Menu
Close
.svg)
A little over a year ago, I wrote an essay for LinkedIn about the prospect of physicians having their own J.A.R.V.I.S., i.e., something akin to Tony Stark's remarkably capable, British-accented AI assistant in Iron Man. The concept was that AI would sit beside the physician, ingesting enormous quantities of information, taking care of the administrative drudgery they detested and, increasingly, helping doctors make better clinical decisions. At the time, physician enthusiasm for AI was escalating, even as understandable concerns about privacy, accuracy and the physician-patient relationship remained.
Here we are, a year later, and that prediction looks both prescient...and somewhat too narrow. Doctors are indeed getting their J.A.R.V.I.S. But so are patients. Health systems are joining the party too, embedding AI into the infrastructure through which care is delivered. And increasingly sophisticated AI models are demonstrating capabilities in clinical reasoning that would have seemed impossible only a few years ago.
The important development, therefore, isn't simply that healthcare is "adopting AI." AI is beginning to insert itself into the chain of information and decisions connecting patients, physicians and health systems. For pharmaceutical companies, and especially commercial teams launching and managing in-line brands, this represents something much bigger than another technological disruption. It has the potential to reshape how patients understand disease, how physicians evaluate treatment options and, ultimately, how medicines are chosen and used.
The numbers certainly suggest that something has changed. The American Medical Association's 2026 survey found that 81% of physicians reported using AI professionally, more than twice the rate in 2023. Other studies measuring enterprise deployment produce lower figures, which is not really contradictory. A physician can start using an AI research tool like OpenEvidence this afternoon. Persuading a health system to integrate one into its EHR requires procurement, security reviews, governance committees and the other rituals by which healthcare ensures that nothing bad happens this afternoon.
What matters more than any single adoption statistic is how AI is being assimilated. One of its first killer apps has turned out to be decidedly under-glamorous: taking notes.
Ambient AI scribes listen to the clinical encounter and produce documentation, summaries and other outputs. In Britain's National Health Service, for example, a multi-site evaluation involving more than 7,000 patients found productivity gains and more time for direct patient care. HCPs involved in the study describe lower cognitive burden and a greater ability to focus on the person sitting across from them rather than the computer sitting beside them.
There is a lovely irony here. For years, one of the fears surrounding AI was that technology would make medicine less human. One of its first significant accomplishments may be getting doctors to look at their patients again.
But the scribe is not remaining a scribe for long.
Consider Abridge for a moment. It entered the healthcare arena as an AI documentation company. It is now evolving toward what it describes as a clinical intelligence platform, operating before, during and after the patient encounter. Its clinical decision-support (CDS) capabilities can recognize the context of a particular patient and surface relevant evidence from sources including UpToDate, The New England Journal of Medicine, JAMA and the Journal of Clinical Oncology.
That distinction is relevant. Traditionally, a physician confronting an unusual presentation might have needed to survey the literature, consult a reference database or phone a colleague. Increasingly, the relevant evidence can find the physician in real time instead.
And this is happening at meaningful scale. Abridge claims its technology is deployed across more than 300 health systems and supports more than 100m clinical conversations annually. Its CDS capabilities have already been employed in cases ranging from unusual rashes to rare neurological presentations. The distinction between AI reducing clinical work and AI influencing clinical work is beginning to blur.
Meanwhile, the technology itself is advancing quickly. A 2026 study in Science by Peter Brodeur et al. tested an advanced reasoning model against hundreds of physicians across challenging clinical cases involving diagnosis and clinical management. Across five experiments, the AI outperformed the physicians against whom it was compared. Importantly, the researchers went beyond clinical vignettes to include a real-world comparison of human and AI second opinions for randomly selected emergency department patients.
This does not mean that "AI is better than doctors", a conclusion that would be both irresistible to headline writers and unsupported by the evidence. Medicine involves physical examination, procedures, empathy, ambiguity, communication, accountability, experience and judgment that cannot conveniently be reduced to a clinical reasoning exercise. But it does mean that AI is beginning to match or exceed physicians on specific cognitive tasks that sit remarkably close to the heart of clinical decision-making.
A recent JAMA commentary by Ezekiel Emanuel and colleagues that I posted about yesterday takes the argument a provocative step further. Reviewing the emerging evidence, he and his co-authors contend that generative AI already rivals or exceeds physicians in five fundamental cognitive tasks: gathering medically relevant information, developing differential diagnoses, selecting diagnostic tests, recommending guideline-concordant treatment and managing some chronic diseases. More provocatively, they argue that once AI becomes better than humans at a particular task, putting a physician "in the loop" may sometimes make the result worse, not better.
There are substantial caveats to all of this. Much of the evidence still comes from simulations rather than everyday clinical encounters. AI remains brittle in certain circumstances, while autonomous systems introduce different risks involving hallucinations, cybersecurity and accountability. And regulation, reimbursement and liability remain formidable obstacles to real-world deployment.
Still, the fundamental questions are evolving. We have spent several years asking whether doctors should trust AI. We may eventually have to confront a more uncomfortable question: when should doctors defer to it?
While much of the healthcare-AI conversation has focused on physicians, something equally consequential is happening on the other side of the exam-room table.
Boston Consulting Group recently surveyed more than 13,000 internet-connected consumers across 15 countries and found that nearly 60% had already used AI-powered tools for personal health. Consumers are using AI to ask health questions, interpret test results, understand treatment options and combine medical information with data from wearables and other sources. This is more than Dr. Google with better grammar. Search engines gave patients access to information; conversational AI can synthesize that information around a person's particular question and continue the conversation.
Health systems are moving in the same direction. Oracle Health's new AI-enabled patient portal allows patients to ask questions about information contained in their own medical records, receive plain-language explanations of diagnoses and test results, examine trends in labs and vital signs and navigate appointments. The AI is integrated with the EHR rather than requiring patients to upload their medical histories into a separate consumer application.
The result could be a new form of algorithmically enabled patient activation. The informed patient is hardly new. What is new is a patient who has access, at virtually any hour, to a system capable of helping them formulate questions, interpret evidence and compare alternatives before and after the HCP encounter.
Put these developments together and the change becomes easier to see.
A patient notices symptoms and asks AI what they might mean. AI helps determine whether and where to seek care. A health-system assistant prepares the patient for the visit. During the encounter, an ambient system captures the conversation while clinical AI surfaces relevant evidence to the physician. Treatment is selected. An AI-generated summary explains what happens next. Later that evening, the patient asks another AI about the prescribed medication, its side effects and possible alternatives.
Not every patient journey will look like this, and certainly not yet. But the direction matters.
AI isn't creating one additional healthcare touchpoint. It is beginning to insert itself between many of the touchpoints that already exist.
That has profound implications for pharmaceutical commercialization.
Pharmaceutical brand teams have become extraordinarily sophisticated at understanding treatment decisions. We routinely map patient journeys, build physician segmentations, conduct ATUs, identify drivers of and barriers to prescribing, test messages and investigate why one patient receives Drug A while another receives Drug B.
We now must add another layer.
First, brand teams should map the AI-mediated treatment journey within their disease. Where are patients using AI? Where are physicians using it? Which platforms are entering the workflow, and at what decision points? AI's influence in a rare disease with a difficult diagnostic journey may look quite different from its role in diabetes, cancer or gout. "AI adoption" is too blunt a construct to be useful.
Second, we should understand AI behavior, not merely AI attitudes. Last year I suggested that pharmaceutical companies might eventually segment physicians according to the degree to which they embrace AI. I now think that understates the opportunity. Knowing that a physician is "positive toward AI" tells us relatively little.
Similar questions increasingly apply to patients themselves.
Third, launch teams should think carefully about evidence in an AI-interpreted world. If a clinical assistant is asked which therapy is appropriate for a patient with a particular biomarker profile, previous treatment history and set of co-morbidities, what evidence will it find? Guidelines, comparative studies, RWE, publication strategy and the clarity with which we define the appropriate patient may become even more important as AI increasingly synthesizes rather than merely retrieves medical information.
This isn't about "marketing to the algorithm." Algorithms, mercifully, cannot be invited to dinner at a congress. It is about ensuring that the clinical evidence supporting a medicine's value is authoritative, accessible and capable of surviving algorithmic synthesis.
Finally, I&A teams need to start to research the algorithm, not just the human. We routinely evaluate how physicians and patients respond to realistic treatment scenarios. Why shouldn't we systematically examine how important AI environments respond to those same scenarios? Which treatments are surfaced? Which evidence is cited? What happens when the patient's age, biomarker, previous treatment or comorbidity changes? How consistent are the recommendations across platforms; and how do they change as models evolve?
As AI becomes more influential in treatment decisions, understanding how it "thinks" may become almost as important as understanding how our customers do.
Over year ago, I imagined AI becoming the physician's J.A.R.V.I.S. I still think that analogy works. But the more interesting development is that J.A.R.V.I.S. appears to be multiplying.
Physicians are gaining AI assistants. Patients are gaining AI assistants. Health systems are embedding intelligence into the infrastructure connecting the two. And the models themselves are becoming increasingly capable of performing some of the cognitive tasks involved in diagnosis and treatment selection.
For decades, pharmaceutical companies have invested enormous resources in understanding how physicians and patients make treatment decisions. In the AI era, that may no longer be enough. We will also need to understand how the intelligent systems sitting between, and increasingly amongst, them shape those decisions.
For commercial leaders preparing tomorrow's launches, understanding this new participant may soon become every bit as important as understanding the traditional ones.
