A nervous energy hung over the airfield at Le Bourget. It was just after ten in the evening, and a crowd of nearly 100,000 was gathered to witness history. A record-setting flight was en route from America. But as darkness settled over Paris, things were not going as planned.

Thirty-three hours earlier, a small, single-engine airplane had departed from Long Island headed east over the Atlantic. Heavy-laden with 450 gallons of fuel needed for the 3,600-mile journey, the custom-modified craft had strained against its weight on takeoff, barely clearing the telephone wires at the end of the runway. Once airborne, news had steadily come through of the flight’s progress. There were sightings over Boston, Newfoundland and then England, and the latest rumor racing through the crowd was that the plane was spotted over Cherbourg. But nothing could be certain, except that the Spirit of St. Louis was three hours overdue.

With the crowds preoccupied by pessimism, Edwin James, chief European correspondent for the New York Times, was taking in the moment. Le Bourget was the only airport in the city, and a brisk wind swept across its expanse, chilling the spectators who had resorted to stamping their feet to stay warm. The airfield itself was well lit by enormous arc lights, unmistakable to any experienced aviator, and it gave James a clear view of the scene. Nearby, he spotted Myron Herrick, the U.S. Ambassador to France, waiting to greet his fellow countryman upon arrival. Flanking him were two companies of French troops - bayonets fixed to their rifles - who had been brought to keep order throughout the spectacle.

The public’s interest in aviation had increased substantially in the past few years, and hopes were high for a successful flight tonight. But in the year 1927, the dangers of aviation were well known to both the press and the public. Just weeks earlier, two of France’s most talented aviators - departing from this very site - had disappeared in their attempt to reach Manhattan. The pain of that loss was still fresh in the minds of the crowd. As the minutes and feet-stamping ticked along, the fate of this flight remained uncertain.

Ignorant of these terrestrial concerns, Charles Lindbergh returned to his instruments, a schooled habit ingrained in him through thousands of hours in the cockpit. As the lights of Paris came into view, he was relieved to see that his engine was running perfectly. His meticulous preparations - calculations, modifications, testing - had not been in vain. He spotted the Eiffel Tower below, circled it once, and then proceeded northeast towards Le Bourget.

Lindbergh grew up in Minnesota the son of a politician, but after dropping out of college, he decided to take to the skies. At the time, there were only a few ways to make a living as an aviator, and Charles pursued them all. In 1923, he bought a surplus Curtiss Jenny and began traveling the country as a ‘barnstormer’ doing paid demonstration flights for spectators. This was followed shortly after by more flight hours gained as a member of the Army Air Service Reserve, a predecessor to today’s Air Force.

But Lindbergh’s career would accelerate fastest when he took his next job as lead pilot for the Robertson Aircraft Corporation in St. Louis, Missouri. Robertson was among the first of a new type of company in America - the ‘airline’ - that had recently arisen to meet a growing economic demand for commercial aviation. And in this new organization, Lindbergh’s role as a pilot was simple:

Fly the mail.

Just two years earlier in 1925, the U.S. government had begun contracting with private companies to provide air mail services throughout the nation, and what followed was nothing short of an aviation renaissance in America. Thousands of airplanes now traversed the skies daily, flying the mail across all corners of the United States. Air mail pilots like Lindbergh had become the most experienced aviators in the world, and just as importantly, the planes used to fly the mail and the infrastructure supporting them had gotten much, much better. Air mail had become far more than a government service - it was catalyzing the development of an entire new industry, one widely recognized as critical to the nation. So important was flying the mail to the future of aviation that Lindbergh viewed it not as an occupation but as his duty to both the country and the world.

Just minutes away from beginning his final approach, Lindbergh pulled out a flashlight and inspected his fuel gauges for a final time. He had taken extreme measures to ensure the Spirit of St. Louis was as light as possible to make room for gas, even going so far as to tear off pieces of excess paper from the notebook he carried in the cockpit. But even when every ounce mattered, he had still made room for a critical piece of cargo traveling with him this night - five pieces of specially marked air mail.

A Problem of Lively Concern

Many clearly promising technologies struggle to gain a foothold, and aviation was no exception. In the early 20th century, there was widespread recognition that a prosperous commercial aviation industry was a transformational capability for any nation, and one that would bring widespread benefits. But the industry was having trouble making progress.

The reality was that aviation didn’t yet have strong economic footing. At the most basic level, the state of aviation technology was a major factor. Airplanes of the nineteen teens and twenties were still rudimentary by any modern standard, and they were expensive to operate. Moreover, they were slow and inefficient, and improvements were capital-intensive. Airplanes were surely useful, but air services were very expensive, and demand was limited. For investors looking to put their money towards innovative projects, aviation appeared to be a money-losing proposition.

The infrastructure supporting commercial aviation was also underdeveloped. With only basic communication and navigation tools, planes could only fly in specific conditions across certain terrain. This made air travel unreliable - a week of bad weather could delay air shipments in a way that other forms of transportation like trains or automobiles were not susceptible to. Moreover, airplanes required places to land and take off - and there was a relative paucity of airfields across the country, limiting the available routes. Even where the technology itself was capable, structural issues prevented expansion.

Finally, standards and regulations were immature, which kept trust low in aviation. The absence of agreed-upon safety criteria meant that airplane crashes were common, and aviation was viewed by many potential passengers as too unsafe. This not only limited demand for air travel, but disincentivized more people from becoming pilots at all - hampering development of a skilled aviation workforce. The relatively high accident rates also created a boring but highly pragmatic problem - it was very hard to get someone to insure an airplane against loss. If an airline was uninsurable, investors were unwilling to put money into the new venture, further stifling innovation.

The gap between the aviation the world did have and the aviation the world could have was clearly visible. At the highest levels of government, action was demanded. “Aviation is a problem of lively concern to every nation,” said President Calvin Coolidge. “Having given the world the practical knowledge of flight, a serious responsibility rests upon us to see that air transport be developed along constructive economic lines.”

Flying the Mail

Members of the government knew that without a business model and investments, aviation would not reach its ultimate potential, and so policymakers began to take action to accelerate aviation’s development. Officials soon found an important lever to pull on - air mail services. The Post Office, as a major consumer of transportation services, had a long history of investing in sending letters and packages by ships and trains, and air mail was a natural extension of this tradition. In fact, air mail was already a service - the Post Office employed its own small corps of pilots. However, members of Congress and the Department of Commerce saw opportunity to use air mail for the benefit of the aviation industry as a whole. In 1925, Congress passed the Contract Air Mail Act allowing the Post Office to pay private companies to fly the mail.

Nearly overnight, the creation of the air mail contracts ushered in major investments in aviation. The Post Office set payments at $3 per pound for the first 1,000 miles, and these rates promoted huge interest in the opportunity. Entire new companies were formed to organize and bid for air mail routes, and these new entities - the first airlines - included major industry leaders like Henry Ford. Bidding was fierce and competitive, with companies jockeying with each other to showcase their airplanes’ performance and offer lower rates. New airplanes were manufactured for the sole purpose of flying the mail, while new technological innovations began making their way into aircraft, all propelled forward by the air mail contracts.

As air mail became the lifeblood of the growing aviation industry, further investments in infrastructure and safety followed. Historically, the government had invested in navigation aids for commercial shipping, and this same precedent was applied to air mail. A nationwide network of lighted beacons soon was installed along air mail routes to guide pilots at night, while new radio-based technologies were developed to allow planes to fly in bad weather. Furthermore, to promote safety, the Post Office - as a paying customer for air mail services - demanded that airplanes and pilots receive a certification of “air worthiness” in order to fly the mail. Not only did this process raise standards for reliability, it simultaneously allowed for airlines to be insurable - a critical aspect for securing investments.

Importantly, the air mail contracts were not viewed as subsidies - the Post Office was a paying customer, bidding was competitive, and air mail was an important and useful service to the nation. The air mail contracts were not a windfall, either - many airlines lost money. But they provided enough support to carry the industry forward.

The first contract air mail route was flown in April 1926 by Varney Air Lines. By 1931, over 85% of airline revenue came from flying the mail. And by 1933, with the release of the Boeing 247, technology had advanced to the point that modern passenger transportation was beginning. Varney would later merge to become part of a new company - United Airlines - and Boeing would go on to become a global leader in airplane manufacturing.

A New Ocean to Cross

A century after the dawn of aviation, we find ourselves in the early days of a similarly transformative technology. The vision of medical AI enabling global access to expert-level medical care is clear and widely held. Aviation had an analogous vision - widespread access to global passenger air travel. But in 1926, the realities of commercial aviation were still far too distant from their promise. Similarly, in 2026, the gap between what medical AI can do for the health of the world and what it is technically and legally capable of doing today is vast.

Medical AI suffers from the same difficult economics as the early aviation industry - it is imperfect and without a viable business model. As a category, general AI clearly is self-sustaining - it is one of the most invested-in technologies ever. However, medical AI which actually delivers care is not widely reimbursed today, and future services like autonomous diagnosis and management have no standard payment pathway. Functionally, this means that investing in building medical AI is highly speculative - just like in early aviation where investors were unwilling to put capital into activities like airplane manufacturing with no viable path to a return. The irony is palpable. Inside a technology category that has trillions of dollars pouring into it, the budding medical AI industry - with all its potential benefits to the health of the world - sits under-developed.

The second order effects of this are also simple to articulate - the know-how and infrastructure upon which to build autonomous medical care don’t get developed or built. After air mail contracts were established, air mail pilots quickly became the best pilots in the world, and their skills and knowledge developed a locus of expertise necessary for subsequent innovation and expansion. Similarly, an infrastructure was put in place to support air mail pilots - nationwide navigation aids, improved airfields, and other advancements - to ensure safe and effective operations. In medical AI, this same type of know-how and infrastructure is waiting to be built, just in a different form - better data, more context, improved interfaces, and a cadre of clinicians and engineers who know how to work with these new tools.

Opening New Routes

The most important lesson contained within the history of air mail is that it worked. The air mail contracts created the basis for an immediate and large increase in investment in aviation, quickly led to major improvements in airplane technology and nationwide aviation infrastructure, and promoted foundational safety regulations. All this led to the achievement of the major underlying objective - in less than ten years, viable passenger air travel had become possible, a new industry had been firmly established, and America was at the forefront. This was not a happy accident, but the direct result of intentional policies undertaken by a visionary cohort of leaders within the U.S. government and actively supported by industry. It is one of the great success stories of American innovation and leadership.

With medical AI, we have a new opportunity for intentional policymaking. We can bring a new industry forward faster than it would on its own, and realize the benefits of medical AI sooner than we would in the absence of action. Just as the humble air mail contracts accelerated development of the aviation industry by decades and positioned the United States to become the global leader in aviation technology, the same set of conditions exist today in the new field of medical AI. With the world looking for its leader in this technology, America should once again step forward.

The easiest mechanism to get started is to begin paying directly for medical AI services - healthcare’s version of the air mail contracts. Like it was with the Post Office, there is precedent for government as a payer. Today, over a third of Americans receive their health insurance through Medicare and Medicaid, and CMS can use this reach to begin reimbursing for medical AI - providing the industry with the types of predictable payments for valuable services that will promote investment and increase both capability and quality of services. Just like with air mail, payments should not be designed as a subsidy, but as fair reimbursement for delivering services that have value. There are many types of decisions, supportive services, actions and monitoring programs that clearly benefit patients and for which CMS - or any other national payer - could provide discrete payments.

Infrastructure for medical AI is the next component that will need to be built out, and institutions also exist to support this work. The Office of the National Coordinator for Health IT plays a major role in shaping national policy around interoperability - the key infrastructure layer needed for effective medical AI - and that influence can be extended to the specific needs of medical AI. For more experimental or high-risk infrastructure projects - especially attempts to design new technologies - government agencies like ARPA-H already serve as a funding source. Although larger infrastructure builds may need congressional appropriations, the good news is that the foundations of interoperability laid down in the last decade - HITECH, FHIR, TEFCA - provide a strong starting point in capabilities rather than requiring a decades-long buildout from scratch.

Finally, medical AI needs its own version of “air worthiness” - the criteria upon which early deployments can be approved. Such a “care worthiness” standard should be rigorous enough that it provides meaningful reassurances for safety, while not being so onerous as to prevent anyone from flying at all. The criteria must also be clear so that compliance with appropriate regulations is not in doubt among good faith actors.

Unlike early aviation, medicine already has many of the institutions available to create the kind of environment within which medical AI can flourish. With intentional policy, the nation would see rapid and dramatic investments and improvements in technology, and the benefits of better care, more access and lower cost can be realized far more quickly. More broadly, it would position the country for ongoing pre-eminence as a global supplier of the advanced medical AI technologies that are poised to become the defining healthcare infrastructure of this century.

New York to Paris

In the years after his flight across the Atlantic, Lindbergh would bear witness to a pace of change that defied even his own extravagant visions. By 1953 - the year he published his autobiography - passengers were flying billions of miles without incident, transcontinental travel was the norm, and the beginning of the jet age was bringing even more impressive capabilities into the skies. “We actually live, today, in our dreams of yesterday…”, he reflected.

The visionaries behind the Air Mail Act in 1925 believed in a future where aviation could reliably reach all parts of the globe, and that such a future would be broadly beneficial to society. Although they could not predict the specific pace of change or the enabling innovations that would follow, they were committed to the underlying principles that real innovation follows real economics.

Today, we have a new vision - of a world with immediate access to expert medical care made possible with AI. With investment and time, that world can become yesterday’s dream - one replaced by a new reality where medical AI is delivering care to billions.

Acknowledgment: I am grateful to Dr. Bob van der Linden of the Smithsonian Institution for his generosity in sharing his expertise for this piece.

The world will soon consider certain forms of medical AI as a similarly essential component of healthcare infrastructure. Just as essential medicines form the basis of treatment for a broad swath of human disease, essential medical AI will form the basis of clinical judgment across that same spectrum of illness. It will be an expectation that the capabilities of medical AI are widely available, and significant resources will be dedicated to ensuring access. But as governments and other seek to invest in this new modality, they will need guidance on where to start - what are the first investments in developing and deploying these newly essential technologies?

A key point is that medical AI capabilities are not monolithic, but are embedded in a series of decomposable skills. These skills are all familiar - history taking, data interpretation, differential diagnosis, management planning, counseling and anticipatory guidance. But performance of each of these skills in the real world will depend on the context available, and performance will differ between countries and communities. Although the skills will ultimately work together, each of them will require dedicated resourcing to build - and the skills may not all be deployable at the same pace in any given setting. Of course, the ultimate goal is to establish all of the essential skills globally as part of a high-functioning health system, and to keep them updated and maintained as medical science evolves. But this will be a step-wise deployment.

The full list of essential medical AI skills will be expansive and cover generalist and specialist domains as well as specific skills and sub-skills. Each skill will also require a minimum information environment to be usable, and performance of skills will need to be validated at the local level. Unlike drugs - which work mostly the same across communities - medical AI will perform very differently based on the deployment setting and require extensive localization efforts.

Building the list of essential medical AI skills will be a large process and require international coordination. The table below presents the foundational framework that can be applied to both generalist and specialist care as these efforts begin in earnest.

Table 1: A Foundational Framework for Worldwide Essential Medical AI Skills

In March of 1942, Anne Miller was dying. An infection raged through her bloodstream, and despite the best efforts of her doctors at Yale for nearly a month, she continued to worsen. For Dr. John Bumstead, Anne’s attending physician, there was only one last hope available.

Through a seemingly impossible set of circumstances, he had managed to acquire a few grams of an experimental substance - penicillin. The minuscule amount he was given represented almost half of the entire supply in the United States at the time - it was so rare that it was delivered to the hospital under police escort. When it arrived, the dose was prepared and administered to Anne while the chaplain said grace.

In the morning, as Dr. Bumstead entered Anne’s hospital room, the prayers of the scientists and the clergymen had both been answered. The effect was immediate and dramatic - her fevers, which had reached as high as 107 degrees, quickly subsided. Soon, she was able to take meals and converse. Eventually, a few weeks later, she would leave the hospital.

Anne’s miraculous recovery moved penicillin out of ‘experimental’ and firmly into ‘transformational’ territory, compelling action to bring the first antibiotic to scale through accelerated industrial production. Dr. Bumstead would continue to practice as an internist, and Yale University would come to play a major role in the expansion of penicillin. Anne went on to live another 50 years.

In the history of medicine, the age before antibiotics seems like a story from a darker era, on the same timeline with leeches and battlefield amputations. But Anne Miller is a contemporary - she only passed away in 1999, at the age of 90. We are not so far removed from a world without the modern cures we take for granted.

A Decade of Delay

Anne’s story is uplifting, but the story of penicillin also contains a cautionary tale. The truth is that penicillin had become known to the world over a decade earlier. Alexander Fleming published his original breakthrough discovery in 1929 - but lack of interest and perceived urgency had caused his findings to sit largely unexamined. It was not until a decade later, in 1939, that researchers at Oxford developed a vision for penicillin as a cure for bacterial infections and aggressively pursued its purification for medical use. The decade of delay between discovery and commitment was a time when millions of people died from infections that could have theoretically been prevented. In a different world, Anne Miller might have been one of them.

With medical AI, we now find ourselves in an analogous situation confronting the implications of a new discovery. But we are not Fleming staring inquisitively into the petri dish. We are Dr. Bumstead, staring in disbelief at what now seems possible.

For all of medical history, clinical judgment has required a human mind. That constraint has now been lifted. Medical AI is now routinely performing at expert levels on clinical reasoning evaluations, and real-world studies are showing serious improvements in quality and outcomes. This is not better software - it is a fundamentally new class of cognitive systems capable of exercising judgment at the highest levels. Most US physicians now use AI in daily practice. The field’s leading voices are urging patients to embrace it. And there is no sign that progress on capabilities is slowing down. Discovery is no longer the bottleneck - it is translation. And this forces us to ask a harder question:

Now what?

A New World In Sight

The world doesn’t have enough doctors, and that fact is responsible for enormous amounts of human suffering. Both at home and abroad, billions of people lack regular access to the skilled labor of the physician workforce. For millions of them, it is the proximate cause of their disability and death.

Doctors are much more than technicians, but their technical skills remain a scarce and valuable resource. One of the most fundamental of these skills is clinical judgment - the ability to make high-quality expert medical decisions. Accurate diagnosis and management are the basis upon which all other medical interventions are deployed. Judgment is upstream of action; without it, care proceeds haphazardly, if at all.

Today, most of the world simply goes without the benefits of good clinical judgment. Globally, we need 6 million additional doctors to meet minimum population requirements, meaning we only have about two-thirds of the doctors required. Studies suggest that lack of high-quality medical services causes more than 8 million deaths a year worldwide, alongside untold amounts of misery. Even in the United States, 100M people lack a usual source of care.

Recent advances in artificial intelligence demonstrate that this status quo can now be challenged. In just a few years, medical AI has progressed from borderline scores on multiple choice exams to expert-level performance across a wide variety of synthetic and real-world clinical scenarios. Taken individually, these studies are narrow and limited. But taken collectively, they represent a body of evidence that is now strong enough to act on with urgency. A different world is now clearly in sight - one where everyone on Earth has access to high-quality clinical judgment in every geography, all the time. The net impact would be millions of lives saved worldwide, and hundreds of millions more unburdened by the scourge of treatable illness.

The worldwide expansion of medical AI would be a generation-defining public health triumph, but decisions made in the coming months and years will determine whether that triumph occurs in this decade or the next. Medicine is now facing a familiar question in bringing forward this new innovation - to choose speed, or accept delay.

And we should choose speed.

The Benefits of Speed

In medicine, “First, do no harm” is a ubiquitous moral refrain, but the harms of inaction are explored less often. We have come to accept long innovation cycles - that the decade of delay is simply the price we must pay for scientific advancement. But that tradeoff is neither morally neutral nor correct.

Imagine knowing about the curative powers of penicillin and withholding it from the world: such an act would be unthinkable. And it was - world leaders acted with urgency to mass produce penicillin during World War II. This was a huge challenge requiring international coordination and significant resourcing, but it was tackled head on - and ultimately succeeded. In 1945, Fleming and his collaborators were awarded the Nobel Prize in Medicine, and millions of soldiers were returning home thanks to these collective efforts.

Even in our own lifetime, we have experienced speed as a choice, not an accident. When the world was gripped by the Covid-19 pandemic, leaders in the medical community called for rapid action, arguing that the cost of waiting was too severe. They were right, and millions of people are alive today because of those decisions.

With medical AI, the world will look back at this point in time and ask what choices we made at this pivotal juncture. Only a few years ago, that choice was murky - the technology wasn’t ready. But now, the accumulated body of evidence supports a very clear choice - speed. That does not mean the path is cleared - It will be hard to bring medical AI forward, just as it was hard to create a world full of antibiotics or a world liberated from a terrible virus. But billions of people are depending on us to make the right choice and march down the harder road with expedience.

Moral urgency is not an argument for recklessness - innovation, speed, and responsibility are compatible. But they must be chosen. Penicillin underwent clinical trial evaluation and was validated by health authorities before industrial scaling - it just happened on the order of a few years. We should not abandon rigor, but we must calibrate that rigor to the opportunity we have before us.

Our Penicillin Moment

We know the harms caused by lack of access to expert clinical judgment. And we know the advantages of speed when translating technology into practice - we have done the math, in public, in our lifetimes, and made the right choice before.

With Fleming and penicillin, the world can be excused for not knowing exactly what to do during the decade of delay. But for us, with the benefits of both hindsight and experience, how can we excuse ourselves?

We now find ourselves standing at Anne Miller’s bedside. The fever has broken, and a new world is possible. We can bring medical AI to scale with urgency and speed, within a decade, for the benefit of billions of people.

Cognition was previously a scarce resource - now it is abundant.

Hesitation was previously neutral - now it is harmful.

This is our penicillin moment. The path forward is clear.

Will we choose it?

Anne Miller with Sir Alexander Fleming (L) and Francis Blake (R), Dean of the Yale School of Medicine. Image Credit

But medicine is complex. Misdiagnosis, inadequate treatment and insufficient systems all conspire against perfection in our quest to ameliorate human suffering. It is for this reason that our performance expectations in medicine are not framed against an impossible bar of flawlessness, but against the high but achievable bar of the “standard of care” - often defined as the degree of care a prudent, reasonable professional would exercise. Although within this definition there is variability, there are many places where the line is clear. A doctor who ignores a life-threatening lab abnormality would obviously violate the standard. But what about the doctor who misdiagnosed a chronic cough as post-nasal drip when it was actually a rare lung cancer? It is cases like this that ask us to summon the standard of care - to ask whether this was a mistake at all, and if so, how bad?

Medicine tolerates imperfection, but that imperfection exists within a defined standard of care. A situation where a 1-in-a-million disease was misdiagnosed as a 1-in-10 common complaint is unfortunate - but it is not inherently negligent. So long as the standard of care was followed, the profession and society accept that mistakes, delays and other bad outcomes are inevitable even when people and systems are performing at their highest levels.

Medical AI Standards of Care

Comparing human beings to each other is an elegant way to create a professional standard - especially in the competitive, high-expectation environment of medicine. But 1:1 comparison with human beings does not hold in medical AI. Patients will interact with AI systems in different ways, AI systems will occupy a different part of the healthcare ecosystem than human actors, and both the success and failure modes of AI systems will be distinct from human beings. We cannot ask “what degree of care would a prudent, reasonable professional exercise” because AI can simultaneously do both more and less than any prudent and reasonable professional. A new standard is needed specific to AI.

The AI Standards of Care separate diagnosis and management explicitly because these functions require fundamentally different standards. Diagnosis is about getting the answer right - does the patient have the disease or not? Accuracy is the indicator of performance, and there are many reasonable paths to generate accurate diagnoses - in the form of deeper history, more testing, and additional consultations. Rather than pre-specify the process, the standard seeks a pre-specified level of outcome.

Management decisions are more about the integration of knowledge and context to make a well-reasoned decision. Unlike diagnosis, management is less about pure accuracy and more more about overall logic - the art of navigating tradeoffs between risks and benefits and translating the nuances and ambiguity of clinical evidence into practice. Being “right” is less important than being “defensible” - that a treatment decision was based on a solid foundation, even if honorable people might disagree. It is for this reason that the AI Standards of Care in management look at the depth of knowledge and overall patient context that was applied to a particular decision. It is not the expectation that any given professional knows every journal article in their field or every minute detail of their patient - but they should reliably take into account the important stuff, and they should be logically consistent in how it is applied.

Notably, the standard of care is not just a professional norm - it is a formal legal concept embedded in state and federal statute. As medical AI evolves, the law itself will need to evolve with it to embed these definitions alongside updates to our practice patterns and technology assessments.

Setting the Standard

Having defined the standard, we will need to set it and measure the performance of medical AI against it - but this is harder than it seems at first glance. The idea of comparing medical AI performance to the baseline performance of a human system is conceptually strong and has been advanced by others, but it hits an immediate limitation - the availability of human performance data upon which to base the comparison. The reality is that most care processes today lack strong instrumentation. We often do not know how good - or not - the current system is at achieving its aims.

The industry will require a fundamentally new performance measurement and monitoring schema in order to operationalize the medical AI standard of care. This is more than just pragmatic - it is about acceptability of medical AI long-term. It is unrealistic to believe that widespread adoption of medical AI will proceed if that AI system is worse than the current system. Rather, systems using medical AI will be expected to be much better than existing human systems. This is the same human psychology that impacts all sorts of technology adoption. People don’t want to ride in self-driving cars that are a little better. They want to feel totally reassured that they are getting exceptional safety and quality. Builders of medical AI should validate this and design for it, not dismiss it.

The scope and scale of the new measurement system remains a larger governance and operational question. Today, medical AI performance measurement compared to humans takes place ad-hoc in the form of supervised deployments, quality improvement initiatives or formal clinical trials. Although this is necessary for early learning, centralized and coordinated standards-setting will become necessary, and those outputs will need to be recognized by a formal authority in order to give the standards authority and legitimacy. This is analogous to how the aviation industry manages complex, domain-specific performance needs - the FAA recognizes consensus engineering standards set by the aviation industry, then holds suppliers to those standards through distributed oversight.

Measuring What Matters

Bringing forward new standards of care will be technically and operationally challenging, but it will also force the profession to directly confront a harder question - what does ‘good enough’ look like in medicine? Today, the answer to this question is opaque and hidden behind our measurement limitations. But with medical AI, we will know exactly how we are doing - and need to decide if it is adequate.

Defining aspirational excellence is not the challenge - the real task is defining what ‘minimum acceptable’ looks like, because it is at the boundary between ‘unacceptable’ and ‘acceptable’ where design choices require the profession and society to accept a known rate of failure.

This tension is not unique to medicine - every high-stakes domain must make hard choices at this junction. For comparison, airplane parts are categorized in a similar fashion - the allowable rate of failure is defined in proportion to the consequence of the failure of the part, and then systems are designed accordingly. Even in flight-critical parts where failure would cause the aircraft to crash, there is a known and agreed on rate of failure. Although that rate is extremely small, it exists and can be designed against. If a plane crashes, everyone can look back and ask “was the standard met”? A crash - while obviously tragic - can then be accurately contextualized, and the humans working in that system can be evaluated on whether they met the standard, not whether they are perfect actors. Without a clear bar to pass, any failure can be interpreted as negligence because honorable people will disagree on the performance threshold.

Defining and measuring these new AI standards of care will ultimately be healthy for the profession - pushing us to bring forward our best thinking to improve the system. We will gain a deeper view into how we are really doing at delivering care, and then be offered the opportunity to do better. If we accept that opportunity, we will have a generational chance to raise the performance floor in medicine in a way that, until now, has been merely aspirational.

Think about the last time you went to the doctor’s office. Did it feel slick and modern? Of course not - it felt like a trip back in time to a world before iPhones and SaaS products, one where paper documents and manual workflows prevailed. Between the clipboard full of intake questions, the scourge of the old magazines, and the perpetual weirdness of the examination table, the clinic feels dated - stuck in a bygone era of medicine where capabilities and expectations were lower.

Clinics are more than just physical spaces. They are dynamic environments that have evolved over time, and there are hundreds of moving parts - idiosyncratic workflows built over years, technologies that are customized for the nuances of a particular practice setting, and a staff that have bespoke training and experience to operationalize it all. This is old build - the medical equivalent of a pre-war apartment building. Everything technically works, but it’s a patchwork of enhancements melded onto a foundation constructed for a different era.

New advancements in AI technology have now created opportunities to totally redesign the infrastructure and operations of the clinic. By integrating operational AI that can automate standard processes alongside medical AI that can automate aspects of care delivery itself, clinic efficiency and effectiveness can improve dramatically while creating more self-service options for patients and a better practice environment for providers.

However, renovating an old build clinic is a challenging undertaking, in no small part because the clinic has to keep functioning during renovations. And because of this inherent limitation, renovations will always carry a huge opportunity cost. When clinic operations are disrupted, patients can’t be seen, revenue drops, and the whole enterprise is put into jeopardy. This reality explains in part why change is so slow in healthcare - when any attempt at improvement comes with a large up front cost, the default response is to delay change as long as possible. “If it ain’t broke, don’t fix it.”

This is a serious opportunity to redesign the core workflows that support healthcare delivery nationally. But the structural barrier to renovating old build presents a major challenge to progress. Change is hard - even minor adjustments to workflows can drag on, and full-scale clinic transformation can take years - and still fail. Moreover, the capacity of clinics to change - and to absorb financial and operational strain during periods of renovation - is very limited in the lower margin, higher-burnout world of outpatient medicine. Clinics that try and move fast will find their staff unhappy and their balance sheets upside down.

The solution is new build. Clinics that start from scratch will be in a much better position to take advantage of new technologies because they don’t need to rip-and-replace anything, and new organizations can design from day 1 to deliver AI-native care the way they want. This takes a different kind of initial investment - in capital, vision and time - but the payoff is a fundamentally new type of clinic built for the modern age.

New build doesn’t mean immediate build. It will take years - even decades - for these new types of AI-native clinics to permeate the system. But this gradual transition in infrastructure is actually good on several levels. First, it allows new build clinics to expand at a more natural pace - tracking with a realistic cadence of patient acquisition and provider hiring. This lets new build clinics install AI-native operations gradually, iterate over time, and not feel as much pressure to scale improved but still sub-optimal systems out of the gate. Second, a gradual expansion of new build clinics lets the system adjust to new staffing expectations without labor market shocks. Rather than cutting staff (which no one wants to do), new build clinics can just hire more slowly and deliberately, and the system can adapt to new staffing structures over decades. And third, it ensures new build clinics adopt better technologies as they are released, not simply at a single point in time. A new build, AI-native clinic ten years from now will look very different than one today - and the structure of the model that will win out over time is not known. If all new build clinics were created over the course of this year, for instance, then a cohort of clinics would exist “stuck” in 2026 rather than having multiple cohorts over time that can compete for the winning model.

Installing AI-native care delivery across the healthcare system is a worthy goal, but renovating the existing infrastructure will be a slow and burdensome process that will ultimately lead to mediocre outcomes. By choosing to start fresh, we can seed the beginnings of a new build that will form the foundation for the next generation of clinics nationwide.

This work has a name: the bottom of license. These are the many things that technically require a professional license - but just barely. The bottom of license includes things like ordering screening labs and imaging, prescribing preventive treatments, straightforward medication titrations and refills, and basic diagnosis and management. In other words, the easy stuff. Medical AI presents a new opportunity to move these offerings into a self-service model. Rather than insisting that huge swaths of low-risk activities continue to require involvement of a human professional, we can just automate it - and we should.

Medicine has already tried everything short of giving up this layer of decision making. Clinics place “standing orders” allowing non-doctors to make decisions. Physicians delegate this work routinely to APPs and pharmacists who themselves follow protocols or, in some places, can simply practice independently. And direct-to-consumer testing and treatment offerings are often effectively a form of self-service - utilizing pre-built pathways to determine appropriateness. And yet, the formalities persist.

The most common arguments against this approach are safety-related - but that assumes that this cannot be done safely. Many of the activities encompassing the bottom of license have already been effectively “automated” in practice, and medical AI simply extends this into a pure self-service model.

Moreover, self-service does not need to achieve “zero risk” to be acceptable. Society allows individuals to self-service many activities with low but non-zero risk. Is self-prescribing a statin any riskier than taking out a loan to buy a car? Both can theoretically cause serious harm - albeit at low rates. We address these concerns with disclosures, guardrails and backstops - the loan terms are clear, a credit check ensures someone doesn’t get overwhelmed with debt, and the requirement for monthly payments provides ongoing monitoring of financial health. There are analogous mechanisms throughout medicine, and a new one as well - the AI itself. The point is that a human need not be involved to proceed. Self-service medicine does not mean “on your own.” It means designing systems that help people do the things they already know they need to do safely and effectively with less impediment.

Moving to a self-service model for bottom of license work is a natural early deployment for medical AI. It provides real value to patients and providers by removing friction that impairs health behaviors and clogs up clinics. Simultaneously, it creates a role for medical AI that allows developers, clinicians and patients to gain experience with the technology that informs higher-stakes use cases.

A license is a powerful tool, and things that require that license should matter. We can use medical AI to say farewell to the bottom of license and raise the floor of what requires physician capabilities in the era of cognitive systems.

This exact scenario plays out thousands of times a day across the country, and the response is the same - the doctors, nurses and staff leap into action. EKG leads are placed; oxygen is started; an IV goes in. We do not admire the problem - we do something about it.

When confronted with an urgent need, our medical system doesn’t wait - it acts. And this bias towards action is not merely an act of goodwill - it is a professional obligation known as the duty to care. When licensed professionals take on their role in society, they also take on a responsibility to act when circumstances demand it. The patient who collapses in the hallway gets immediate attention; the critical lab result gets reviewed; the frantic phone call from a patient gets returned.

Duty to care is an important accountability mechanism in medicine. Failure to take appropriate action - a violation of the duty - can be punished. Doctors who fail in their duty can face lawsuits or, at the extreme, have their license revoked. But these actions are uncommon. More often, the punishment is informal - in the form of consternation from peers and patients. Legal recourse is not a substitute for strong professional norms.

On the surface, the duty to care exists as a forcing function, but it is much more than that. Underlying the duty to care is a deeper purpose - it is a mechanism by which an entire professional workforce internalizes its obligations to society, and then reflects those obligations back behaviorally and structurally. From the earliest days of medical school, trainees are taught about the existence and importance of their duty. At first, this must be explicit, conveyed through reading and lecture. We then learn to recognize situations that trigger the duty - things like urgent and emergent presentations, or complex social and ethical situations. But most importantly, we learn that the duty compels us to do something. In medical school, this initial period of professional acculturation marks the transition from bystander to actor. Then, during residency training, the duty is reinforced through practice. The STAT page gets immediate attention; the patient struggling to breath gets seen first. At some point, the duty no longer needs to be referenced - it is simply understood.

AI systems in care delivery are new actors in the healthcare ecosystem, and their duty is currently undefined. If concerning symptoms are entered into an AI chat, what duty applies? If a grossly abnormal lab test is uploaded to AI, what should come next? If an individual reveals to AI a desire to self-harm, should someone intervene?

When patients tell their human doctors that they are in trouble, that doctor has a duty to respond - but does AI?

The AI Duty to Care

The profession needs a new framework for defining the duty to care across medical AI applications. Like in human care, that duty is not unlimited, and it must be different between care settings. Our expectations of a good Samaritan are different from our expectations of an emergency medicine attending - and we should apply a similar approach to AI.

In AI-Native Care Delivery, there are four pillars, each with their own level of duty: the AI Chat, the AI Consult, autonomous systems, and human care. As systems become more clinically-focused, the duty becomes more demanding.

In AI Chat - which is intended as a general health Q&A system - the duty is most similar to that of a medical assistant or nurse. Although a chat cannot make a diagnosis, it can recognize abnormal symptoms or findings, express appropriate concern, and recommend a next action - including proceeding to an AI Consult.

Within the AI Consult - the autonomous visit conducted by the AI - the duty begins to extend further. The AI Consult should perform at a high level in differentiating symptoms that require urgent or emergent care from those that can be seen routinely or managed at home. A reasonable site of care recommendation based on the information provided becomes the duty - with specific levels of accuracy as a separate performance standard that must be agreed upon later.

Autonomous care presents a more complex exploration of the duty, because by its very nature, the autonomous system should operate freely within its boundaries. However, the most important duty of the system is to reliably identify times when human judgment is required. The duty only ends when a situation is escalated and a human is brought into the loop. How this happens will be a specification of each autonomous system and is outside the scope of this piece. Suffice it to say, this is an area with significant unknowns.

The last pillar - human care - represents the most familiar version of the duty - to make a comprehensive assessment and plan, and to implement that with the patient. This is consistent with our existing conception of our duty, but with a key nuance - doctors will now need new standards for how we react to AI-generated escalations. Although these escalation pathways are familiar - akin to pages or portal messages - they will contain different information and require a new set of norms.

Finally, it should be noted that defining the duty sets a floor, not a ceiling. Systems should exceed minimum requirements wherever possible.

Action Versus Performance

A final distinction is worth noting - the duty to care is different from the standard of care. The duty compels us to take action, while the standard compels us to achieve a minimum level of performance. One is obligation; the other is competence. Although they work together, they serve different purposes. Medical AI systems will also need their own standards of care - and that will be the subject of future work.

Defining The Boundaries

It’s one in the morning, and someone tells an AI system they are in distress. Who is responsible for what happens next? Today, that responsibility is opaque and diffuse - but that is not a sustainable system.

By defining the AI duty to care clearly, we can constructively assess whether any given system is adequately meeting that duty, while avoiding the unrealistic expectation that any AI must be all things to all people. Moreover, we can move towards better explanations for our patients about the limitations of various AI systems, allowing for clearer understanding by the public of how to use this technology in their own lives.

Our professional obligations are not infinite - but they are defined. With medical AI, we can apply the obligations we already have internalized towards this new and powerful technology.

What does this mean in practice? Picture this - a new patient comes into clinic and has 1000 pages of medical records spanning 40 years and dozens of doctors in different cities. It is a lifetime of information.

How many of those 1000 pages should their doctor read before they can begin to make medical decisions for their patient?

This is a real choice that doctors are faced with every day, and knowing when you have enough information to start making decisions is not always obvious. When patients are complicated, reading a single note is never enough to get up to speed. But scouring 1000 pages of primary source documentation becomes practically untenable, with time and efficiency costs outweighing benefits to patients. Somewhere between 1 and 1000 pages, there is a stopping point - but where is it?

“Deciding when to decide” turns out to be a thorny issue without an agreed upon standard and high levels of variability. And like many things in medicine, the standard we do have is normative, not clearcut. Throughout training and into independent practice, we observe how our peers gather information, figure out what types and amount of data we need to make different types of decisions, and then go forth into the real world and apply our findings. Along the way, it becomes apparent that there is a floor - meeting the “standard of care” - and then a large grey area within which we work. For most medical decisions, the decision to decide is made with incomplete context and meaningful uncertainty.

We accept this normative framework in humans, but AI systems delivering care will require explicit standards for when they can act. We need a mechanism for encoding this meta-cognitive process into AI systems delivering care.

The Judgment Sufficiency Threshold

The Judgment Sufficiency Threshold is a new framework for defining the boundaries within which AI systems in care delivery can take action. It works by setting a minimum amount of information required to make a decision. Like human decision making, all available context is not a pre-requisite to act - but unlike human decision making, the threshold is explicit.

JST and “deciding to decide” is not something we go around thinking of every day, so let’s start with a concrete example. A 50 year old man comes to clinic with new low back pain, and his doctor needs to decide if this is serious and needs imaging, or low-risk and needs conservative care. How much context does the doctor need to decide that? Here’s an example of the JST for this decision:

JST for Imaging New Onset Low Back Pain

• Review of past medical and surgical history

• Evaluation for red flag symptoms (paralysis, incontinence, weight loss)

• Prior history and pattern of low back pain

• Any recent trauma?

• Focused review of systems for low back pain to exclude rare causes (malignancy, autoimmune)

With this set of information, our doctor has enough context to decide with confidence whether there is anything big and bad going on that needs imaging or whether conservative care is appropriate - and to defend that decision to peers. More information might theoretically be helpful, but the costs of information gathering start to outweigh the benefits.

What ends up on this JST is not settled science. For instance, some would argue a physical exam belongs on this JST - and they might be right. But a 50-year-old with new low back pain and a completely unremarkable history is getting conservative management. The exam may be helpful - but it also introduces friction and delay into the decision to decide. Suddenly, action cannot be taken until a physical exam is completed - an exam which won’t change management in most cases. Whether or not the physical exam sits above the JST in a low back pain patient is not the point - the point is that we will need to decide how we want to decide. The JST is a framework for making explicit what has always been implicit, and reasonable clinicians may draw the line in slightly different places. What matters is that the line exists.

Types of Decisions

The JST is a dynamic threshold - it changes with the types of decisions being made. To understand where the JST falls for any given decision, we need two axes: urgency - is there time pressure to act? - and consequence - what happens if this decision is wrong, or isn't made at all? For high-urgency, high-consequence decisions like a cardiac arrest, minimal JST is needed to act - an AI system need not wait to process the patient’s entire medical record before activating an emergency response and prescribing CPR. At the opposite end, low-urgency, high consequence decisions like selection of chemotherapy would require deep context before an AI will recommend a regimen - we might actually want to read all 1000 pages, and then some.

There will be many different types of decisions that AI will need to make, and in order to begin defining JSTs, we will first need to classify decision types. Broadly, decisions can be placed into four main categories:

Action: High urgency, high consequence. Think CPR for cardiac arrest, TPA in acute stroke, or crashing the OR for a trauma patient. The JST is minimal because delay itself is the danger to the patient: a short, non-negotiable checklist might be required, then go.

Accuracy: Low urgency, high consequence situations where time is on your side, but the cost of a wrong decision is severe. The JST is high because more and detailed information is usually still helpful. In cancer, more pathology, more genetics and deeper values conversations all add up to a better decision - even if it takes time.

Efficiency: Low urgency, low consequence decisions where the volume is high and risk per encounter is small. Here, society prioritizes throughput in order to meet population needs with available capacity. The JST is standardized so we can routinely handle volume while accepting some uncertainty.

Clearance: High urgency, low consequence situations where the patient has an acute need that merits a timely response, but is most likely not in a true emergency. The JST is confirmatory - the clinical task is to rule out serious conditions and verify that the situation is under control and does not belong in another quadrant.

Knowing what type of decision you’re making is an important component of applying the JST. Clinical scenarios that are misclassified as needing clearance when what they really need is action will lead to errors in AI just as they lead to errors in human-driven care today. Even very severe tension headaches are benign and respond to acetaminophen - unless they are an acute intracranial hemorrhage. Clinicians are taught early on to recognize “sick” from “not sick;” similarly, we need AI to accurately discriminate between various decision types so that the correct JST can be selected for a given situation.

Finally, it is important to recognize that although the JST establishes the floor for information gathering, there is also a ceiling. Over-gathering information leads to harm in each quadrant. When applying technology to decision making, there is an instinct that “more is more” because the marginal costs of gathering additional information appear small against the risk of ‘missing something’ within an automated process. Although well-intentioned, that viewpoint neglects the costs of gathering more data, costs that human actors also typically navigate implicitly. The CT scan that offered an extra margin of safety for low-risk abdominal pain felt like good medicine - until it led to an anaphylactic reaction to IV contrast.

A normative specification

The point at which an AI has ingested enough information to move forward with a decision is not a technical specification, but a normative one - set by physicians who design AI systems and reinforced through societal accountability structures that will need to be built. Every clinical decision AI makes will require navigating uncertainty and balancing risk against benefits - but at some point, inaction is itself a form of decision making when we accept the consequences of the status quo.

The JST acknowledges that more information might be helpful, and that there are real tradeoffs between comprehensiveness and accuracy. But instead of accepting decision paralysis, it defines the threshold at which the benefits of waiting to collect more information are outweighed by the harms of waiting to act.

Anyone using modern AI tools based on large language models recognizes that they are something different. It is a dramatic and foreign experience to type instructions into a chat interface and receive an elaborate response - one that not only looks like it came from a human being, but reads like it came from a particularly smart one.

The difference is not subtle, either. Many technologists I have spoken with over the years independently share the same story: the first time they used ChatGPT, usually in November of 2022. The common theme in these stories is that their first interaction with a modern LLM was an epiphany moment. For people who had spent years working with traditional software or even sophisticated ML tools, there was an immediate recognition that a transition had occurred. This was something different.

As LLM-based systems advanced even further, the gap between them and all of our prior computer-based technologies has only widened. GPT-3.5, the model powering the original public release of ChatGPT, was remarkable for its ability to effortlessly handle natural language, but it was still relatively rudimentary. It was often wrong on the facts, hallucinated responses that weren’t based in reality, and generally did mediocre work. It still felt different, but some chose nonetheless to dismiss these systems as parlor tricks or “stochastic parrots.” Sure, they might look different, but the LLM emperor had no clothes and all would be revealed as the AI hype cycle wound down - or so it was thought.

Now here we are a few years later, and the differences between LLMs and everything else that came before them have become large and undeniable. The best models now achieve world-class performance across a wide variety of domains, and new agentic systems are capable of carrying out complex work at scale. Each week, over a billion people turn to LLMs for help with all aspects of human life - as thought partners, research assistants, teachers, mentors, even companions.

These are not tools of automation or computation - these are a new form of cognition. LLMs might still run on electrons and silicon wafers, but that is where the comparison with traditional software ends and a new horizon begins.

A senior technology leader who recently started using modern AI more often in his own work described his adoption experience to me very succinctly:

“AI is now the smartest person I know.”

Airplanes Are Not Flying Cars

Technological transitions can look blurry at first, but accurate identification and classification of technological shifts becomes hugely important in appropriate design and management of new capabilities. This goes beyond semantics - by making a clear determination about what something is and what it isn’t, we are implicitly making decisions about how we manage and control that technology in the real world.

Let us run a thought experiment and explore the similarities between airplanes and automobiles. The first car was invented in 1885, and the first powered flight would not come until 20 years later in 1903. On the surface, one might say these technologies look similar - and this is a comparison worth exploring.

Similarities between airplanes and automobiles:

• Forms of transportation carrying passengers and cargo

• Powered by engines and liquid fuel

• Operated by trained human beings

• Capable of high speeds

• Require specialized infrastructure to function

• Economically transformational

• Harmful if misused

Taken at face value, airplanes and automobiles seem incredibly alike. One might even conclude that airplanes are just derivatives of the preceding category of innovation. It would be easy to come to the conclusion that airplanes are just flying cars.

Of course, that conclusion would be incredibly wrong. Airplanes were in fact an entirely new category, and society spent the next 50 years installing new agencies, systems and operating standards for ensuring that aviation could function safely and effectively.

For the past three years, the technological transition towards LLM-based systems has blurred the line with everything that has come before them. In many ways, we’ve been treating LLMs like “flying cars” - as if they were an incremental derivative of the software and consumer apps that came before them. This perception is in no small part because LLMs fit into our lives through the channels we are used to - we download ChatGPT from the App Store, log in using our Google account, and type commands to generate Word and Excel files. All of this feels familiar - and yet, what comes out of these tools on the other end could not be more different.

Cognitive Systems

LLM-based technologies are best classified as cognitive systems capable of both general and specialized intelligence. This represents a fundamentally new category that has never before existed on Earth. It is not an incremental advance - it is a giant leap.

Defining LLMs as cognitive systems starts to clarify much of the debate about the role of AI in society. Rather than arguing technicalities, we can base what happens next on actual capabilities. We now have technology systems capable of human-level cognition - what do we do now?

Managing aviation the way we managed automobiles would have led to widespread dysfunction - a new set of standards, controls and norms was needed. Similarly, attempting to manage cognitive systems using the standards, controls and norms used for older technologies will fail for a simple reason - those criteria are no longer fit for purpose. An entirely new system is needed, and without it, entire industries will continue to be held back.

These systems are undeniably powerful, and like any powerful object, they must be controlled and contained. Society places limits on the use of every technology proportionate to its potential for misuse, and it would be reckless to suggest that we should let cognitive systems operate wholly unconstrained. We don’t just let anyone step into the cockpit of a 747 or take care of patients in the hospital.

On the flipside, aggressive attempts to limit the applications of cognitive systems will slow the innovation cycles and natural learning required to understand what elements actually need controls, how those controls should be designed, and how to effectively weigh the impact of those controls against access to benefits. Honorable people may disagree on the exact balance, but they cannot debate whether that tradeoff exists - it is observable and consistent over time.

Cognitive systems are both artificial and intelligent. Seeing them clearly for what they are lets us stop debating what they aren’t - and start building the new regime under which these technologies can be deployed.

Update: 2/27/26

Since publishing this article, I have been asked for a definition of cognitive systems. Here it is.

Cognitive Systems are artificial systems that perform sustained, context-dependent reasoning across complex information environments to produce judgments, recommendations, or actions that would otherwise require human expert cognition.

Recent advances in technology have forced us to reconsider this question - one we have not had to ask in a long time. New AI systems are proving to be exceptionally skilled at knowing what to do when you are sick, and even pretty good at helping you do it. If computers can best doctors in technical knowledge and even carry on some form of the relationship with patients, we find ourselves suddenly asking why people need to be involved at all. We find ourselves asking for the first time in a hundred years, what is a doctor?

An honest answer to this question does not start by romanticizing a bygone era of medicine. That would simply be an exploration in nostalgia and do little to reflect on the world as it actually is today and where it is inevitably going. Longing for a world that no longer exists only serves to answer a different question - what was a doctor?

Medicine has encountered large role transitions before, as have other professions, and we should not shy away from that. But exploring our past to inform the present is a pre-requisite to understanding where the future of medicine is headed and how to shape it. We must seek to find the deeper insight rooted in history, and tailor that insight to reflect the present moment.

So - what is a doctor? Fundamentally, a doctor is a moral agent of society - a locus of trust, authority and accountability. The state, acting on behalf of its citizens, licenses doctors to independently carry out activities intended to improve health - collectively known as the practice of medicine. Doctors are then held accountable for the results of that practice through a combination of personal standards and professional oversight. Although there are many expectations embedded in this definition of a doctor - technical excellence, interpersonal skill, duty to care - the role itself is anchored to a moral charge.

When we recognize a doctor as a moral agent whose goal is to distinguish right from wrong in matters of human health, the impact of technological change can be correctly framed. We can allow ourselves to disaggregate raw medical knowledge and clinical reasoning from the role of doctor, and instead see doctoring through its original lens - as a steward of judgment and arbiter of correctness. Of course, such stewardship must rely on knowledge - but knowledge has always been only one part of being a doctor.

If medical knowledge becomes less important in the role of a doctor, some have suggested that the ‘relational’ aspects of medicine will become more important. This is also true - but it too must be correctly framed. The doctor-patient relationship goes beyond friendly rapport between two people. Relational medicine is anchored in mutual trust and duty towards others - trust that your doctor has your best interests at heart because it is their moral prerogative, and a duty to take action on your behalf to maintain health. As a doctor, every relationship carries moral weight.

It is now clear that technology will advance to the point that computers will be more knowledgeable than humans. This is already true in many other domains, and it is coming true in medicine. But a doctor does something no computer ever can - carry the weight of another person’s life on their conscience. This makes the indispensability of doctors again self-evident. If we view medicine through this perspective, we can move away from the hysteria around “doctor replacement” and take on the more grounded view of “doctor reassignment.” If we no longer need doctors to complete basic tasks like diagnosing simple infections or refilling common medications, so much the better - their moral agency can be applied to new ends.

Reminding ourselves what doctoring is really about can bring a renewed clarity of purpose to our profession and commitment to our practice. With the knowledge that doctors will always be indispensable, we can begin the exciting work of reimagining what medicine looks like in the years and decades ahead. There is no doubt that modern technology will fundamentally change what doctors do - but no amount of technology can ever change what doctors are.

Additionally, evaluation criteria need to approximate something that is both understandable and valid as representative performance in a domain and care setting. An extremely talented generalist will struggle in the cardiac ICU, just as a seasoned CT surgeon will be unequipped to manage a complex primary care panel. In medicine, our evaluations accomplish this by structuring assessments first by general medical knowledge (USMLE), then by domain (specialty boards), then by site of care and patient population (CME). With AI, a similar paradigm can be employed.

MedEval is a framework for structuring the evaluation of AI systems in care delivery to assess technical performance and identify failure modes in the system in which they are expected to exist. At its core, it answers two questions:

1. How often and under what circumstances does AI perform well?

2. How often and under what circumstances does AI underperform?

MedEval has 5 components which allow for flexible design of evaluations and which map to current professional assessment.

Domain refers to the knowledge base expected to be assessed. This may range from general to highly specialized.

Site of Care then modifies that assessment to include the types of medical conditions which are encountered in different care settings.

Patient Population further tailors the evaluation by varying the types of patients in which these medical conditions manifest in order to capture how disease affects different populations differently.

Case Mix adjusts the acuity, complexity and commonality of examined cases.

Modality adjusts for how an AI is expected to interact with a patient (text, voice or multi-modal).

With this system in place, AI systems can now begin to be evaluated for performance on representative scenarios that start to look a lot more like those they will encounter in the real-world.

The engine that powers MedEval is the AI Patient. These are synthetic cases that encode a “ground truth” upon which to base the interaction with the AI system being tested. These AI Patients are conceptually the same as the standardized patients encountered by clinical trainees throughout their career. They have a complete backstory and a known diagnosis upon which to base the subsequent evaluation of performance. Because the cases can be reused, performance in individual cases and in aggregate can be examined over time. The evaluation uses an LLM-as-judge to score the case series, with focused human review utilized selectively.

It is tempting to ask how MedEval should be “scored,” but that is not the intent of this piece. There are many valid approaches to measurement. Some will be obvious - like diagnostic accuracy or treatment concordance. But other factors become more nuanced to “measure” per se versus observe - especially complex, cascading failure modes and nuanced situational deviations. Furthermore, measurement schema will also need to introduce many new domains beyond diagnosis and treatment including communication quality, comprehensiveness of the history, ability to surface information, ethical behaviors, readability, empathy, etc.

The more important question is what kind of performance should we expect? But this is not a binary question. AI doesn’t “pass the test” - it just does better or worse. It is not unreasonable to define minimum thresholds and to base them on some heretofore undefined standard, but that also creates opportunity for gamification. To quote Don Wheeler from Understanding Variation: Managing Chaos:

“When people are pressured to meet a target value, there are three ways to proceed:

1. They can work to improve the system

2. They can manipulate the system

3. They can manipulate the data”

MedEval should encourage good faith improvement of the system, not manipulation to hit a target. This is particularly true because AI does not exist in isolation, but within complex systems. An AI with relatively “high” diagnostic concordance may be unsuitable for deployment if diagnostic accuracy is not materially important. Alternatively, an AI with low expected diagnostic concordance might actually represent step change improvement within the current system and merit deployment. These questions are situational and varied, not predictable and routine. And most importantly, technical performance is just one input into the ultimate decision to deploy when using a Care Worthiness standard.

MedEval is a measurement framework for others to build upon, not a prescription for how the evaluation must be carried out. As we migrate from static to dynamic benchmarks, the MedEval framing can be applied to diverse types of AI systems in care delivery, with results tailored directly to the specific usage expected of the technology.

In Utah, a new service now allows patients to use an AI system to refill their chronic medications. This has opened a deeper conversation about where medicine is headed - one deserving of serious dialogue.

No one denies the problems this service is solving:

• Patients waiting days for a response or weeks for appointments before a refill can be processed

• Harm from missed medications, including disease exacerbation and therapeutic abandonment

• Loss of trust in a system that never quite seems to work when you need it.

Similarly, no one denies the risks of inadequate solutions either:

• Missed symptoms that threaten safety

• Labs that go dangerously unmonitored for years

• New drug interactions that slip through the cracks

But the real dialogue surrounding Utah is about more than just medication refills. Some view this as the first step in a descent into a world without doctors - one in which medicine is delivered by machines and physicians become relics of a bygone era. It is true that a small and vocal minority desires that future. I do not count myself among them.

Medicine is an inherently moral practice with human relationships at its center - not technology. But moral practice does not mean manual practice. Medicine has long acknowledged a responsibility to use better tools when they reduce harm. Even valid concerns about the impact of those tools on the profession are not wholesale reasons to reject their use - including AI, one of the most useful tools ever created.

Self-service and automation are now the norm across much of modern life. Medicine remains different, and for good reason. It takes expertise, skill, and vigilance to care for people and keep them safe. Every physician can vividly recall a moment when apparent simplicity gave way to complexity - when a refill was no longer just a refill.

Whether it’s renewing medications, conducting a routine office visit, or even just responding to a short portal message, all of these encounters can become “something more” the moment uncertainty enters the picture - when new symptoms, missing data, or a clinical change mean that proceeding safely requires judgment, not just execution.

Until now, we have lacked good tools to manage that complexity beyond human effort alone. At the same time, our existing tooling and processes create gaps. In many practices, refills are managed through checklist-based protocols and asynchronous chart review. These informal clinical encounters are performed with variable scrutiny and often limited data about a patient’s current context. Even when patients are asked to come into clinic before receiving a refill, delays in scheduling or lab work frequently lead to medication gaps, and trust is lost when patients feel as though their medications are being held hostage. None of these are neutral outcomes. They cause real harm today.

Improving the system is the goal. Whether an AI-enhanced clinical service refilling medications represents an improvement is an empirical question - one that is being answered through rigorous evaluation under regulatory oversight. But the expectation is that this workflow improves outcomes beyond the status quo - by making clinical uncertainty more visible, and identifying more often when human judgment is required. 

The service as designed today is not built for “convenience.” It is architected for robustness, layering multiple safety gates that every patient must pass through before a medication refill is deemed safe and reasonable. Only then can the service be delivered. AI adds a new dimension by creating a structured opportunity for patients to report changes in their health - something they often do not have the chance to do in current practice. And when uncertainty arises, a clinician remains available for escalation at all times. No one is forced to bypass human interaction.

The view that our relationships could be eroded by AI is well-founded, and it is one I share. Like most modern physicians, I have spent far more time in my career staring at an EHR when I would rather have been engaging with patients. That pain is real - and it has nothing to do with AI.

Now, we have an opportunity to shepherd in a new class of technology into our work. If done well, it will create more room for the relational care we all desire. Technological advancement is not an end unto itself, but a mechanism of improvement. Complex systems evolve by design or by drift. The task now is to choose the former.

Any advancement in medicine requires physicians at the helm and patients at the center. For new clinical services leveraging technology, that means making judgment explicit, uncertainty visible, and escalation reliable. That is what is being done in Utah.

Doctor is a role, ascribed by society, with responsibility for carrying out functions which offer the potential for benefits and the risks of harms to human health. These functions constitute the “practice of medicine” - the various forms of reasoning, judgement and actions under uncertain conditions undertaken with direct responsibility for the outcome.

To accomplish these functions, doctors have always turned to an increasingly advanced set of tools which aid them in that practice, including physical, chemical and cognitive instruments. The stethoscope let us listen to the heart and lungs, penicillin helped us treat disease, and reference sources like UpToDate improved decision making.

Now, a new set of tools are so highly capable that there is confusion about whether they somehow can become “a doctor.” This confusion is rooted in mistaking capability for responsibility.

Doctor is a role.

Practicing medicine is a function.

AI is a tool.

Tools, no matter how capable, cannot serve in a role. That is because a role is a vehicle for the expression of morality - or what is right. An antibiotic might cure your infection, but it did not decide whether you would be better off taking it.

Conflating a technical capability with a moral responsibility treats performance as if it were permission. This creates unnecessary confusion about why doctors exist, and gives rise to the flawed discourse about whether they still need to.

Doctor is a role.

Practicing medicine is a function.

AI is a tool.

Clinical reasoning is everywhere. AI can now generate differential diagnoses, assessments, and plans on demand. The labor to produce clinical reasoning is basically zero. With a simple prompt, AI can create something that looks thoughtful, thorough, and clinically fluent - just like a doctor wrote it.

This has given rise to a new problem: notes, plans, and recommendations that are factually accurate, plausible and mostly reasonable - but which lack contextualization. They are clinically hollow - the medical equivalent of “AI Slop.”

But clinicians who use AI seriously know something important.

AI doesn’t just generate medical content.

When used deliberately, it generates clinical possibilities - diagnostic and treatment options that can reasonably be applied.

And these clinical possibilities, when examined through experience, transform into clinical judgements. AI can now surface ideas at a scale and speed never before accessible in medicine. That isn’t clinical slop. That is cognitive leverage.

Put another way:

• Behind every great clinician is access to a deep bench of specialty knowledge.

• Behind every sound diagnosis is disciplined synthesis.

• Behind every reasonable plan is a risk-benefit calculation.

These are all functions that AI can now assist with - and often excel at.

Suddenly, the question is no longer:

“Can AI generate an assessment and plan?”

It is:

“Is this assessment and plan worthy of action?”

Clinical judgement is the responsibility to decide which signals are allowed to become care.

And it is learned - sharpened through exposure, error, responsibility, and sustained engagement with real patients over time.

Having clinical judgement requires experience, intuition, and skill.

But exercising clinical judgement means deciding what crosses the threshold into care.

Understanding this distinction has become central to resolving a new dilemma in medicine.

For clinicians, AI raises an uncomfortable question:

If an AI suggests a diagnosis or plan, is the clinical judgement still my own?

If the system surfaces the right idea, can the clinician claim ownership - or have they suddenly outsourced their expertise?

That tension is real.

One response is to reject AI entirely. But that creates a different failure: refusing the most powerful clinical reasoning amplifier ever built.

The solution is not abstinence.

Ownership in medicine has never come from generating ideas. We call consults and reference literature daily. It comes from deciding which ideas are relevant and appropriate for a specific human being.

The correct response to working with AI is to double down on exercising clinical judgement:

• When AI suggests something unsafe, you stop it.

• When AI suggests something interesting, you explore it.

• When AI tries to do too much, you restrain it.

• When AI makes a plan you can’t stand behind, you discard it.

Because in medicine, curation is now more important than generation.

Judgement is the act of discernment under responsibility.

And clinical judgement remains the final arbiter of what becomes care.

In medicine, we communicate more than just facts. The things we share with our patients carry meaning - the story a person constructs from information.

Like many things, meaning exists on a continuum:

• Factual Meaning: “What is this?”

• Interpretive Meaning: “What does this means for me?”

• Emotional Meaning: “How should I feel about this?”

• Existential Meaning: “How does this change my life?”

As the weight of a message gets heavier, the trust required to deliver it increases. And until now, human beings remained the only actor capable of delivering complex meaning.

But that has now changed in a world with AI - which can deliver complex meaning instantly, at scale. How do we approach this new reality? It will not be sufficient to show that AI can be right on the facts - we need to decide what types of messages - and meaning - that AI can deliver.

To answer that question, modern healthcare needs a new framework: Top of Relationship

As healthcare professionals, we often talk about practicing at the “Top of License” - making decisions only we are qualified to make based on our training and credentials.

• MDs diagnose and recommend treatment

• RNs triage and counsel

• MAs administer assessments and provide basic guidance

But if a license gives you permission to decide the right thing, a relationship gives you permission to say the right thing.

A relationship lets you convey meaning and be met with gratitude, not insult.

Think about two different scenarios:

Scenario 1:

• Your trusted doctor diagnoses you with cancer.

• They bring you to the office and tell you with care and empathy.

• You feel grateful to have a doctor who knows you and shows up.

Scenario 2:

• Your trusted doctor diagnoses you with cancer.

• They ask the front desk to call you with the result.

• You feel livid. Betrayed.

Practicing at the Top of Relationship means matching the importance of the message to the depth of trust between the messenger and the patient. Because when trust is high, you can say a lot. And when trust is low, you might need to hold back and wait for a better resource.

A hug from your mother means something different than a hug from a stranger.

A hard truth from a lifelong friend lands differently than the same words from an acquaintance.

The question isn’t just whether information is correct -

It’s who has earned the right to deliver it.

AI systems are increasingly embedded directly into care delivery and are performing advanced functions including history taking and clinical reasoning. There is enormous potential to expand the scope of these tools. However, organizations face increasingly complex decisions about when, where and how AI can or should be deployed.

Numerous groups have developed frameworks to help organizations make decisions about whether to use AI in healthcare applications. However, these frameworks suffer from three major limitations that hamper effective decision making:

• Current frameworks overwhelmingly focus on evaluating technical performance relative to other important domains

• They do not specify what value system applies when making adoption decisions

• There is no defensible standard for basing an adoption decision

The consequences of these constraints are predictable - evaluations focus on whether AI ‘passed the test,’ and then stakeholders apply different values and norms to debate whether the results met a threshold that was never defined.

In this environment, progress understandably stalls. At the organizational level, no one can agree on what ‘good’ looks like leading to endless internal debate. At the industry level, variation in adoption decisions create fear that first-movers will be unfairly punished. At the national level, this leads to a slowing in the pace of advancement in technologies that could greatly benefit patients.

There is a need for a better method that can overcome these limitations and promote action from organizations and their leaders. We don’t just need better metrics, we need better decision logic.

Towards Care Worthiness: A New Decision Standard for AI Adoption

Care Worthiness is a new method for making decisions on AI adoption in healthcare by anchoring deployment to commonly accepted standards of professional entrustment. This allows for richer conversations about the specific role of an AI system in care and a determination of whether that role is acceptable based on the application of a moral standard, not simply technical one. Teams applying a Care Worthiness lens are able to see decisions in a new light because the standard governs overall permissibility, not raw performance. If a system is deemed Care Worthy, teams can feel confident that they are acting in a way that is aligned with societal values.

Care Worthiness extends the logic of professional entrustment to AI systems. When evaluating care worthiness, organizations define the role that a given AI system is intended to take on in the care journey, then assess whether the AI system is worthy of that role given current knowledge on performance, consideration of the degree of risks involved in deployment, and the existing safeguards to manage and mitigate those risks. Any determination of acceptability must meet not only a technical standard, but the moral one - it must feel like a responsible, grounded choice given the sum of available knowledge.

This is not a relaxation of standards - professional entrustment is a demanding process that includes multiple layers of accountability including patients, peers, supervisors and regulators. Care Worthiness applies medicine’s own high standards for delegating clinical work to AI systems. Additionally, it also recognizes that knowledge about any system will always be imperfect, but that clinically-grounded, morally responsible and institutionally defensible decisions can still be made even without perfect information

Why a Moral Standard is Needed for Seemingly Technical Decisions

Technical evaluations can tell us important information such as accuracy, robustness, bias and failure modes. But they cannot answer moral questions like how much variation or potential for error is acceptable, who should bear the risk and responsibility of system performance, what level of supervision is sufficient, and whether the role itself is appropriate. These are normative questions, and in medicine normative questions are settled by a combination of professional codes, accountability structures and individual moral reasoning.

The lack of a moral lens through which to interpret current AI performance evaluations is why so many organizations feel stuck ever after the evaluations are in. Without a clear moral lens, each decision feels existential - that even minor disagreements on acceptability might snowball into major ruptures within organizations, or worse, that participants might be on the wrong side of history.

There is no doubt that technical evidence must be inputs into any framework evaluating AI, but they must be in service of moral reasoning, not a substitute for it. Technical evidence may illuminate what is possible, but Care Worthiness determines what is permissible.

Key Input Domains to Determine Care Worthiness

These domains are not a checklist to determine Care Worthiness - rather, they inform a judgment about whether delegating a role to AI is morally responsible given the existing knowledge about the system.

Determining Care Worthiness: The Entrustment Bar

An AI system passes the entrustment bar for Care Worthiness if a reasonable clinician and organization could defend the decision to entrust it with this role - before patients, peers and regulators - given the known risks, safeguards and consequences of failure.

This determination can be made by applying the 3-part test below.

The Entrustment Bar

Towards a Care Worthiness Standard

The bigger problem in determining whether AI can be used in care delivery is not simply lack of evaluation, but decision paralysis. When standards are opaque and devoid of moral authority - especially in healthcare - well-meaning differences in perspective become hard-stop blockers to progress.

The good news is that medicine has already solved this kind of problem through entrustment. Clinicians and other healthcare professionals are subject to evaluation on multiple domains - both technical and subjective - before being deemed as worthy of a clinical role. Even then, their scope is defined, and they exist within robust systems that protect against variation and error.

The definition of Care Worthiness can evolve, as can evaluation criteria. But as a standard, it offers a path to responsible progress. Care Worthiness offers a new way forward that neither freezes innovation nor abandons responsibility - by grounding decisions about AI participation in the same logic of entrustment that has long framed how medicine delegates care.

The reality is that healthcare has always been governed by moral standards first, and technical standards second. Care Worthiness restores the correct order when it comes to deploying AI in care delivery.

Here’s the system behind it: AI-Native Care Delivery.

AI-Native Care Delivery is an operating model for clinical care. Not a feature, not a workflow add-on, but a fundamentally different way of organizing how care is delivered.

The idea that the shift towards this care model is inevitable rests on a simple premise.

It can answer questions, reason over clinical data, and act.

Design care around those capabilities, and an entirely new delivery model becomes possible.

In an AI-Native Care Model, care is structured around four modes.

1. AI Chat: Handles common clinical and administrative questions.

2. AI Consult: Conducts the visit before the visit. Synthesizes history and context and conducts the first layer of clinical reasoning into an assessment and plan.

3. Autonomous Care: Manages well-bounded, low-risk problems that should be fast and efficient.

4. Human Care: Manages unbounded, higher-complexity problems that require judgment and trust.

The key shift isn’t that doctors do less.It’s that doctors focus on the parts of care where human connections matter most.

What makes this model durable is that it aligns incentives naturally.

• Patients get faster access and fewer dead ends.

• Clinicians get relief from transactional work.

• Organizations expand capacity without expanding headcount.

Importantly, AI-Native does not mean AI-First. The latter is a different concept rooted in dictating patient flow, while the former is a concept rooted in patient enablement.

AI-Native Care Delivery isn’t speculative. Pieces of it already exist in production today.

What’s new is recognizing them as a unified system that works together.

That’s the model.

I spend a lot of time evaluating AI systems for care delivery, and the current methods have reached their limit.

Today, we still grade clinical AI with knowledge exams and structured cases.
MedQA, MedR Bench, etc - these are helpful but constrained.

Useful for stress-testing a single model.
Useless once the system becomes agentic, conversational, and deployed in the messy reality of care.

Because real patients don’t hand us well-structured clinical information.
They don’t offer their labs up front.
And they almost never say the one sentence the model “needs” to get the question right.

Current evaluation methods are also designed for an outdated purpose - testing a model’s medical knowledge.

But new agentic systems aren’t built to “take a test.”

They’re built to interact with real patients - iteratively, imperfectly, like clinicians do. And we need to know well how they’re doing at that.

This is why we need the AI Patient.
An AI that can behave like a patient across thousands of known cases, each with ground-truth diagnoses and management paths.

It selects a case, then starts a conversation.

The system goes to work case-by-case, repeating new presentations across easy, hard, and rare situations.

When the visits are done, an LLM-as-a-judge reviews the interactions to flag safety issues, missed cues, or reasoning errors. These get funneled back into a dashboard for review by the development team.

Test → re-test → refine.

This becomes a continuous loop that costs a few hundred dollars and runs in minutes.

It also mirrors how human clinicians learn:
See more patients.
Get expert feedback.
Close the gaps.

I believe the AI Patient will become the standard for evaluating AI systems in care delivery.

What capabilities would you want the AI Patient to have?

This isn’t just data gathering. An AI Consult is a clinical-grade visit - including history, preliminary assessment, and initial plan - between a patient and an AI system. The output becomes a progress note the clinician reviews before the visit, then uses to ask clarifying questions and finalize the plan.

Think of it another way - it’s like the best Chief Resident you ever worked with prepping every case.

Why is the AI Consult transformative? Because overnight we will take 60-70% of the visit time that used to be spent in synchronous data gathering and chart review and free that time up for more discussion on the major issues and the plan.

And the benefits are huge and immediate. Visit quality goes up because doctors and patients are now spending most of their time focused on making decisions, not just exchanging data. And efficiency improves because what used to be a 30 minute visit comprised of mostly low-quality interaction is now a 15 minute visit comprised of almost all high-quality interaction

.

Patients feel like their doctor finally has the time and space to pay attention to them. Because they do.

For practices, the implications are straightforward and dramatic: throughput doubles while operating expenses only increase by ~20%. Overnight, primary care turns from a loss leader to a profit center, and major capacity issues get resolved.

For patients, there will be more opportunity to share a full set of concerns before every visit - so conversations with doctors are richer and more informed.

And for clinicians, we will have more bandwidth than ever to focus on our patients, not the chart.

It would be naive to think that such a profound shift would happen quickly, but it will be impossible to ignore - and it will come faster than anyone is expecting.

If the visit starts before the visit, what changes next in the system around it?

Always On Call is an exploration of these topics at the furthest edges of AI as applied to human health and wellbeing.

I hope these articles are helpful to others in sparking new ideas. It goes without saying that they are only possible because of the thoughtfulness and intellectual contributions of countless others.

Subscribe now