The Primary Care Office After the EHR

A note on this piece. This is the flagship essay of a personal series about how the location of medical knowledge has changed in one physician’s lifetime. Several of its sections are written to grow into standalone pieces, and a map of those branches appears at the end.

Two rooms, thirty years apart. In the first, a young man sits at a wooden table in Amritsar in 1995 with a textbook of physiology open in front of him. Whatever he does not memorize tonight will not be available to him tomorrow. There is no device in his pocket, no search box, nothing he can ask. If he wants to know how the kidney handles potassium, the knowledge has to move from the page into his head, and it has to stay there.

In the second, a physician walks toward an examination room in California in 2026. On the screen behind him is a chart containing more information about the patient he is about to see than any human being could read in a week: thousands of laboratory values, hundreds of prescriptions, twenty years of notes written by people he has never met. Before he reaches the door, a piece of software has already read all of it, summarized what it thinks matters, drafted the orders it expects him to sign, and flagged a message the patient sent at 2 a.m.

The young man and the physician are the same person. That is the only unusual thing about this essay. Everything else in it is happening, or about to happen, to every doctor and every patient in America.

If you have visited a doctor’s office recently, you may already have seen the beginning of a transformation without realizing that you were witnessing one. Your physician may have asked permission to let a phone or computer listen to the conversation so that software could prepare the medical note. You may have received a MyChart message that was unusually polished, with a disclosure saying that artificial intelligence helped generate it. Your doctor may have walked into the room seeming to know more of your history than anyone could reasonably have read in the few minutes between patients. Somewhere behind the scenes, a machine may already have summarized part of your chart, examined a referral, suggested a billing code, prepared a reply, or decided which piece of information deserved a human being’s attention first.

These developments are easy to dismiss one at a time, because each looks small: a better note, a better search function, a smarter patient portal, a quicker refill. Taken together, I do not think they are small at all. I think they are the first visible pieces of the largest reorganization of outpatient medicine since the paper chart gave way to the electronic health record.

I have come to this question not as a technologist but because the arc of my own life happens to cross almost the entire transition. I began pre-medical studies in Punjab in 1994–1996 and studied medicine in Amritsar. I completed my intern year at the University of Arkansas for Medical Sciences in 2007–2008, while American electronic records were still fragmented and evolving. I then completed my second and third postgraduate years with UCSF Fresno at Community Regional Medical Center from 2008 through 2010. I began practicing in California just before Community Medical Centers made its major Epic transition in 2011.

I have practiced through the years in which Epic and other large electronic health record systems moved from useful software to something closer to the operating environment of American medicine. Now, thirty years after I first entered the sciences that led me here, I am watching that architecture begin to shift again.

This essay is therefore about artificial intelligence. Beneath the technology, though, it is about something much older: where knowledge lives, who controls it, how faithfully it is preserved as it passes from one set of hands to the next, and whether information ever actually becomes understanding.

For much of medical history, the physician carried an extraordinary amount of the system inside his or her own mind. The paper chart eventually became an external memory. The electronic record became an almost limitless one. Artificial intelligence is now beginning to do something different in kind rather than degree: it can read that memory, reorganize it, interpret it, and increasingly act upon it. That is where the promise begins. It is also where the danger begins.

Khadur Sahib: The Written Word, the Shared Meal, and the Living Body #

My family’s roots are in Khadur Sahib (ਖਡੂਰ ਸਾਹਿਬ), in Punjab’s Majha region, not far from Amritsar. To someone unfamiliar with Sikh history it is simply a town on a map. Within Sikh memory it is inseparable from Guru Angad Dev, the second Guru, who made Khadur an important center of the early Sikh community.

Guru Angad received the guruship from Guru Nanak in 1539. Khadur Sahib is therefore, among other things, a place of succession: the place where a founding teaching had to pass faithfully from its first voice into a second one, and from there into a community that would outlive them both.

I did not think of that as a medical idea when I was young. I do now. Every physician knows that the most dangerous moments in care are often the handoffs: the night sign-out, the hospital discharge, the referral letter, the transfer between clinics, the moment when one clinician’s understanding must become another’s. Information is not usually lost where it is created. It is lost where it is passed on.

The history associated with Guru Angad includes the strengthening and dissemination of Gurmukhi (ਗੁਰਮੁਖੀ), the script in which Punjabi is written. Khadur is also closely associated with Mata Khivi, Guru Angad’s wife, and with the tradition of langar (ਲੰਗਰ), the communal kitchen in which people sit together without distinction of wealth or social rank. Sikh tradition also remembers Guru Angad’s attention to disciplined physical life, including the akhara (ਅਖਾੜਾ), the ground for wrestling and bodily training.

As a child, I did not think of any of this as a framework for understanding medicine. Looking back from 2026, I find three ideas in that inheritance that seem unexpectedly suited to the technologies now entering health care.

The first is the faithful keeping of words. Gurmukhi gave Punjabi speech a durable written form and helped make teachings reproducible beyond the memory of any single person. That is a problem every civilization eventually meets: how do we preserve what happened without allowing it to change each time it is retold?

Medicine has wrestled with precisely that problem for centuries. A physician observes something, and then it is written. Another physician reads the writing. A later clinician copies the summary. Years afterward, the patient’s story may contain a mixture of observation, interpretation, and inherited assumption, and no one can easily tell which is which. The medium has changed from handwriting to databases. The obligation has not: keep the record faithfully.

The second is represented by langar. Mata Khivi is remembered for her stewardship of the communal kitchen at Khadur, where nourishment was extended to everyone rather than according to hierarchy. If you have ever visited a Sikh gurdwara in Fresno, Yuba City, Houston, or Queens and been handed a plate, you have eaten in that tradition.

Medicine uses different words for the same moral question: access, equity, disparities, social determinants of health. If technology makes health care more powerful, who receives that power? Does the affluent patient receive an invisible AI assistant that gives a physician more time to listen, while the poor patient receives an algorithm in place of a physician? Does automation level the floor, or does it build another tier?

The third idea is physical. Medicine can easily forget the body when data become seductive. The patient is not the chart, the risk score, or the prediction. There remains a body in front of us, and sometimes the most important information is visible only there: in a face, a gait, a tremor, an effort of breathing, a pause before answering, or the expression of the family member who has not yet spoken. Those three ideas give us a useful test for health care AI: fidelity, equality, and embodiment.

• Can it preserve the record without distorting it, including across every handoff?
• Can its benefits reach people without deepening the inequalities that already exist?
• Can it help us see the patient more clearly, rather than replacing the patient with data?

The verse at the head of this essay has stayed with me for that reason. Its point is not that light is useless. Its point is that illumination and understanding are not the same thing. A hundred moons and a thousand suns can produce enormous brightness and still fail to tell us what anything means. Modern medicine is now meeting that problem at a scale no previous generation of physicians faced.

Amritsar: When Knowledge Had Weight #

I grew up in Amritsar, a city of well over a million people that most Americans know, if at all, as the home of the Golden Temple. It is also the city founded by Guru Ram Das, the fourth Guru, whose name my medical school would later carry.

I attended Spring Dale School. Spring Dale’s own institutional history traces its beginning to a remarkably small group of children, just seven, and to the later establishment of Spring Dale Senior School in 1981. Those details interest me more now than they did then, because in retrospect I can see how different the act of learning was before universal digital access. Knowledge had weight. It occupied shelves.

A book had to be obtained and opened. A teacher had to be heard. A difficult concept had to be wrestled with long enough to become part of memory, because there was no expectation that the answer would be instantly available whenever it was needed. There was no smartphone in a student’s hand, no search engine waiting for a sentence, and certainly no large language model willing to explain the same idea at ten levels of difficulty until one of them finally connected.

Many American readers who grew up before the internet will recognize the feeling. It was the era of the encyclopedia set on the living-room shelf, the card catalog, handwritten notes, and the teacher whose explanation could suddenly make an impossible subject coherent. You did not simply “access” information. You went and found it. Because retrieval was difficult, internalizing knowledge had unusually high value.

That scarcity had substantial costs. Geography mattered more. Wealth mattered more. Access to excellent teachers and institutions mattered more. Brilliant students could be separated from opportunity by distance or circumstance. Digital information has democratized knowledge in ways that should never be dismissed, and a student today can learn in an afternoon what once took a semester and a library card.

But scarcity trained the mind differently. It encouraged a person to carry more of the intellectual apparatus internally, to build a structure inside the head into which new facts could be fitted. That would matter when I entered medicine.

Khalsa College, 1994–1996: The Physician’s Mind as the First Database #

In 1994–1996 I studied pre-medical sciences at Khalsa College in Amritsar, completing that pre-medical class in 1996. Khalsa College was founded in 1892 through the educational ambitions of the Singh Sabha movement, which sought to bring modern scientific education into dialogue with Punjab’s own intellectual and cultural traditions rather than simply replacing them. I did not appreciate the symmetry at the time. I was preparing for a scientific profession in an institution born during one period of cultural and technological transformation, just as medicine itself was approaching another.

For American readers, the path differs from the familiar one of a four-year undergraduate degree followed by medical school. In India, students complete intensive pre-medical science studies and then enter a combined medical program directly. The structure differs; the purpose is the same: a hard grounding in the sciences before clinical training begins.

In those years the most important information-processing device in medicine was still the human brain. No one spoke of it in such mechanical terms. We simply called it knowing medicine.

You learned anatomy until the geography of the body existed inside you. Physiology stopped being a set of equations and became an explanation of why the body behaved as it did. Pharmacology meant knowing not only the names of medications but their mechanisms, interactions, and consequences. Pathology meant seeing patterns. History-taking meant learning that the first answer to a question is not always the important one. Examination meant using eyes, ears, hands, and judgment together. Books were not merely references. They were the technology through which one loaded information into memory.

Senior physicians were another kind of database. A teacher who had seen thousands of patients carried patterns no textbook could fully reproduce. One learned by watching where the experienced physician looked, which question came next, which apparently minor clue changed the differential diagnosis, and, perhaps most importantly, what did not provoke unnecessary action.

Because information was harder to retrieve, physicians were expected to possess more of it internally. The physician’s mind functioned at once as database, search engine, pattern-recognition system, decision-support tool, and audit trail. It was imperfect. Memory fails. People tire. Experience can create blind spots as easily as wisdom.

But that system had one enormous advantage: cognition and responsibility were joined in the same person. The one who made sense of the information was usually the one who made the decision and answered for it.

We are now beginning to separate those functions. That may prove to be the most consequential change of all, and much of what follows is an argument about how to separate them without losing the accountability that their union once guaranteed.

Medical School in Amritsar: The Paper Chart and Its Honest Limitations #

I went on to Sri Guru Ram Das Institute of Medical Sciences and Research in Amritsar and completed my MBBS in 2002. The MBBS, Bachelor of Medicine and Bachelor of Surgery, is the primary medical degree in India and many other countries, the equivalent of the American MD.

The world was already digitizing rapidly, but everyday clinical medicine still had an architecture earlier generations would have recognized. The patient spoke. The physician listened and examined. The laboratory and the radiology department supplied more pieces. The record preserved some portion of what had happened. Then the physician tried to turn those fragments into a coherent story. Paper medicine had many weaknesses, and I have no desire to romanticize them:

• Charts disappeared, or their pages ended up out of sequence.
• Handwriting could turn a medication order into a hazard.
• A laboratory result could be overlooked.
• Important information might sit at another hospital with no practical way to obtain it quickly.
• A consultant’s recommendation could be buried dozens of pages deep.
• If nobody remembered to ask about something, it could effectively cease to exist.

These were not charming inconveniences. Patients could be, and were, harmed. But the paper chart’s failures had one redeeming quality: they were honest. You knew when you did not have the chart. You knew when the handwriting was illegible. You knew when records from another hospital were unavailable. The incompleteness was visible, and visible gaps provoke questions.

Digital medicine would solve many of those failures while creating a subtler kind. The electronic chart can give the appearance of completeness even when its information is duplicated, outdated, contradictory, or wrong. A gap you can see makes you cautious. A gap disguised as a complete record makes you confident. That is a far more dangerous illusion.

There was also one quality of the paper era I did not appreciate until much later: the record remained subordinate to the encounter. We practiced medicine, and afterward we documented what had occurred. The record could be burdensome, but it had not yet become the environment through which almost every clinical act had to pass. That relationship was about to reverse.

Interlude: A Short History of Writing Things Down #

Before going further, it is worth pausing on how strange the medical record is as an object, because nothing about the AI debate makes sense without it. Physicians have written down what happened to patients for as long as there has been writing. The Hippocratic Epidemics, compiled roughly twenty-four centuries ago, contain case histories that read almost like modern progress notes: the day of illness, the fever, the sleep, the stools, the outcome. What is missing from them is the thing we now take for granted. They are not organized around the patient across time. They are organized around the physician’s observation.

For most of history, that was the pattern. The record belonged to the doctor, lived in the doctor’s notebook, and died with the doctor’s practice. A patient who saw three physicians had three records, or none.

The modern idea, that one patient should have one record that follows that patient everywhere, is surprisingly young. At the Mayo Clinic in 1907, Henry Plummer replaced the ledgers in which each physician kept his own notes with a single dossier per patient, filed under a unique number, that any physician in the institution could pull. It sounds like a filing decision. It was a philosophical one. The record stopped being the physician’s diary and became the patient’s biography.

Sixty years later, Lawrence Weed, a physician in Vermont, looked at the chaos those dossiers had become and proposed that the record should be organized not by who wrote what but by what was wrong with the patient. His problem-oriented medical record, published in 1968, gave the world the problem list and the SOAP note (Subjective, Objective, Assessment, Plan), which every American medical student still learns. Weed’s insight is the one AI now has to inherit: a record is not a pile of facts. It is an argument about what the facts mean.

Artificial intelligence in medicine is also older than most people think. In the 1970s at Stanford, a system called MYCIN could recommend antibiotics for blood infections by reasoning through several hundred hand-written rules, and in tests it performed about as well as the specialists it was compared with. It was never used on a single real patient. The reasons had nothing to do with accuracy. Nobody knew who would be responsible if it was wrong, the computers of the day could not reach the bedside, and physicians did not trust advice from a box that could not explain itself in terms they accepted.

Half a century later, we are asking exactly the same three questions about systems that are incomparably more capable. Who is responsible? Can it reach the point of care? And can it show its work? The technology changed beyond recognition. The questions did not.

UAMS, 2007–2008: Medicine Between Two Worlds #

American medical training after medical school is counted in postgraduate years, or PGYs. The first, PGY-1, is what most people still call internship. Mine was at the University of Arkansas for Medical Sciences in Little Rock, from 2007 to 2008. Electronic medical records were already a normal part of clinical work there. They were not Epic.

The vendor name itself matters less than what the distinction reveals: the history of digitization is usually oversimplified. American medicine did not wake up one morning, remove every paper chart, and install Epic. The transition was long and messy. We lived among systems layered on systems:

• a laboratory application knew one part of the patient, and a radiology system another;
• orders might be electronic, and some documentation might be electronic;
• billing lived somewhere else entirely;
• paper still survived, and fax machines certainly survived.

A physician often needed practical knowledge not only of medicine but of where a particular piece of information lived. Yet the change was unmistakable. A result that once had to be physically found could increasingly be retrieved. A previous note could appear on a workstation. Orders could travel without relying on handwriting. Information could exist in several places at once. The chart was becoming less an object and more a network.

For an intern, the computer quickly became as unavoidable as the pager. We were learning medicine during the adolescence of the EHR: clearly permanent, increasingly powerful, but not yet integrated enough to disappear into the background.

It was also my first sustained lesson in the handoff problem in its modern form. Every morning and every evening, interns pass patients to one another. In that transitional world, the handoff carried not only clinical judgment but a map of which system held which fact. Lose the map and you lost the fact. Khadur Sahib’s old question, how a teaching passes faithfully from one set of hands to another, turned out to be a question about hospital sign-out.

UCSF Fresno and Community Regional, 2008–2010: Learning Medicine Before the Integration #

I completed PGY-2 and PGY-3 from 2008 through 2010 with UCSF Fresno at Community Regional Medical Center, the large teaching hospital that anchors care for much of California’s Central Valley.

Fresno taught me something no technology demonstration can: systems matter most where patients have the least margin for systems to fail. The Central Valley cares for farmworkers, working families, immigrants, and people with limited access to preventive care. Many patients arrive late in the course of illness, and their medical decisions are inseparable from employment, transportation, family obligations, language, and money. It is also a region of many languages: English and Spanish, but also Punjabi, Hmong, and others. In such a setting, the elegant treatment plan written in a chart matters only if the patient can understand it, afford it, reach it, and carry it out.

Residency also placed me directly inside the transitional world of health information technology. We moved between applications, searched for information in different places, and learned the practical geography of the hospital’s information systems. Residents everywhere become experts in these things, because the hospital never pauses long enough to wait for perfect infrastructure. We learned which screen held the truth, which held an older version of it, and which telephone number actually reached a human being. I finished residency in 2010. Community Medical Centers moved to Epic in 2011. Medicine, at least in my corner of it, entered another era.

The Epic Era: When the Record Became the Room #

When large integrated electronic health records arrived, their advantages were so substantial that it is easy to understand why they spread:

• the laboratory history, medication list, and allergies were all visible in one place;
• hospital notes and imaging reports could be found and retrieved;
• prescriptions were legible, and orders traveled electronically;
• patient messages increasingly entered the same ecosystem;
• the chart could not disappear because another department had borrowed it;
• a decade of laboratory values could be graphed in seconds;
• a physician seeing a patient at night could know what had happened that morning.

The improvement in accessibility was profound, and it saved lives in ways that rarely make headlines: the allergy that was seen, the duplicate prescription that was caught, the result that could no longer be lost in a stack of paper.

Epic Systems, based in Verona, Wisconsin, became one of the defining companies of this transformation. Its story is itself a piece of American medical history: the company was founded in 1979, in Madison, by a computer scientist named Judy Faulkner, and grew for decades without venture capital or a public stock listing into the software that now holds the records of a very large share of American patients. It never held a literal monopoly on American EHRs, and that word should be used with care. But among large health systems it became so prominent that many physicians began to experience it not as an application but as infrastructure, like the building’s plumbing.

MyChart entered the vocabulary of patients. The In Basket, the physician’s electronic inbox, entered the vocabulary of physicians. Order sets, medication reconciliation, health maintenance, SmartTools, problem lists, and results review became ordinary features of clinical work. An entire generation of physicians learned not merely to use an electronic chart, but to practice medicine through one.

Federal policy accelerated the change. The HITECH Act of 2009 created billions of dollars in incentive payments for hospitals and practices that adopted electronic records and demonstrated what the government called “meaningful use” of them, with penalties to follow for those that did not. Within a few years, adoption that had crawled for decades became nearly universal. It is worth remembering that the great digitization of American medicine was not simply a market choice. It was, in large part, a policy decision, and the shape of the record physicians now live inside was formed as much by regulation and billing as by clinical need. Later rules under the 21st Century Cures Act generally required that patients be able to access much of their own record electronically, including their clinicians’ notes. The record was no longer only the physician’s working document. It had become a shared, legal, billing, regulatory, and personal document all at once, read by clinicians, auditors, insurers, lawyers, and patients. The record had become powerful enough that we reorganized medicine around it. That was the bargain. And the price of the bargain was attention.

One Question, Three Eras #

Consider a single, ordinary clinical question, the kind that arises a dozen times a day: Has this patient ever had a bad reaction to a sulfa drug? In Amritsar in 2002, the answer lived in three places. It lived in the patient’s memory, which might be accurate, vague, or wrong. It lived in a paper chart, if the reaction had happened at this hospital, if someone had written it down, if the page had survived, and if I could find it in the next four minutes. And it lived in the physician’s mind, if the physician had happened to be the one who saw the rash years ago. Any of those sources could fail silently. If all three failed, the patient got the drug.

In Little Rock in 2008, the answer might be in the allergy field of one system, or in a scanned document in another, or in a discharge summary from a hospital across town that our software could not see. It was probably somewhere. Finding it required knowing where to look, and the intern who knew was on the other service.

In Fresno in 2026, the answer is almost certainly in the chart. So is everything else. The allergy field says “sulfa: rash,” entered in 2013 by a nurse who was transcribing what the patient’s daughter said. It also says “no known allergies,” entered in 2019 by a locum physician who did not scroll down. The two entries have coexisted for seven years. Nobody has reconciled them, because reconciling them is nobody’s job and every physician who noticed was already eleven minutes behind.

Now imagine the fourth era. Software reads the whole record, notices the contradiction, traces each entry to its source, tells me which is more likely to be true and why, and asks me to decide before the prescription is written.

That is what AI in medicine actually looks like when it works. Not a robot doctor. A very fast, very patient reader that never forgets to scroll down, and that is required to tell me where it got its answer. The rest of this essay is about what happens when that reader also learns to act.

Why Your Doctor Keeps Looking at the Computer #

Patients noticed before the profession fully admitted what was happening. The physician entered the examination room and increasingly faced two patients: the human being in the chair and the digital representation glowing on the screen. The human being had a story. The screen had requirements. A single visit could demand all of this:

• open the last note, then the older notes;
• review and reconcile the medication list, and review allergies;
• check today’s laboratory results, then the old ones;
• find the specialist’s letter and open the imaging report;
• respond to the warning, and dismiss the irrelevant warning;
• update the problem list;
• place the order, write the prescription, and choose the diagnosis;
• create the referral and satisfy the quality measure;
• write the note, select the billing code, and close the encounter.

Then answer the message that arrived while doing all of that. Then the next message, and the next. The profession eventually developed its own rueful vocabulary for what followed: documentation burden, after-hours EHR work, “pajama time.” Many physicians finished their charts at the kitchen table after their children were asleep. Burnout research began to treat the electronic record not as a neutral tool but as one of the working conditions of the job. The irony is almost perfect. The EHR solved information scarcity so effectively that it produced information overload.

A patient who has lived with several chronic diseases for twenty years can now generate an extraordinary digital wake:

• thousands of laboratory values and hundreds of prescriptions;
• imaging studies, hospitalizations, and specialist notes;
• portal conversations, referrals, and scanned documents;
• an accumulating problem list that may contain true diagnoses, resolved diagnoses, suspected diagnoses, and errors copied from note to note until repetition begins to masquerade as corroboration.

The note itself changed shape. Templates and copy-forward functions made it easy to carry yesterday’s words into today’s document, and billing and legal pressures rewarded length. Many notes grew longer while becoming harder to read. The physician looking for the one sentence that mattered often had to dig through pages of text that were technically accurate and practically useless.

Five minutes before the visit, the physician may have access to almost everything. The practical question remains ancient: What actually matters? That is where Guru Angad’s verse took on a meaning for me that it did not have when I was young.

ਜੇ ਸਉ ਚੰਦਾ ਉਗਵਹਿ ਸੂਰਜ ਚੜਹਿ ਹਜਾਰ ॥ A hundred moons. A thousand suns. Enormous light, and still darkness. Modern medicine is no longer dark because information is absent. Sometimes it is dark because all the information is shining at once. Data are not understanding. A complete chart is not necessarily a coherent story. A thousand signals are not a decision. The next technological revolution is beginning precisely there.

AI Arrives Quietly: The Computer That Listens #

Artificial intelligence did not enter ordinary outpatient medicine wearing a white coat or announcing itself as an autonomous diagnostician. It began by offering to write the note. Ambient AI systems, often called AI scribes, listen to the encounter with the patient’s consent and create a draft for clinician review. The appeal to physicians is immediate, because it attacks an absurdity embedded in modern practice: after conducting the medical encounter, the physician often spends additional time reconstructing that encounter for the machine.

Ambient documentation begins to invert that relationship. Instead of physician practices medicine, then recreates medicine for the computer, we move toward physician practices medicine, and the computer constructs the record around the encounter.

That is more than a convenience. It is an architectural shift. For the first time since the EHR arrived, the record begins to follow the conversation rather than the conversation following the record.

But excitement must not outrun evidence. A 2026 Journal of General Internal Medicine scoping review screened 3,203 manuscripts and identified 61 studies meeting its criteria:

• 26 focused on model development;
• 8 reported clinical-trial results;
• 11 involved real-world implementation.

The authors concluded that primary care AI remains largely developmental, with limited real-world use beyond areas such as ambient scribing, clinical decision support, and workflow automation. For the same reason, I do not repeat the widely circulated claim that AI scribes save a fixed number of minutes per visit. Published results vary with the setting, the specialty, the tool, and how carefully clinicians review what the software writes. Some physicians report that they can look at their patients again. Others find that editing a fluent but imperfect draft carries its own burden. Both experiences are real.

That is the correct place from which to think about AI in 2026: neither a skepticism that refuses to see the transformation, nor an enthusiasm that mistakes a demonstration for a mature health care system.

The technology is moving faster than the evidence. Medicine has seen that before, with new drugs, new devices, and new procedures, and the profession’s hard-won habit in each case has been the same: welcome the promise, then demand the proof.

The More Important Distinction: Answering Versus Acting #

The next leap is larger.

Generative AI answers. Agentic AI acts.

That difference sounds like vocabulary until one follows its consequences. A generative system, the kind many Americans now use as a chatbot, responds when asked. It can examine the chart and tell me that a patient’s hypertension appears poorly controlled.

An agentic system can be assigned an objective: help determine why this patient’s hypertension remains uncontrolled, and move the case toward resolution. In pursuit of that objective it can:

• retrieve recent home blood-pressure measurements and judge whether there are enough of them to mean anything;
• review the medications and examine refill history;
• check creatinine and potassium, and identify whether required laboratory monitoring is missing;
• ask the patient for additional readings;
• determine whether a proposed next step fits a pre-approved clinical pathway;
• prepare, but not necessarily sign, the medication order and the laboratory tests;
• send instructions and schedule follow-up.

Then, critically, it can remember the objective. Two weeks later it can ask whether the laboratory test happened, whether the patient obtained the medication, whether the readings improved, whether potassium rose, and whether the patient reported dizziness. It can then decide whether the case now requires physician review. The important property is not merely intelligence. It is persistence.

The traditional EHR waits. It waits for the physician to open it, for somebody to find the result, for a medical assistant to route the request, for somebody to notice that the referral was never completed, for the patient to return, for the next click.

The agent continues. That changes everything, including the handoff problem. An agent does not go off shift. It does not forget, at 7 a.m., what the night team meant to do.

Persistence cuts both ways, however. A human being who makes a mistake usually makes it once. A persistent system built on a mistaken premise can carry that mistake forward with perfect diligence, every day, for every patient it touches. The same quality that makes agents valuable makes their errors durable.

Epic and the Strange Possibility of the Post-Epic Epic #

For physicians already working in Epic Ambulatory, the outpatient version of Epic, this transition is no longer purely theoretical. Art, Epic’s clinician-facing AI, already includes functions such as:

• outpatient note drafting, and proposed orders and diagnoses derived from the encounter;
• chart summaries and patient-message assistance;
• medication insights and patient instructions;
• coding suggestions;
• natural-language questions asked directly of the chart.

Ergo Visit came next. In August 2026, Epic announced that Ochsner Health had become the first health system to use it. It is an AI-supported outpatient workflow designed to help clinicians arrive prepared, surface relevant information during the encounter, and have documentation and orders lined up for review before the visit ends.

That deserves a pause. The important feature is not a better note. It is the idea that the record begins anticipating the work the encounter will require. Emmie, Epic’s patient-facing system, points in the other direction. Within MyChart, it can answer patient questions using chart information, help explain clinical information, generate reminders, and support scheduling.

Agent Factory points toward yet another layer: health systems building their own agents to complete multi-step tasks rather than merely generate text. Epic has publicly described examples involving medication-refill protocols, patient outreach, and medication-conflict detection.

These are company descriptions, not independent proof that every deployment works equally well. They should be read the way a physician reads a manufacturer’s brochure: as a statement of intent, pending evidence.

As a map of where the industry is heading, though, they are revealing. Epic spent decades building the place where health care information lives. It is now trying to build systems that can work on top of that information. Which leads to a fascinating paradox: Epic may become more important by requiring physicians to consciously use less of Epic.

A Visit in 2030 #

Imagine a routine morning several years from now. My next patient is 67, with type 2 diabetes, coronary artery disease, stage 3 chronic kidney disease, and hypertension. Today, I might open several parts of the chart, scan old notes, compare laboratory values, inspect the medication list, and reconstruct what has happened since the last visit, all in the minutes between patients. A mature system might instead give me a concise briefing:

• The A1c, a blood test reflecting average blood sugar over roughly three months, has risen from 7.1 to 8.0 percent.
• Kidney function is approximately stable, with an eGFR around 52.
• LDL, the “bad” cholesterol, is 83 on current treatment.
• Home blood-pressure readings average approximately 142/82.
• Pharmacy information raises concern about inconsistent use of one blood-pressure medication.
• The diabetic eye examination is overdue.
• He has mentioned getting up at night to urinate, which physicians call nocturia, in two recent messages.
• There has been no hospitalization since the last visit.

Every sentence links to its source. That requirement is essential. I do not want an oracle. I want a navigator. I enter the room oriented, but not prejudged. Then the patient tells me something the database does not know. He has not been taking the blood-pressure pill consistently, because when he takes the full dose before work he becomes light-headed. He works a physical job, starts early, and has been afraid of falling on the site.

The computer saw nonadherence. The conversation reveals a clinical problem: perhaps blood pressure dropping when he stands, perhaps a dose or timing problem, perhaps a medication that simply does not suit his day. The plan changes entirely.

That distinction contains much of the future of medicine. Data can tell me that a medication was not taken. A relationship may be required to tell me why. Suppose we decide to modify the regimen. The system can immediately bring forward kidney function, potassium, allergies, prior intolerance, drug interactions, insurance information, and monitoring requirements. The rest of the plan follows from what we decide:

• if we agree on a change, the prescription is prepared for my review;
• if I want a chemistry panel in two weeks, the order is prepared;
• if we discuss colon cancer screening, the referral is prepared;
• if I ask to see him again in six weeks, scheduling begins;
• the after-visit instructions are generated from what we actually decided, in plain language, and in Spanish or Punjabi if that is what he reads.

I review. I correct. I approve. Then I leave the room. I have practiced medicine. I have not begun a second shift recreating medicine for software. That is far more consequential than dictation.

Medication Management: Where Agents May Mature First #

Medication management may become one of the earliest genuinely agentic clinical domains, because the surrounding information is unusually structured. The relevant facts include:

• renal and liver function;
• allergies and prior intolerance;
• interactions and current dose;
• laboratory monitoring and treatment history;
• insurance coverage and patient cost;
• adherence and clinical targets.

Much of that can be retrieved automatically. The physician’s job should increasingly be to decide what the information means, not to spend minutes locating it. Consider today’s prescription workflow. A physician and patient decide on a medication, and the prescription is sent. Hours or days later, the office learns that the drug is not covered, or requires authorization. Staff investigate. A preferred alternative is identified. The physician reviews the issue again. A second prescription is sent. Meanwhile, the patient has gone without treatment, or has paid a price at the pharmacy counter that nobody discussed in the exam room.

That workflow exists because the administrative facts become visible too late. The technological direction is toward making them visible at the moment of decision:

• if a medication will cost the patient hundreds of dollars, I want to know while the patient is still sitting in front of me;
• if the insurer requires a different medication first, I want to know before I prescribe;
• if prior authorization is required, the system should already know what documentation will be necessary.

Administrative reality should move upstream into clinical decision-making, instead of ambushing the plan afterward.

The Refill Queue Should Become an Exception Queue #

Refills illustrate the same principle even more clearly. A stable patient requesting an unchanged thyroid medication with appropriate monitoring is not the same clinical problem as a patient requesting a medication after kidney function has deteriorated. Yet the electronic workflow may present them as two nearly identical boxes waiting for a signature.

That makes little sense. An agent should be able to determine that one request satisfies a defined protocol while another does not. A future refill environment could quietly sort requests into categories:

• routine;
• missing monitoring;
• needs patient information;
• dose discrepancy;
• potential interaction;
• high risk;
• physician judgment required.

The purpose is not to let software prescribe indiscriminately. The purpose is to stop spending physician cognition on transactions that do not need it. I do not want an AI that simply renews everything. I want an AI that reliably recognizes which request I actually need to think about. That is a fundamentally different goal, and it should be designed, tested, and measured as one.

The In Basket May Eventually Stop Being an Inbox #

Primary care physicians talk about the In Basket as though it were a force of nature. It is not. It is the place where fragmentation becomes someone’s labor. It fills with items of every kind:

• laboratory results;
• patient messages;
• refill requests;
• hospitalization notices;
• specialist notes;
• forms;
• insurer requests;
• pharmacy clarifications;
• and then another laboratory result.

Each item exists because some system generated information that another human now has to interpret or route. AI changes this most when it stops merely summarizing the pile and begins executing its routine parts:

• a normal screening mammogram produces the appropriate letter to the patient;
• a stable abnormal laboratory value is compared with prior results before a response is drafted;
• a potassium of 6.1 is escalated immediately;
• a hospital discharge triggers medication reconciliation and follow-up;
• a specialist’s recommendation is compared against the current care plan;
• a patient asking whether ibuprofen is safe triggers a review of kidney function, blood thinners, allergies, and medication history before any answer is prepared.

The future inbox may then cease to be a pile of transactions. It may become an exception console. The distinction matters because human beings are better at exceptions than at repetition, and machines are better at repetition than human beings. Health care has spent decades assigning the work backward.

Handoffs: Where the Record Breaks #

If there is one place where I would ask AI to prove itself first, it is at the handoff, the old question of Khadur Sahib in its clinical form. Much preventable harm in medicine happens not because nobody knew something, but because the person who knew it was not the person who needed it. Patients leave the hospital with new medications that never reach the primary care list. Specialists recommend repeat imaging that nobody orders. A test result arrives after the ordering physician has rotated to another service. A patient changes clinics and brings a stack of paper, or nothing at all.

The electronic record reduced some of these failures and disguised others. Information now usually exists somewhere. The question is whether anyone who needs it will encounter it in time. A well-designed agent is, in principle, an excellent custodian of handoffs. It can watch for the discharge summary and compare its medication list against the outpatient list. It can notice that a recommendation in a consultant’s note never became an order. It can track a pending result across a change in the ordering clinician. It can hold the thread when the humans holding it change.

That is also where it must be most carefully governed. A handoff is not only a transfer of facts; it is a transfer of responsibility. The system must make clear, at every step, which human now owns the problem. An agent that carries the information but blurs the ownership has solved half of the problem and hidden the other half.

Language Access: A Promise and a Test #

In the Central Valley, where I trained and practice, a patient’s language can decide the quality of care as surely as any laboratory value. Interpreters are essential and often scarce. Written instructions are frequently available only in English, or in a translation of uneven quality.

AI could change this more than almost any other feature. It could make after-visit instructions in a patient’s own language routine rather than exceptional, and portal messages readable by the patients who receive them. It could support a real-time conversation that previously required waiting for an interpreter.

It is also a place where errors are unusually hard to see. A physician reviewing an English note can catch a mistake. A physician reviewing a translation into a language he or she does not read cannot. A mistranslated dose or warning may be invisible to everyone except the patient, who may have no way of knowing it is wrong.

The standard should therefore be demanding. Translation tools used for clinical instructions need validation in the specific languages and populations they serve, not only in the handful of languages where testing is easiest. Qualified human interpreters remain necessary for consent, serious diagnoses, and any conversation where misunderstanding carries real consequences. Language access is where the langar principle becomes concrete: a technology that works well only for patients who speak the dominant language is not a neutral technology.

When the Insurance Company Has AI Too #

There is another side of the transaction that deserves much more attention. Physicians will have AI. Patients will have AI. Insurers will have AI. Those systems will not necessarily share the same objective.

Prior authorization, the requirement that an insurer approve a test, procedure, or drug before paying for it, offers an early glimpse. Nearly every American has met it. CMS-0057-F, the federal interoperability and prior authorization rule, sets decision deadlines for affected payers. For covered medical items and services within its scope, payers must decide expedited requests within 72 hours and standard requests within seven calendar days. Drugs are excluded from these particular requirements, and medication prior authorization follows different regulatory pathways.

As standardized electronic interfaces expand, provider-side software can increasingly assemble the clinical material supporting a request, while payer-side software evaluates whether the requirements are met. That sounds efficient, and it could be. It also creates the possibility of algorithmic friction:

1. The provider’s agent submits the justification.
2. The payer’s algorithm rejects a required element.
3. The provider’s agent supplies additional documentation.
4. The payer’s software identifies another criterion.

The argument becomes faster but not necessarily more humane. Machines may conduct a bureaucratic dispute thousands of times faster than people, and the patient can still be waiting at the end, now with less visibility into why. Technology does not eliminate incentives. It amplifies whatever incentives are already present. That is why law and policy will matter as much as engineering.

California Has Already Drawn Some Lines #

California has begun to answer several of these questions. AB 3030 took effect January 1, 2025. In specified physician offices, clinics, and health facilities, when generative AI is used to create communications about a patient’s clinical information, the patient generally must be told that AI generated the communication. The patient must also receive clear instructions for reaching a human. There is an important exception: the requirements do not apply when the communication has been read and reviewed by a licensed or certified human health care professional.

That exception tells us something. The law recognizes a meaningful difference between assistance and delegation. AI can help produce the communication. Human review changes the structure of accountability. SB 1120 addresses another boundary. It regulates health plans’ and insurers’ use of AI, algorithms, and other software in utilization review and management. Such tools may not independently make medical-necessity determinations that deny, delay, or modify care. Those determinations must remain with appropriately qualified licensed professionals.

Again, the principle is revealing. Software may assist, organize, and analyze. But when the consequence becomes sufficiently serious, California has begun insisting that a human professional remain in the chain of responsibility.

The American Medical Association has moved in the same direction. In 2026 it adopted policy emphasizing that AI should support rather than replace physician judgment. The policy treats physician oversight, transparency, evidence of safety, and accountability as central to clinical use. The AMA has also argued that liability should not simply be placed on the physician regardless of where an AI failure originated. On its view, accountability should follow the party best positioned to understand and reduce the relevant risk, including developers in appropriate circumstances. That debate will become far more important as AI moves from advice to action.

Who Is Responsible When the Machine Is Part of the Chain? #

Imagine an adverse event several years from now:

1. An AI system summarizes a chart incorrectly.
2. A second agent relies on that summary to prepare a medication change.
3. The order reaches the physician, who reviews it quickly during a crowded clinic session and approves it.
4. The patient is harmed.

Who failed? The physician? The health system? The software vendor? The person who configured the workflow? The institution that required its use? The model developer? The data pipeline? Some combination?

Traditional malpractice law developed around human professional decisions. Agentic medicine creates chains of mixed human and machine activity. The meaningful legal question cannot simply be “Was AI involved?” It will be a series of harder questions:

• What was the system designed to do, and what evidence supported its use?
• What did the vendor represent, and what did the health system know?
• Could the physician reasonably have identified the failure?
• Was the recommendation explainable?
• Was the model being used within the setting in which it had been validated?
• Was human review meaningful, or merely ceremonial?
• Did the organization’s productivity expectations make genuine review realistic?
• Did the software actually malfunction, or did the physician make an independent clinical error?

Those are the questions any careful reviewer already asks of a chain of human error. The difference is that some links in the chain will now be software, and software does not testify. It leaves logs, if someone thought to keep them.

Those distinctions will determine whether “human in the loop” becomes a genuine safety measure or merely a contractual mechanism for placing liability on whoever clicks “sign.” Medicine should decide that before the litigation decides it for us.

The New Clinical Skill: Asking the Machine “How Do You Know?” #

Diagnostic AI may ultimately become a significant part of this transformation. The machine has an ability humans do not: simultaneous attention to an entire longitudinal record at enormous scale. It can notice:

• that a pulmonary nodule was never followed up;
• that platelet counts have remained elevated for two years;
• that a chronic cough began after an ACE inhibitor was started;
• that a specialist recommended repeat imaging and the test never happened;
• that apparently resistant hypertension coincides with erratic medication filling;
• that a diagnosis copied through years of notes was never actually established;
• that a hemoglobin has drifted slowly enough to escape notice at any single visit.

These are meaningful capabilities. But there is a dangerous asymmetry between human and machine error. A confused human often looks confused. A computer can be wrong beautifully. It can produce a polished, logically organized explanation built on a false premise. It can mistake repetition for corroboration, treating a copied diagnosis as independent confirmation. It can convert an inference into a fact, or omit a contradictory piece of evidence without announcing the omission. The resulting recommendation can look more authoritative than it deserves.

That creates automation bias, the human tendency to defer to the machine. The response should not be to reject AI. It should be to demand provenance:

• show me where the diagnosis came from;
• show me the laboratory trend, the note, and the imaging report;
• show me what supports the conclusion and what argues against it;
• show me what information is missing;
• separate what the patient said from what a clinician observed and from what the model inferred.

The most important button in future clinical AI may not be “Accept.” It may be “Show me why.” A record repeated is not a record corroborated. A prediction is not a diagnosis. A fluent answer is not evidence. Those principles will need to become as instinctive for the next generation of physicians as confirming a patient’s name before signing an order.

AI Literacy Will Become Part of Medical Literacy #

Future physicians do not need to become software engineers, any more than a physician must know how to build an MRI scanner to understand when MRI helps, where it fails, and how to interpret its limits. AI will be similar. Physicians will need to understand enough to recognize its failure modes:

• What is hallucination, and how do I detect it?
• What is automation bias, and how does it show up in my own behavior?
• What population was this model trained on, and does it perform equally well for the patients in my clinic?
• Was it validated prospectively, or only on historical data?
• Does its output reflect correlation or causation?
• How does it communicate uncertainty, and what happens when its input data are incomplete?
• Can its performance drift over time, and who monitors it after deployment?
• What happens when two automated systems disagree, and when must the physician override them?

The AMA now explicitly treats education of physicians and patients as part of responsible AI deployment. The hardest skill may be psychological rather than technical. Physicians are trained to seek expertise, and AI speaks with the grammar of expertise. It does not hesitate the way a human consultant hesitates. It does not look uncertain unless it was designed to express uncertainty. It can produce a beautifully structured answer in seconds.

The physician of the AI era must develop an unusual reflex: fluency is not evidence; confidence is not calibration; and a machine able to read everything can still misunderstand the patient sitting three feet away.

The Patient Will Have AI Too #

For most of medical history, there was a profound asymmetry of information between doctor and patient. The physician had the books, the language, and the record. The internet narrowed that gap. Generative AI may change it far more dramatically. Patients will increasingly arrive having already asked a machine:

• what a laboratory result means;
• what diseases could cause a symptom, and whether it could be cancer;
• why they are taking a medication, and what alternatives exist;
• whether they should ask for an MRI;
• what questions to ask the doctor.

Epic’s Emmie is one early institutional version of the same change, offering conversational assistance around information already inside MyChart. This does not make physicians irrelevant. It clarifies what physicians are actually for. The physician becomes less valuable as the patient’s exclusive source of information, and more valuable as the person who determines what information means for this person:

• which possibility is plausible, and which is important;
• which recommendation applies to this patient;
• what we are willing to risk, and what we should ignore;
• what to do first;
• what the patient actually wants;
• what the real problem is underneath the stated problem.

Those are higher-order tasks. AI may strip away some of the simpler work and reveal them more clearly. I think often about how different this is from my education in Amritsar. When I began pre-medical studies at Khalsa College in 1994–1996, information was scarce enough that obtaining it was itself part of education. You went to the book, the library, the teacher. You remembered because retrieval was expensive.

In 2026, information is nearly free. What is scarce now is attention. Judgment is scarce. Context is scarce. Trust is scarce. Responsibility is scarce. Those may become the currencies that matter most in medicine.

What Happens to Nurses, Medical Assistants, and the Front Office? #

The same reasoning applies to the health care workforce. Predictions that AI will simply eliminate medical assistants, nurses, and office staff misunderstand why so much of their work exists. Health care employs enormous numbers of people to compensate for broken information flow. Humans call, fax, and leave messages. They check whether the fax arrived. They log into insurer portals, re-enter information, retrieve records, route requests, schedule, remind, and chase missing information.

In many offices, people function as human integration software between digital systems that do not communicate properly. They are often highly capable people doing work far beneath their abilities, because the machines around them cannot finish a transaction on their own.

That is precisely the kind of work agents will increasingly absorb. The question is what health care does with the recovered human capacity:

• a medical assistant can spend more time guiding patients through their care;
• a nurse can spend more time on meaningful triage, education, and complex coordination;
• front-office staff can solve difficult access problems rather than repeating routine scheduling transactions;
• clinical pharmacists can manage higher-level medication problems.

The workforce does not necessarily disappear. The value of human work changes. But nothing guarantees that organizations will use the opportunity wisely.

The Efficiency Trap #

This may be the central operational risk. Suppose ambient AI saves a physician meaningful documentation time. What happens next? One possibility is that the physician goes home earlier, thinks more clearly, listens longer, and burns out less.

Another is that the organization immediately shortens appointments and adds more patients. Then AI reduces typing but increases the density of clinical decision-making. The physician moves from patient to patient faster. Each AI-generated note must still be reviewed. Each automated recommendation must still be supervised. Several agents may be performing tasks at once, and exceptions arrive continuously.

The nature of fatigue changes. Keyboard fatigue becomes micro-cognitive fatigue: hundreds of small approvals, rapid context switches, and repeated judgments about whether a machine-generated action is safe. That fatigue may be less visible than documentation burden, and potentially more dangerous.

If a person is expected to review fifty AI-generated decisions rapidly, “human oversight” becomes a legal fiction. A signature is not necessarily supervision. A click is not necessarily judgment. This is why productivity should not be the only measure used to evaluate clinical AI. Health systems should also be asking:

• Does the physician spend less time on after-hours work?
• Does the physician look at the patient more?
• Are fewer important follow-ups missed?
• Do authorizations happen faster?
• Are medication errors reduced?
• Are unnecessary tests increased?
• Does the technology create new alerts?
• Does it change diagnostic accuracy?
• Do patients understand when they are dealing with a machine?
• Is the physician still thinking independently?

These are harder questions than “How many more visits can we schedule?” They are also more important.

If AI Gives Us Time Back, We Do Not Have to Sell All of It #

This deserves to be said plainly. If technology gives a physician twenty minutes back, the health system does not have to monetize every recovered minute. Sometimes the return on technology is time itself:

• time to sit down;
• time to hear the second concern, the one the patient almost did not mention;
• time to examine carefully;
• time to explain why another MRI is not necessary;
• time to discover that “noncompliance” is actually fear, cost, dizziness, transportation, or a spouse who disagrees with the plan;
• time to recognize that the chief complaint is not the real complaint;
• time to call a family member;
• time to think;
• time to practice medicine.

For years, health care technology was sold with the promise of efficiency, while physicians often experienced it as additional work. AI has an opportunity to reverse that. If every saved minute simply becomes another unit of throughput, it will not.

The Equity Question: Which Future Does the Patient Receive? #

This is where my mind returns to Khadur Sahib and to langar, where everyone sits on the same floor and eats the same food. The same technology can create two very different health care systems.

In one, AI handles the clerical work so that a physician has more time with the patient. That patient experiences artificial intelligence as something almost invisible: the doctor sits down, already understands the history, and listens.

In the other, AI is used to reduce how often the patient reaches a physician at all. That patient meets a chatbot, then an automated triage system, then an asynchronous protocol, then escalating layers of contact before ever seeing a doctor. The software may be similar. The economics are different. It is not difficult to predict which population is most likely to receive the second model:

• safety-net patients and rural patients;
• patients with poor insurance;
• patients who speak less dominant languages;
• patients with transportation problems;
• patients living in places like the Central Valley, where access has always required more effort.

That is why equity cannot be bolted onto AI policy after deployment. It has to be part of the architecture. The question is not only whether AI performs accurately. It is what kind of human care the technology makes available afterward, and to whom.

The old ideal of langar remains unexpectedly relevant: technological abundance has little moral meaning if access remains stratified. A thousand suns that shine only on some patients are not progress. They are a new kind of shadow.

Privacy: Whose Record Is It? #

A system that reads everything raises an obvious question: who else is reading? Within a health system, patient information is governed by HIPAA and by state law, and the vendors that process it on a health system’s behalf are generally bound by contract to protect it. Much of the AI now entering clinical care operates inside that framework.

Consumer AI tools are different. When a patient pastes laboratory results into a general-purpose chatbot at home, that information is usually governed by the company’s own terms and privacy policy, not by the rules that bind the patient’s doctor. Many patients do not realize this.

There are institutional questions as well. What happens to recordings made by ambient scribes, and how long are they kept? Is de-identified patient data used to train commercial models, and on what terms? Who audits what an agent did, and can a patient see that record? These are not objections to AI. They are conditions of trust. A technology that asks patients to let it listen to the most private conversations of their lives must be unusually clear about what it keeps, what it shares, and what it forgets.

Continuous Primary Care: The Patient Between Appointments #

One of the most promising changes may be the least visible. Primary care has traditionally been organized around encounters because human attention is episodic. The patient comes in, we evaluate, we make a plan, and the patient leaves. Then dozens of other patients occupy our attention.

But the disease continues. Blood pressure does not wait for the next visit, and neither does diabetes. Kidney disease does not stop progressing because the appointment is three months away. A referral can remain unfinished for months. A medication can quietly stop being filled. A laboratory test can remain undone. The physician may rediscover all of this at the next visit, or not at all. Agents make another model possible. The system can notice:

• that the A1c is rising;
• that the retinal examination remains overdue;
• that renal monitoring was never completed;
• that a prescription was never filled;
• that home blood pressure remains high;
• that a specialist’s recommendation never became an order;
• that a follow-up was missed;
• that a patient discharged from the hospital never came back.

Then the software can take the appropriate permitted step: remind, request, offer scheduling, prepare, or escalate. The fundamental unit of primary care begins shifting from the appointment to the patient over time. That could be one of the most important changes in outpatient medicine in a generation.

The Danger of Automated Anxiety #

It could also become intolerable. The ability to identify every deviation does not mean every deviation deserves intervention. Medicine already struggles with alert fatigue. Imagine what happens when AI can discover an almost unlimited number of possible concerns:

• a slightly abnormal result generates another message;
• a theoretical interaction triggers another alert;
• a weak statistical association produces another recommendation;
• a risk score suggests another screening test;
• a data stream drifting slightly from baseline produces another task.

The patient receives more warnings. The physician receives more work. The health care system generates more medicine. That is not necessarily better health. It may simply be automated anxiety. Experienced medicine requires restraint. There are abnormalities to investigate, abnormalities to watch, and abnormalities to ignore. There are guidelines to follow, and patients for whom following the guideline mechanically would be foolish.

The best AI will not simply detect. It will help contextualize, and it will understand that the value of information depends on what we do with it. Sometimes the most intelligent action is no action at all.

AI May Make the Average Physician Encounter Harder #

There is another consequence almost nobody advertises. If automation succeeds, the easy work disappears first:

• the routine refill;
• the normal laboratory result;
• simple scheduling;
• straightforward education;
• the preventive reminder;
• stable, protocol-based follow-up.

What remains on the physician’s schedule is harder:

• multimorbidity and polypharmacy;
• frailty;
• diagnostic uncertainty and treatment failure;
• conflicting specialist recommendations;
• psychiatric complexity and social instability;
• nonadherence with complicated reasons underneath it;
• end-of-life decisions;
• symptoms that refuse to fit the textbook;
• patients whose priorities do not fit the guideline.

The physician may spend less time on clerical work while the remaining clinical work becomes more concentrated and more difficult. That does not mean AI failed. It may mean it succeeded. Perhaps medicine becomes less clerical and more medical, more physician-like again.

But we should design staffing, schedules, and expectations accordingly. A physician seeing increasingly complex cases cannot automatically be expected to see more of them simply because routine work has been automated.

The Things the Machine Still Cannot See Easily #

After enough years in medicine, one develops a respect for information that does not fit neatly into a field in the EHR.

• The patient says everything is fine, and it is not.
• The medication list is technically excellent, and the frail person taking those medications is miserable.
• The patient asks for a CT scan, and the real request is reassurance.
• Another patient refuses medication. The chart says noncompliance. The patient is afraid.
• Someone agrees to the entire plan, and before the visit ends, years of experience tell you none of it will happen.
• The spouse finally speaks, and the history changes.
• The patient walks into the room differently from the last time.
• A tremor appears when the patient reaches for something.
• A talkative person has become quiet.
• A daughter looks away when her father answers the alcohol question.
• The patient’s goals have changed, and nobody has asked.

Medicine happens partly in that space. The database sees variables. The physician sees a person living among other people. This is the akhara principle in the examination room: the body, and the life around it, is not an abstraction.

Perhaps AI will become better at some of these things. It almost certainly will. But there is a difference between detecting a pattern and entering into the responsibility the pattern creates. That is why the central question should not be whether AI can imitate more physician tasks. It should be which tasks ought to remain human even after machines can take part in them.

Medicine Needs a Philosophy of Work More Than an AI Strategy #

Every health system now wants an AI strategy. I increasingly think that is the wrong starting point. Medicine needs a philosophy of work. That philosophy starts with a set of questions:

• Which tasks truly require a physician?
• Which require another human being?
• Which require clinical judgment, and which require empathy?
• Which require physical presence?
• Which exist only because software has historically been unable to do them?

That distinction should govern automation. Searching, retrieving, transcribing, comparing, routing, scheduling, reminding, tracking, and repetitive reconciliation should increasingly belong to machines. Other work should remain with people. Listening belongs to humans. Examining belongs to humans when examination matters. Interpreting ambiguity, negotiating goals, counseling, explaining uncertainty, and accepting responsibility all remain deeply human.

There will be overlap, but the principle matters. The goal should not be maximum automation. It should be the maximum removal of work that does not require the human beings currently forced to perform it.

Epic May Disappear by Becoming Ubiquitous #

For an Epic Ambulatory practice, I would therefore be cautious about surrounding the existing EHR with a new collection of disconnected AI applications. The old problem was fragmentation. There is little wisdom in rebuilding fragmentation out of smarter products.

Integration matters because context matters. A narrower system can be safer and more useful than a spectacular general-purpose model operating outside the clinical record, provided it knows four things:

• exactly which patient it is dealing with;
• where every piece of information originated;
• which actions it may perform;
• when it must return control to a clinician.

The advantage of integration is not convenience. It is provenance and governance. This is why Epic’s position is so interesting. If its AI strategy succeeds, Epic may become more deeply embedded in medicine while its traditional interface becomes less visible. Today I operate Epic. I click, search, open, route, reconcile, and close. Tomorrow I may increasingly express intent:

• What happened to this patient’s kidney function after the medication was started?
• Why was this medication discontinued?
• What remains unresolved in three years of evaluation for dizziness?
• What did cardiology recommend?
• Did the patient complete the follow-up imaging?
• Which parts of today’s plan have not happened six weeks from now?

The system searches, retrieves, organizes, and shows the evidence. The physician decides. In that world Epic remains the record, the database, the interoperability layer, and the transaction engine. But the physician no longer spends the day consciously operating it. The EHR recedes because it has become more capable. That would complete an extraordinary circle.

From Khadur Sahib to the AI Clinic #

When I think about that circle, I sometimes travel back farther than Epic, Fresno, or Little Rock. I think about Khadur Sahib. I think about Amritsar and Spring Dale. I think about entering Khalsa College in 1994 to study pre-medical sciences, with no conception that thirty years later I would be practicing medicine in California while machines capable of reading enormous medical records began to sit invisibly inside the clinical encounter.

I think about the path from Khalsa College to medical school at Sri Guru Ram Das Institute. Then residency in America: PGY-1 at UAMS in 2007–2008, and PGY-2 and PGY-3 with UCSF Fresno at Community Regional Medical Center from 2008 through 2010. I think about the patchwork systems at UAMS before Epic, about Fresno before Community’s Epic transition, and about watching the electronic record become indispensable.

And I think about how difficult it would have been, thirty years ago, to explain what a modern physician now takes for granted:

• a patient sends a message from home;
• the physician sees laboratory results obtained somewhere else;
• the medication travels electronically to a pharmacy;
• the patient reads the physician’s note;
• a computer listens to the visit;
• another computer summarizes twenty years of records;
• an AI assistant answers questions inside the patient portal;
• software prepares clinical work for review.

We call all of this normal remarkably quickly. That should make us modest about predicting what another thirty years will bring. It should also make us attentive to what must not change.

The Patient Is Still the Point #

The central encounter in medicine is remarkably stable. A person is worried. Another person has been trained to help. Everything else is infrastructure around that relationship: the textbook, the teacher, the paper chart, the laboratory slip, the X-ray, the prescription pad, the computer terminal, the electronic health record, the smartphone, the neural network, the agent.

Every generation becomes fascinated by the newest piece of infrastructure and risks confusing it with medicine itself. It is not. The purpose of the infrastructure is to make the encounter better. That is the test, and it can be asked in plain terms:

• Does the technology preserve the truth of the record, including across every handoff?
• Does it help the physician understand, rather than merely see more?
• Does it make care more accessible, rather than creating a lower tier of automated medicine for people with fewer resources?
• Does it give clinicians attention back?
• Does it improve follow-through?
• Does it expose uncertainty, and show its sources?
• Does it know when to stop?
• Does it allow a human being to remain responsible where responsibility matters?

Those questions matter more than whether an AI can outperform a physician on an examination.

Let the Machine Carry What It Can #

I began studying medicine in an era when the physician’s mind carried an enormous amount of the system inside it. I practice today in an era when the system carries an enormous amount of information but demands an enormous amount of the physician’s attention in return. Neither arrangement is ideal. The next era can be better. Let the computer carry what it can:

• let it remember everything it can, and search faster than I can;
• let it reconstruct the timeline, compare the laboratory values, and reconcile the lists;
• let it watch the unfinished tasks, and find the forgotten result;
• let it hold the thread across every handoff;
• let it prepare the paperwork, check the coverage, and fight with the authorization process;
• let it remind the patient, follow the referral, and prepare the note;
• let it show me what I may have missed, and tell me exactly where the evidence came from;
• let it keep working when the clinic becomes busy, and perform the repetition that humans do badly precisely because it is repetitive.

But do not mistake any of those abilities for the whole of medicine. When a frightened patient looks across the room and asks, “Doctor, what do you think I should do?”, the purpose of all that computing power should be to give the physician more evidence, more clarity, and above all more attention with which to answer.

That is the measure of success. It is not whether the AI can generate a note, or whether a health system can schedule three more appointments. It is not whether an insurer can process another thousand authorizations, or whether the machine can imitate the language of a physician.

The question is whether technology can make the encounter more human after thirty years in which technology often made it less so. If it does, the achievement worth wanting will not be that a machine learned to imitate a physician. It will be that medicine learned, after three decades of digitization, which work never required a physician in the first place.

Return to Khadur Sahib #

That brings me back to where this story began. The traditions associated with Khadur Sahib now seem to me almost uncannily suited to the questions confronting medicine. Each speaks to one of them:

• Succession asks whether what one person knew can pass faithfully to the next.
• ਗੁਰਮੁਖੀ, the written word, asks whether we preserve the record faithfully.
• ਲੰਗਰ, the shared kitchen, asks who shares in the abundance.
• ਅਖਾੜਾ, the discipline of the body, reminds us that the person is not reducible to abstraction.

And Guru Angad’s verse warns against confusing light with understanding. ਜੇ ਸਉ ਚੰਦਾ ਉਗਵਹਿ ਸੂਰਜ ਚੜਹਿ ਹਜਾਰ ॥ A hundred moons. A thousand suns. Imagine them now as data points:

• laboratory values, radiology images, and genomic sequences;
• home blood-pressure readings and continuous glucose measurements;
• prescription histories and hospital records;
• portal messages and wearable devices;
• clinical notes, risk scores, and population models.

Millions upon millions of observations, illuminating the patient from every direction. ਏਤੇ ਚਾਨਣ ਹੋਦਿਆਂ ਗੁਰ ਬਿਨੁ ਘੋਰ ਅੰਧਾਰ ॥ And still, without the capacity to make meaning from them, darkness.

The modern medical record is approaching a thousand suns of information. Artificial intelligence may become an extraordinarily powerful instrument for organizing that light. But organization is not wisdom. Prediction is not judgment. Information is not understanding. And understanding a disease is still not the same as understanding the person who has it.

Thirty years after I entered pre-medical studies at Khalsa College, that is the future of artificial intelligence in medicine that interests me. It is not the disappearance of physicians, nor the autonomous clinic, nor medicine surrendered to an algorithm. It is something at once less dramatic and far more important:

• a physician who no longer has to carry the entire information system in his mind;
• an electronic record that no longer demands constant human operation;
• machines that absorb the repetition, retrieval, tracking, and clerical friction they are increasingly able to handle;
• and a patient who once again receives the most valuable thing technology can return to medicine: the undivided attention of another human being.

That human being understands the evidence and its limits, understands the person in front of him, and is willing to take responsibility for what happens next. Perhaps that is the real primary care office after the EHR.

The record will still be there. The algorithms will be everywhere. The intelligence behind the encounter may exceed anything earlier generations could have imagined. But the computer will have moved quietly into the background. And in the foreground, where medicine began, there will still be a patient and a physician, talking to one another.

What This Means for Patients Now #

That future is not fully here, but patients are already meeting pieces of it. A few practical principles are worth knowing. They are general information, not legal or medical advice for any individual situation.

Ask about ambient recording. If an office wants to use ambient AI during your visit, it is reasonable to ask what is being recorded, why, how it is stored and for how long, and whether participation is optional.

Know your right to a human in California. If you receive an AI-generated clinical communication that was not reviewed by an appropriately licensed or certified health care professional, state law generally requires a disclosure and instructions for reaching a human. If something in a message worries or confuses you, use them.

Read your own record. Most patients can now read their notes and results through a patient portal. If a note states something important that is wrong, report it early, because errors in electronic records tend to reproduce themselves.

Keep your own medication list. Include nonprescription drugs and supplements, and bring it to every visit. Artificial intelligence can reason only from the information available to it. A beautifully designed system working from bad input is simply a beautifully designed way of being wrong.

Understand consumer chatbots. Information you paste into a general-purpose AI tool at home is generally governed by that company’s terms and privacy policy, not by the rules that protect your records inside your doctor’s office. Decide deliberately what you share.

Bring the questions, not just the answers. Patients who use AI to research their health should not be embarrassed about it. Bring what you learned. Ask what the AI got right and what it got wrong. And ask your physician the same question physicians should increasingly ask their own systems: How do we know that?

Where This Series Goes Next #

This essay is the trunk of a series. Each of the branches below is a question this piece opened and could not close. They will be published separately on personal.kpsgill.com and linked back here.

• A Short History of Writing Things Down. From the Hippocratic case histories to Mayo’s single dossier, Weed’s problem list, and the first medical AI that no one dared to use.
• Pajama Time. What the electronic record did to the physician’s evening, and why the note grew longer as it became less readable.
• The Handoff. Why medicine loses information where it is passed on rather than where it is made, from Khadur Sahib’s succession to the 7 a.m. sign-out and the hospital discharge.
• Show Me Why. Provenance as the first principle of clinical AI: why the most important button is not “Accept.”
• Software Does Not Testify. Who answers when a chain of machines and humans harms a patient, and how AB 3030, SB 1120, and the AMA’s 2026 policy begin to answer.
• Langar and the Algorithm. Equity as architecture, and the two futures the same software can build.
• Ten Languages, One Clinic. Language access in the Central Valley, and why translation is where AI errors are hardest to see.
• What the Chart Cannot See. A field guide to the information that never makes it into a database: the gait, the pause, the daughter who looks away.
• The Efficiency Trap. What health systems should measure instead of visits per hour.
• The Patient Will Have AI Too. A guide for patients on chatbots, portals, and the question to bring to every visit.
• Where Knowledge Lived: Amritsar, 1995. A memoir chapter on learning before the internet, and what scarcity taught.

Evidence and Editorial Note #

This essay deliberately combines personal memoir, historical reflection, documented current technology, and forward-looking analysis. To keep those categories distinct, its claims are graded in four tiers.

Documented findings.

• A 2026 Journal of General Internal Medicine scoping review of primary care AI screened 3,203 manuscripts and included 61 studies: 26 on model development, 8 reporting clinical-trial results, and 11 concerning real-world implementation. The authors characterized primary care AI as still largely developmental.
• California AB 3030, effective January 1, 2025, requires specified disclosures and a route to a human for certain generative-AI patient clinical communications, unless an appropriately licensed or certified human health care professional has reviewed them.
• California SB 1120 regulates health plan and insurer use of AI and similar tools in utilization review. It preserves qualified human decision-making for medical-necessity determinations that deny, delay, or modify care.
• CMS-0057-F’s 72-hour expedited and seven-calendar-day standard prior-authorization timeframes apply to covered medical items and services within the rule’s scope, and exclude drugs.
• The American Medical Association’s 2026 policy direction emphasizes physician oversight, transparency, auditable evidence of safety and efficacy, education, and the principle that AI should support rather than replace physician judgment. It also holds that responsibility for AI-related harm should rest with the party best positioned to understand and reduce the risk.
• The epigraph appears on Ang 463 of Sri Guru Granth Sahib under the Second Mehl.

Vendor representations. Descriptions of Epic’s Art, Emmie, Ergo Visit (including Epic’s August 2026 announcement of Ochsner Health as its first user), and Agent Factory reflect Epic’s public information. They are product descriptions, not independent verification of clinical effectiveness.

Author’s recollection and tradition. The author’s educational and professional history is presented as personal experience. The account of Khadur Sahib, Guru Angad Dev, Mata Khivi, langar, Gurmukhi, and the akhara reflects Sikh historical tradition and standard reference sources. The connections drawn between those traditions and medicine are the author’s reflection.

Analysis and forecast. Future clinical workflows are the author’s analysis as of September 21, 2026, not descriptions of established clinical practice. These include the illustrative 2030 encounter, persistent agents, exception-based inbox and refill management, AI-supported handoffs and language access, and conversational interaction with the EHR. Current peer-reviewed literature does not establish that primary care has already become “AI-first.”

Independence. This article is not sponsored by Epic Systems Corporation, Microsoft, any artificial-intelligence developer, electronic-health-record company, insurer, pharmaceutical manufacturer, or health system. Mention of a company or product is for analysis and does not constitute endorsement.

About the Author #

Kanwar Partap Singh Gill, MD, traces his family roots to Khadur Sahib in Punjab’s Majha region and grew up in Amritsar. He attended Spring Dale School and completed pre-medical studies at Khalsa College in 1994–1996. He earned his MBBS from Sri Guru Ram Das Institute of Medical Sciences and Research in Amritsar in 2002.

He completed PGY-1 at the University of Arkansas for Medical Sciences in Little Rock in 2007–2008, and PGY-2 and PGY-3 with UCSF Fresno at Community Regional Medical Center from 2008 through 2010. He has practiced medicine in California since 2010. His career has crossed four distinct information eras: predominantly paper-based medicine, early and fragmented electronic records, the integrated-EHR era, and the emerging age of generative and agentic artificial intelligence.