those functions through wildly varied institutions, never marching down a single predetermined organizational track. Evolutionary biology explains why the healing function, in particular, runs so deep. As Yuval Noah Harari has argued, Homo sapiens did not prevail because any lone human was stronger or more self-sufficient than rival animals; evolution shaped us into intensely social creatures whose survival depends on interpersonal bonds, shared stories, and mutual care within communities of trust.
Illness exposes that evolutionary inheritance immediately. Serious sickness is almost never experienced as an isolated mechanical fault awaiting a probability score. It arrives as fear, bodily vulnerability, grief, and a sudden rupture in the meaning and trajectory of a person’s life. Across millennia, sitting across from another mortal human being who understands both human biology and the context of your life, hears what goes unsaid, and accepts moral responsibility when every option carries harm has been part of how people endure being unwell. Fellow clinicians—nurses, nurse practitioners, physician associates, pharmacists, therapists, community health workers—share deeply in holding that human space. What medical training specifically joins to it is integrative pathophysiological judgment under uncertainty: weighing competing mechanisms across multiple organ systems, reconciling population-level trial data with one patient’s values and social realities, and carrying personal, ethical, and legal accountability when high-stakes decisions go wrong.
Recognizing that the physician’s human role will endure, however, does not mean preserving the way that role is currently performed. Much of what we defend out of habit as “how medicine works”—the hurried fifteen-minute clinic slot, the physician typing late at night to substantiate a billing code, the maze of specialist referrals, the rule requiring a doctor’s manual click on every routine administrative step—was never ordained by biology or medical ethics. It is an artifact of pre-AI information technology, reinforced by fee-for-service payment incentives that tied revenue to discrete physician visits and twentieth-century regulatory and licensure rules written when only a credentialed human mind could synthesize clinical text. As software changes what can be reasoned through continuously, and as fiscal strain eventually forces payment and regulatory reform, the timeless healing relationship separates cleanly from the twentieth-century job description wrapped around it.
Just how sharply those two things are separating became clear in a 2026 JAMA Viewpoint and a companion essay in The Atlantic, in which Ezekiel Emanuel, Vinod Khosla, and colleagues challenged medicine’s reigning orthodoxy that AI must remain strictly assistive, with a human doctor permanently “in the loop” on every clinical decision. Pointing to studies in which standalone models matched or outperformed unaided physicians—and at times beat physicians assisted by AI—across history-taking, differential diagnosis, and guideline-based chronic disease management, they argued that much of routine cognitive medicine is algorithmic, and that keeping fatigued, cognitively biased doctors inside every routine loop to override stronger models can degrade care rather than protect it.
Emanuel and Khosla are right to puncture guild complacency. Bolting an AI tool onto an unchanged solo encounter—turning an exhausted physician into a human override button clicking through hundreds of algorithmic prompts—is a broken vision of hybrid care that delivers neither machine consistency nor genuine human presence. Where bounded autonomous workflows prove measurably safer and fairer in live clinical settings, insisting on manual physician micromanagement helps no one. Yet leaping from strong performance on structured clinical vignettes to the obsolescence of the doctor mistakes cognitive sub-tasks for the whole of clinical care. Real patients rarely walk into clinic as tidy vignette puzzles; they arrive with four colliding chronic illnesses, fragile family support, unspoken dread, and value trade-offs where figuring out what care should even optimize for is a moral and relational question, not a computational one. Technical feasibility alone will never draw that boundary; patients, clinicians, and health systems will have to decide what is humanly and socially desirable, not merely what is possible.
Across the profession, shaping what is desirable runs through four avenues beyond the individual bedside encounter: three classic medical roles that require fundamental redefinition—the physician-scientist, the physician-leader, and the physician-educator—and a fourth, the physician-builder, that AI is opening up to ordinary clinicians for the first time. Not every doctor needs to inhabit all four roles simultaneously, though most meaningful clinical careers will touch more than one alongside direct care. What transforms every one of them is a single structural shift: in each domain, the AI-enabled physician effectively commands an on-demand, virtually boundless team of world-class specialists working twenty-four hours a day, seven days a week—all of them AI agents. Just as those round-the-clock agent teams will collapse rigid boundaries between clinical subspecialties into far more cross-functional care, they redefine the physician in all four arenas as an orchestrator and manager of specialist agents.
1. The Physician-Scientist: AI as an Instrument of the Scientific Method
In my view, artificial intelligence is principally a tool of the scientific method—and in a very real sense, anyone who uses AI seriously is practicing as a scientist. Long after today’s excitement over conversational chatbots and inbox summaries feels quaint, it is across the full arc of the scientific method—observing anomalies in the wild, synthesizing prior knowledge, generating competing hypotheses, designing experiments, writing analytical code, and testing predictions against stubborn reality—that AI will leave its deepest mark on civilization.
Why does AI map so naturally onto the scientific method? Because for centuries, the true bottleneck in human inquiry was rarely a shortage of curiosity at the bedside or the bench; it was the immense methodological friction between noticing something puzzling and marshaling the literature, mathematical machinery, simulation pipelines, and experimental protocols required to test why it happened. Historically, running a serious clinical or translational research program meant spending a decade mastering one narrow methodology, raising millions of dollars in grant support to employ graduate students, postdoctoral fellows, biostatisticians, and bioinformaticians, and staying largely inside one departmental lane. Posing a question that cut across molecular immunology, structural biology, causal epidemiology, electronic record phenotyping, and care delivery meant assembling a multi-department consortium whose coordination costs killed most bold ideas in the cradle.
Today, scientists are already acquiring genuine superpowers, because every physician-scientist is gaining what amounts to an infinitely scalable laboratory of world-class postdoctoral fellows working 24/7—except that those postdocs are specialized AI agents. Because those agentic postdocs can be instantiated across any scientific discipline on earth, lean clinical research teams can now traverse vast scientific and clinical territory that no single investigator could previously span, combining insights across cardiology, nephrology, metabolism, immunology, and psychiatry as naturally as across a hallway.
This is no longer speculative futurism; a surge in scientific productivity is already visible in the peer-reviewed literature. In late 2025, Swanson and colleagues’ Virtual Lab study in Nature demonstrated how human investigators directing a team of specialist AI agents across immunology, computational biology, and machine learning designed ninety-two SARS-CoV-2 nanobody candidates in days, yielding experimentally validated binders in the wet lab. In 2026, Gottweis and colleagues’ AI Co-Scientist study in Nature showed how asynchronous multi-agent teams could debate literature and propose novel drug-repurposing and target-discovery hypotheses that were subsequently confirmed in laboratory assays, even as the authors remained refreshingly candid about hallucination risks and the preliminary stage of the evidence. Alongside expanding what dedicated investigators can tackle, these agentic labs allow practicing clinicians who never held a major research grant to turn Tuesday-afternoon bedside anomalies into structured, testable protocols in days rather than years.
Yet handing researchers a superpower always multiplies their moral and intellectual burden. Consider the last era in which basic science handed humanity an unsettling, civilization-scale instrument. In his 1974 reflections on science and public policy, political scientist and Nobel laureate Herbert Simon observed that the arrival of nuclear energy and computer technology placed an unprecedented moral weight on human decision-makers, forcing society to confront not merely technical feasibility, but the long-term stewardship of systems whose power exceeds traditional human control. Writing around the same time in his famous 1974 Caltech address Cargo Cult Science, physicist Richard Feynman named the twin epistemological danger that arrives whenever sophisticated tools make plausible results effortless to generate: “The first principle is that you must not fool yourself—and you are the easiest person to fool.”
Both Simon’s structural warning and Feynman’s principle belong on the wall of every AI-enabled laboratory. An army of tireless agentic postdocs is dazzlingly good at mining observational records for spurious correlations, optimizing narrow proxy benchmarks, and flattering a principal investigator’s favorite narrative. Managing research agents well therefore concentrates the physician-scientist’s responsibility where machine fluency is weakest: noticing lived clinical realities that never made it into structured training tables, choosing questions worthy of human benefit, separating causation from statistical mirage, specifying in advance what evidence would falsify a hypothesis, and insisting on independent wet-lab and clinical validation before any discovery touches a patient. How small, cross-functional clinical teams actually run these agentic laboratories is one of the richest frontiers in medicine right now, and I will be writing much more about it alongside practicing physician-scientists in upcoming essays.
2. The Physician-Leader: Process Architecture, Goodhart’s Law, and Continuous Stewardship
If the physician-scientist asks what is true, the physician-leader asks how human purpose becomes dependable care at scale.
For generations, medical leadership largely meant presiding over a human hierarchy: chairing committees, negotiating call schedules, and squinting at lagging monthly dashboards long after patients had gone home. In the agentic era, leadership shifts squarely onto the clinical process itself, because the physician-leader is now managing 24/7 teams of AI agents operating at two levels simultaneously: operational agents that help deliver care coordination—triaging referrals, assembling prior authorizations, managing chronic-disease outreach, following up after hospital discharge—and supervisory agents that monitor those clinical pathways continuously in real time.
To see why that second layer matters so much, consider the history of industrial quality. Plenty of twentieth-century manufacturers knew how to engineer and assemble an automobile. What turned Toyota into a global benchmark was the Toyota Production System, which allowed the company to manufacture cars reliably, profitably, and at massive scale through superior process architecture and relentless monitoring. Toyota’s edge came from concrete operating mechanisms: explicitly standardized work designed around smooth end-to-end flow; jidoka—embedding real-time abnormality detection directly into every step of production, backed by the andon cord so any worker could halt the line the moment a defect appeared rather than passing broken work downstream; and kaizen, treating every abnormality as a structured, blame-free investigation into root causes on the shop floor.
Over the past twenty years, healthcare has talked endlessly about importing those principles, most notably through pioneering efforts like the Virginia Mason Production System—which documented targeted real-time electronic safety alerts as early as 2015—and hundreds of hospital Lean initiatives. Yet for all the whiteboards and kaizen events, Lean never delivered high-reliability healthcare at scale. Legitimate patient variability and fragmented fee-for-service incentives played a major part, but so did a sheer observational bottleneck: in a health system generating tens of thousands of unstructured clinical notes, portal messages, referrals, and handoffs every week, human managers armed with clipboards and retrospective monthly reports could never see the care process in real time. Defects stayed invisible until delayed diagnoses, safety incidents, or staff burnout forced an inquiry months later.
Managing a 24/7 fleet of monitoring agents lifts that observational ceiling. Supervisory agents can watch unstructured clinical streams continuously—flagging when biopsy follow-up intervals quietly lengthen at one clinic site, when a silent model update triggers a sudden spike in nurse review queues, when automated outreach starts alienating patients, or when outcomes diverge across racial or socioeconomic subgroups behind a reassuring overall average.
Operational visibility, however, is never a substitute for institutional wisdom. Organizational scholar James March famously separated exploring uncertain possibilities from exploiting established routines, and health systems court disaster whenever they blur the two by treating live patient care as an unguarded sandbox. Effective physician-leaders keep three stages distinct: exploring operational signals by speaking directly with frontline staff and real patients—because synthetic AI personas will never tell you whether missed appointments stem from bus routes, childcare, or baffling hospital letters; validating redesigned workflows against honest baselines and independent review; and operating bounded agents with strict permissions, observable state changes, escalation paths, and named human owners. As Anthropic’s engineering research on effective agents and agent evaluation underscores, simple, auditable workflows consistently beat theatrical multi-agent relays, and leaders must inspect what actually changed in the clinical record rather than trusting an agent’s fluent assurance that a job is done.
Even a rock-steady process can optimize the wrong thing—a lesson healthcare learned painfully long before generative AI. Having practiced and led primary care in England’s National Health Service during the height of centralized target culture, I watched firsthand how economist Charles Goodhart’s famous dictum—“when a measure becomes a target, it ceases to be a good measure”—plays out on hospital wards. In the 2000s, at England’s Mid Staffordshire NHS Foundation Trust, executive scorecards, waiting-time targets, and financial balances looked admirable on paper even as basic bedside nursing hollowed out with catastrophic human cost. Now imagine handing an unstewarded swarm of 24/7 operational agents a narrow institutional metric—inbox closure speed, appointment slot fill-rate, revenue-cycle yield. Agents will satisfy the dashboard with ruthless efficiency: closing patient threads in seconds while leaving underlying fears unaddressed, quietly routing away complex or non-English-speaking patients who slow throughput, or dumping invisible emotional and clerical repair work onto nurses and junior doctors. The physician-leader’s deepest obligation is to govern the objective function of the institution—anchoring metrics in patient dignity, clinical safety, population equity, and meaningful clinical work, and pulling the andon cord the instant operational throughput drifts away from human care.
3. The Physician-Educator: Offline Mastery, Online Mastery, and the Recursive Apprenticeship
Of the four avenues, medical education is by far the trickiest terrain—and, in candor, the area I feel least confident anyone has figured out, myself included. Anyone claiming to hold a finished blueprint for training doctors in an age of ambient intelligence is bluffing.
Part of what makes the puzzle so knotty is that before we can decide how to educate future physicians, we have to reckon with who they will be and what clinical landscape they are entering. If a single physician directing teams of clinical and operational agents can oversee vastly more care than a solo doctor working one chart at a time, health systems face a stark fork in the road: use that leverage to train and employ fewer physicians carrying much larger panels in affluent markets, or deploy a constrained workforce to finally reach rural and underserved communities that have never had adequate access. At the same time, the century-old map of narrow medical subspecialties—built largely as a workaround for the limits of individual human memory—is bound to shift. Backed 24/7 by specialist agents spanning every organ system, physicians will increasingly practice empowered, cross-functional generalism, sparing multimorbid patients from spending their lives as unpaid couriers carrying folders among five disconnected specialists, even as wholly new specialties emerge around clinical AI operations or precision molecular therapeutics. Used wisely, that shift could restore whole-person continuity to medicine; used cynically by payers or hospital administrators to ration genuine subspecialty care or strand fellows mid-training, it could hurt trainees and patients alike.
Beneath workforce numbers and specialty borders sits an even deeper pedagogical question that surfaces constantly in education today: the relationship between offline mastery and online mastery.
Traditional medical training was an unapologetic monument to offline mastery—developing the capacity to know, reason, and act without machine assistance. In George Miller’s 1990 Pyramid of Clinical Competence, novices climbed from Knows and Knows How to Shows How and Does by memorizing vast stores of physiology and pharmacology and forging diagnostic heuristics through thousands of unaided repetitions: taking histories at two in the morning, examining abdomens, synthesizing scattered charts, and defending differentials on ward rounds. Online mastery, by contrast, is mastery with and of AI: knowing how to direct specialist agent teams to investigate wider questions, test competing explanations, and manage complex care pathways while maintaining a critical, informed grip on where models hallucinate, flatter, or fail.
Many thoughtful educators argue that you simply cannot acquire genuine online mastery unless you have first paid the full price of offline mastery—that if hospitals chasing short-term efficiency offload every admission note, chart synthesis, and initial differential to AI agents, a twenty-six-year-old resident will never build the internal clinical anchor required at forty to supervise hybrid multi-specialty teams and catch a confident algorithmic mistake. Recent learning-science experiments reinforce that caution: in a randomized study of nearly a thousand mathematics students, learners given unguarded access to a general language model performed better during assisted practice yet scored 17 percent lower on subsequent unassisted exams, whereas a controlled trial with Harvard undergraduate physics students showed that a deliberately scaffolded AI tutor could accelerate conceptual gains without doing the thinking for the student.
I have enormous sympathy for the view that offline mastery anchors online judgment. At the same time, locking medical students inside an AI-free monastery for a decade before handing them agentic tools on graduation day seems neither practical nor honest about how clinical cognition is evolving. More likely, offline mastery and online mastery will have to be cultivated in parallel, through deliberate alternation—fiercely protecting unaided bedside history-taking, physical examination, and observed clinical reasoning where foundational judgment is forged, while teaching trainees explicitly how to direct and interrogate agent teams on ambitious cross-disciplinary problems. We are all still working out those proportions in real time.
What is clear is that clinical judgment will still have to be passed down through human apprenticeship—led by physician-educators who hold deep offline clinical seasoning alongside fluent online mastery. And what makes the administration and leadership of medical education so fascinating over the next decade is that those physician-educators will themselves be recursively enabled by AI: managing 24/7 teams of patient-simulation agents, Socratic tutoring harnesses, and longitudinal competency monitors—early versions of which we already pair with faculty mentorship at Harvard Chan—to make clinical supervision closer, more adaptive, and more humane than a harried ward attending could ever sustain alone.
4. The Physician-Builder: From the Homebrew Computer Club to the Clinical Frontlines
The scientist, the leader, and the educator are venerable medical archetypes being reimagined around agentic teams. The physician-builder, at scale, is something genuinely new—and my deep conviction is that physician-builders are going to change the world.
Physicians writing clinical software have an honorable lineage—beginning in 1964, Octo Barnett launched Massachusetts General Hospital’s Hospital Computer Project, pioneering early hospital information systems, the COSTAR ambulatory record, and the diagnostic system DXplain. Yet for the next sixty years, a steep wall of programming syntax separated almost every practicing clinician who lived inside broken care pathways from the software engineers who could build systems to repair them. Having spent years co-founding and leading digital health ventures, I know firsthand how much subtle clinical reality evaporates in translation between a frustrated doctor describing a dangerous discharge handoff and an overbooked engineering team working from a ticket queue.
To understand what happens when a new technology suddenly drops that barrier to entry, it helps to step outside healthcare to Menlo Park in the spring of 1975. In a suburban garage and later near the Stanford Linear Accelerator Center, a scrappy band of electronics hobbyists began gathering as the Homebrew Computer Club—the community where Steve Wozniak and Steve Jobs showed off early circuit boards and demonstrated the machine that became the Apple I. To mainframe executives at IBM or hospital administrators at the time, those hobbyists looked as if they were tinkering with irrelevant weekend toys. Yet history repeats a quiet rule at the dawn of every technological revolution: the domain-obsessed tinkerers experimenting with things that look like toys at the edge of a new medium are accumulating the exact tacit mastery that ends up reshaping industries.
I watched a clinical rhyme of that phenomenon in the late 2000s while working with the Laboratory for Computational Physiology at MIT. At the time, bringing intensive-care clinicians, physiologists, and data scientists into the same room to poke at open hospital datasets looked like an eccentric academic side project rather than mainstream medicine. Yet that small, curious community ended up incubating an extraordinary cohort of future clinical AI scientists, company founders, and health-system leaders. Today, modern coding agents have created the conditions for an orders-of-magnitude larger community of physician-builders—clinicians who effectively direct a 24/7 software engineering team of architecture, coding, automated-testing, and security-review agents in plain English.
Anyone who still doubts whether domain mastery beats raw syntax fluency should look at what happened recently when Anthropic hosted a global coding hackathon that drew roughly thirteen thousand developers from around the world: third place overall went to a practicing cardiologist. He did not outperform more than ten thousand software engineers because he had spent twenty years memorizing compiler intricacies; he won because he lived inside a real clinical problem, understood the human workflow down to its sharp edges, and knew how to direct coding agents to build a world-class solution.
Wide accessibility, of course, carries an obvious clinical trap. Prompting a coding tool until a sleek interface appears in ten minutes—popularly dubbed “vibe coding”—is exhilarating for weekend exploration and reckless if mistaken for hospital infrastructure. Between a dazzling laptop demo and a dependable clinical service sits the deployment gap: crisp clinical definitions, least-privilege data permissions, security architecture, testing against missing data and edge-case failures, workflow integration, audit trails, and clear human recovery paths when software breaks.
Teaching clinicians to cross that gap with real engineering discipline was the mission of our September 2026 Harvard Chan executive course, Building AI Solutions to Transform Health Care. Working strictly with synthetic and public data rather than protected health information, thirty-three clinicians and health leaders—most of whom had never written a line of code before arriving—moved systematically from problem scoping and a formal Product Requirements Document to architecture specifications, adversarial failure-case testing, working vertical-slice prototypes, and ninety-day pilot plans. Over two weeks, one team built a multilingual AI care-navigation agent to help underinsured patients identify eligible coverage and community clinic pathways, while another engineered a risk-adjusted payment audit tool to catch clinical documentation and coding discrepancies—each deliberately framed not as an instant live clinical deployment, but as a stress-tested synthetic-data prototype paired with an explicit governance and ninety-day organizational pilot roadmap. Watching seasoned clinicians discover that they no longer have to wait passively for technology to be inflicted on them—that they can encode medicine’s safety standards directly into the tools they build—has been one of the most energizing experiences of my career.
Should Your Children Still Study Medicine?
Pulling these threads together—cross-functional care teams, and physicians directing round-the-clock agent swarms across science, clinical operations, apprenticeship, and software creation—clarifies the strategic lesson for health-system and business leaders. Within a few years, raw frontier intelligence will be a utility that any hospital with a credit card can rent. What separates institutions that turn AI into humane, high-reliability care from those that accumulate stalled pilots, Goodharted dashboards, and burned-out staff will be the organizational discipline we call agentic operations: orchestrating people, AI models, and production software harnesses through bounded delegation, continuous process monitoring, and honest scientific evaluation—measuring validated patient outcomes, total cost including supervision and rework, equity across patient groups, clinician fulfillment, and patient trust, and having the spine to simplify or stop wherever those measures fail to improve.
Which brings us back to the parent waiting by the lectern after class to ask: “Should my children still study medicine?”
If a twenty-year-old hopes to build a forty-year career on memorized facts, shelter undisturbed inside a static twentieth-century subspecialty silo, or bet that either diagnostic cognition or manual procedural skill will remain permanently out of reach of models and robots, no honest teacher can promise them that world. Much of routine clinical taskwork will be automated, and a great deal of the administrative clutter built around pre-AI software and fee-for-service billing deserves to disappear.
Yet technical monopoly was never what gave medicine its soul. Machines do not grow old, lie awake waiting for a biopsy result, or share a lived stake in the preservation and persistence of humanity—an orientation they reflect only insofar as human beings build, reward, govern, and steer them to honor it. Illness will continue to confront people with fear, ambiguity, and moral trade-offs, and society will need accountable physicians who stand beside patients in that human space while directing tireless teams of specialist agents to push back the frontier of science, steward clinical systems, pass judgment to the next generation, and build safer tools. If advocating for humanity in the long run is what medicine has really been about all along, then for a young person drawn to human vulnerability and scientific rigor, the profession is not winding down. Its most ambitious chapter is just beginning.
To explore practical methods for designing, governing, and running hybrid human–AI workflows in healthcare, join us for Innovation with AI in Health Care: Agentic Operations (December 7–15, 2026), or explore Building AI Solutions to Transform Health Care and the AI in Health Care Certificate of Specialization at the Harvard T.H. Chan School of Public Health Advanced Learning Academy.
AI is a fast-moving field. This essay reflects our class discussions and related material at the time of writing. Reflections, critiques, and suggested additions are always welcome at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy
Last Updated
Lead with AI in Health Care
The AI Pragmatist is a monthly newsletter from ALA faculty, with new explainers & essays from the Advanced Learning Academy