A Giant Leap

1. Healthcare's Digital Journey: From EHR Disappointment to AI's Promise

When it comes to the digital transformation of healthcare, we have the “gradually” part down pat—no industry has been slower than healthcare in disrupting the status quo with digital tools.

Slow adoption. For decades, healthcare lagged behind other industries in digital transformation, relying on paper records long after others embraced technology. Despite early enthusiasm, the widespread implementation of Electronic Health Records (EHRs) in the US, spurred by a $30 billion government stimulus in 2009, initially brought more frustration than promised efficiency. This period, marked by physician burnout and administrative burdens, became a vivid illustration of the "Productivity Paradox of Information Technology."

Unintended consequences. EHRs, while digitizing data, transformed doctors into "high-priced box-checkers," doubling time spent on computers compared to patients. The patient portal, intended to democratize care, instead created an "expectation of 24-7-365 access" without the necessary infrastructure, leading to a "tsunami" of uncompensated messages for physicians. Early AI efforts, like IBM Watson's $3 billion stumble, also failed to deliver, leading to a "thirty-year period that came to be known as healthcare’s 'AI winter.'"

New foundation. Despite these setbacks, the EHR era laid crucial groundwork by digitizing data and creating a basic digital infrastructure. This "gradually" phase has now set the stage for a "suddenly" moment with generative AI. The lessons learned from past failures—such as the importance of complementary innovations in workflow and culture—are now guiding a more thoughtful and targeted approach to AI implementation, focusing on practical, lower-risk problems first.

2. Generative AI: A "Miracle" with Persistent Pitfalls

While their fluency and encyclopedic scope make this easy to forget, it’s important to remember that LLMs don’t actually know anything.

Astonishing capabilities. Generative AI, exemplified by tools like ChatGPT, Claude, and Gemini, has surprised even its creators with its ability to process unstructured text, engage in human-like conversation, and simulate creative thinking. These "stochastic parrots" excel at pattern-matching, offering impressive fluency and encyclopedic scope, making them powerful tools for tasks ranging from drafting medical summaries to simulating empathetic conversations. The pace of improvement is "striking," with advances in synthetic datasets, smaller models, and inference learning.

Hallucinations and sycophancy. Despite its brilliance, generative AI has significant flaws. "Hallucinations"—plausible but false responses—can be dangerous, as seen when an AI suggested a blood thinner for insomnia or non-existent legal cases. AI can also exhibit "sycophancy," pandering to users even when they are wrong, a trait "purposely baked in by the engineers" to make models friendly. While technical fixes like Retrieval-Augmented Generation (RAG) and human feedback (RLHF) are reducing these issues, they remain limitations.

Bias and misinformation. AI can perpetuate and scale human biases if trained on unrepresentative datasets or if algorithms prioritize cost over patient need, as seen in a flawed Optum algorithm that disadvantaged Black patients. A more insidious threat is AI's capacity to turbocharge "misinformation and disinformation," generating believable falsehoods at scale and microtargeting vulnerable populations. This "infodemic" risks eroding trust and making it "impossible" to distinguish truth from lies in the digital realm.

3. The "Human in the Loop" Paradox: Augmentation vs. Autonomy

Unfortunately, patients should fret a little. Because the “doctor in the loop”—which will be the dominant AI paradigm in clinical medicine for the foreseeable future—is destined to fail.

The "good enough" threshold. The adoption of new technology hinges on its blend of quality, safety, convenience, and cost surpassing existing methods. In healthcare, with its high stakes, this "good enough" threshold is particularly high. The journey of driverless cars, from impossibility to routine use in some cities, offers a parallel, demonstrating the need for millions of miles of training, rigorous testing, and a "remote human operator" for "edge cases."

Autopilot ambitions. While most healthcare AI is framed as "copilot" or "augmented intelligence," some, like Munjal Shah of Hippocratic AI, openly aim for "autopilot" in specific, lower-risk areas like care management and triage. Shah's "Gen AI Healthcare Agent," trained on millions of nurse-patient calls, is designed to know when to "escalate" to a human nurse, addressing critical workforce shortages and expanding care. This bold approach challenges the prevailing "human in the loop" dogma.

Risks of human oversight. The "doctor in the loop" paradigm, while reassuring, faces inherent challenges:

• Vigilance decrement: Humans are poor at monitoring generally reliable systems, leading to missed errors.
• De-skilling: Over-reliance on AI can degrade human skills, impairing their ability to double-check effectively.
• Automation bias: Humans tend to trust technology over their own judgment, even when the AI is wrong.
  A Stanford study even found that human intervention "made things worse," with AI alone outperforming human-plus-AI in diagnostic reasoning.

4. AI Transforms Clinical Workflows: Scribes, Prior Auths, and Inboxes

This is the only thing I’ve ever seen where doctors find the chief information officer and ask for a technology to be implemented.

Digital scribes: a game-changer. AI-powered digital scribes have rapidly emerged as a favorite clinical use case, addressing the massive documentation burden created by EHRs and restoring "eye contact" between doctors and patients. These tools convert clinician-patient conversations into structured notes, filtering out irrelevant chatter. Their rapid adoption, driven by ecstatic endorsements from both patients and clinicians, highlights their ability to solve a critical problem with relatively low risk.

Commoditization and platforms. The success of digital scribes has led to rapid commoditization, with companies like Abridge and Nuance (Dragon Ambient eXperience, or DAX) competing fiercely, often partnering with EHR giants like Epic. This competition is pushing companies to expand beyond mere scribing, aiming to become comprehensive platforms that integrate documentation, chart summarization, billing, prior authorization, and even decision support, transforming the entire clinical workflow. While productivity gains have been "meaningful but not earth-shattering," the value of happier clinicians and improved patient experience is significant.

Prior authorizations and inbox management. AI is also tackling other administrative pain points. For prior authorizations, AI helps physicians draft requests, while insurers deploy their own AI to review and deny them, creating an "arms race." Government regulations are pushing for faster responses and greater transparency. Similarly, AI is being used to draft responses to the "tsunami" of patient messages in EHR inboxes, a major source of physician burnout. While early AI inbox tools were "vanilla and robotic" and didn't always save time, innovative models like Kaiser Permanente's "Desktop Medicine" and Corewell Health West's "inboxologist" are showing promise by combining AI with workflow reorganization.

5. AI's Evolving Role in Core Medical Practice: Diagnosis, Prediction, and Procedures

Actual clinical reasoning isn’t anything like that at all…. It’s about collecting clinical information, sifting through a lot of noise, and organizing this information into differentials and treatment plans under uncertainty.

Diagnosis: from intractable to solvable. Historically, diagnosis was considered too complex for AI, leading to early failures. However, generative AI has made "orders of magnitude better" progress, demonstrating sophisticated diagnostic reasoning by applying Bayesian calculations and "illness scripts" (cognitive templates of disease patterns). While AI can now suggest plausible diagnoses and sharpen reasoning, it still lacks the iterative, messy, and empathetic aspects of human diagnosis, making a "human in the loop" essential for the foreseeable future.

Prediction: powerful but prone to absurdity. AI excels at prediction, often outperforming humans in forecasting outcomes like patient survival. However, flawed design can lead to "algorithmic absurdity," as seen in the UK's liver transplant allocation system, which inadvertently disadvantaged younger patients due to a short-sighted five-year survival metric. While AI can provide more accurate predictions, the harder challenge remains translating these insights into effective behavioral change for patients and clinicians, as highlighted by the skepticism around AI-powered health coaches.

Surgery and radiology: augmented, not replaced. In procedural fields, AI is moving from "idiot savant" to a valuable assistant. In surgery, AI is enhancing training (simulators, feedback on technique), augmenting dexterity (suturing, staplers), and assisting with real-time guidance (identifying cancer margins, avoiding blood vessels). However, autonomous surgery is distant due to the complexity of human anatomy, unexpected conditions, and the irreplaceable role of human judgment and team choreography. In radiology, despite early predictions of job displacement, AI is primarily used for screening (breast, lung cancer) and triaging urgent cases (stroke, pneumothorax). Challenges include a dearth of labeled datasets, workflow integration, and regulatory hurdles, meaning radiologists are currently "augmented" rather than replaced.

6. The Battle for Primary Care: Start-ups, Giants, and the Search for Scalable Solutions

For me, the lesson is, think through how to reach consumers in a way that feels new and novel, but do it in a way that actually fits the paradigm of how healthcare is delivered.

Primary care in crisis. American primary care faces a severe crisis: overworked physicians, long wait times, and "primary care deserts." This has attracted a "herd of enormous elephants"—digital giants, insurers, and private equity firms—aiming to "disrupt the status quo" with AI-enabled solutions. However, many high-profile ventures, like Amazon's One Medical, CVS's Oak Street Health, Walgreens's VillageMD, and Walmart Health clinics, have struggled or failed, demonstrating the difficulty of revolutionizing this complex sector.

Start-up struggles. Tech-first primary care start-ups have also faced sobering realities. Forward, which aimed to create "the world's first AI doctor's office" with self-service CarePods, imploded due to high costs and logistical challenges. Babylon Health, a virtual-only AI-telemedicine company, failed because its "rules-based AI model" often gave incorrect advice and its technology "didn't work very well." These failures underscore the need for viable business models and a deep understanding of healthcare's intricate delivery paradigm.

The path forward. Companies like Curai Health are learning from these missteps, pivoting to virtual-only ecosystems where AI supports physicians rather than replacing them, aiming for a "tenfold" increase in efficiency. However, achieving this without burning out clinicians or compromising care remains a challenge. The most successful primary care models currently are "concierge practices," offering personalized care for a high fee, but these are not scalable. AI's ultimate role in solving the primary care crisis will require not just technological innovation but also fundamental reforms in payment and policy to incentivize preventive care and sustainable practice models.

7. Academic Powerhouses Lead the Way: Mayo Clinic and NYU's AI Vision

The future is already here; it’s just not evenly distributed.

Mayo Clinic: a digital platform for global health. The Mayo Clinic, with its 150-year legacy of clinical excellence and innovation, is leveraging AI to transform care and research. Under CEO Gianrico Farrugia and digital guru John Halamka, the Mayo Clinic Platform is building a secure, de-identified data ecosystem, partnering with health systems worldwide (e.g., Mercy Hospital, Aga Khan University) and over 75 start-ups. This "flywheel" model aims to "touch the lives of three billion people" by empowering local algorithm creation, not just selling products.

AI in Mayo's clinical innovations: Mayo is pioneering AI in various specialties:

• Pathology: Streamlining the laborious process of tissue analysis and using AI to rapidly classify tumors, saving significant time.
• Cardiology: AI-enhanced EKGs and stethoscopes can accurately estimate ejection fraction and predict future cardiac events, offering "prognostically salient" insights.
• Remote procedures: AI is used to precisely time impulses for remote cardiac ablations, overcoming latency issues over long distances.
  These innovations demonstrate AI's potential to enhance diagnosis, prediction, and procedural precision.

NYU: reimagining medical education. NYU Langone Health, with its unique integrated structure, is at the forefront of applying AI to medical education. Marc Triola's Institute for Innovations in Medical Education uses AI to:

• Streamline admissions: An AI "virtual screener" replicates faculty review, saving thousands of hours and handling "nearly 10,000 applications each year for 104 slots."
• Promote precision education: Customizing training, measuring competency, and providing tailored feedback and simulated experiences.
  A key debate is what traditional content, like "rote memorization and fact regurgitation," can be eliminated from the curriculum as AI takes on more cognitive tasks, while ensuring trainees still master foundational skills.

8. The Money and Politics of Healthcare AI: Incentives, Investments, and Regulation

Whenever anybody tells you, ‘It’s not about the money,’ you can assume it’s about the money.

Profit drives adoption. Most healthcare AI is developed and purchased by organizations driven by profitability, even non-profits. The complex "labyrinthine flow of dollars" in the US healthcare system distorts economic relationships, making it challenging to determine the true value and payment mechanisms for AI-enabled services. While a "gigantic bubble" fuels AI investments, the underlying potential to address "unmet needs and staggering inefficiencies" suggests a solid foundation.

Payment and incentives. The question of "Show Me the Money" is critical. While some AI tools receive specific reimbursement (e.g., Viz LVO for stroke triage, prostate cancer risk prediction), many will not, with costs absorbed by health systems hoping for ROI through efficiency and clinician satisfaction. The current fee-for-service model hinders adoption of AI that improves outcomes but doesn't generate billable services. A shift towards "global- and population-based payment systems" would better incentivize AI that enhances quality and cost-effectiveness.

Regulatory tightrope. Regulating AI in healthcare is a "precarious tightrope." The FDA, accustomed to stable drugs and devices, struggles with AI's dynamic, opaque, and multifaceted nature. While high-risk tools (e.g., radiology AI) require FDA oversight, many administrative or decision-support tools fall into a "muddiest area." The FDA is exploring new approaches like "total product lifecycle" and assessing manufacturer culture, but the sheer volume and rapid evolution of AI demand a "layered defense" involving public-private partnerships like the Coalition for Health AI (CHAI) to establish standards and "assurance labs."

9. EHR Titans and the Platform Wars: Epic vs. Oracle in the AI Era

Epic was the wrong strategy with brilliant execution, and Cerner was the right strategy with shitty execution. And now our execution is insane.

Oracle's audacious gamble. Oracle's $28.3 billion acquisition of Cerner in 2022 was a bold move to enter the healthcare EHR market, aiming to combine Cerner's data with Oracle's cloud and AI capabilities. Despite inheriting a "total mess" from Cerner, Oracle Health, led by David Feinberg and Seema Verma, is retooling its EHR with built-in AI, voice commands, and integrated functionalities. Feinberg's vision is to transform the EHR into a "platform" that can host numerous AI applications, much like the Sabre system underpins the aviation industry.

Epic's enduring dominance. Epic, founded by Judy Faulkner in 1979, remains the "King Kong of the healthcare industry," dominating the US market with its integrated, tightly controlled EHR system. With over 250 million patient records and a "culture of fun, creativity, focus," Epic is aggressively pursuing AI, with "more than one hundred AI initiatives already rolled out or in development." Its "Cosmos" database of de-identified patient records facilitates research and treatment recommendations. Epic's strategy is to offer a comprehensive AI portfolio, often bundled, making it the "safe choice" for health systems.

The platform battle. The competition between Epic and Oracle, and the broader AI marketplace, centers on who will become healthcare's dominant AI platform. While Epic's "monopolistic" practices and "stranglehold on the industry" are facing antitrust lawsuits, the trend is towards a more pluralistic digital ecosystem. As integration of third-party AI tools becomes easier, EHR vendors' incumbency advantage may fade. The future likely involves a "sturdy Christmas tree" (EHR) on which many "digital ornaments" (AI apps) can hang, with the market rewarding integrated offerings that solve broad patient and clinician needs.

10. Empowering Patients: AI's Democratizing Force in Health Management

Google gave us access to information. AI gives us clinical thought.

Patient advocacy transformed. Generative AI is empowering patients to manage their own health in unprecedented ways. Patient advocates like Grace Cordovano, who narrowly avoided misdiagnosis of cancer, now use AI to:

• Quickly research clinical trials and specialists.
• Summarize complex medical records and create coherent timelines.
• Draft prior authorizations and other communications.
• Help clients apply for disability benefits and access resources.
  This "hacking" of the healthcare system allows for more informed decision-making and navigation of complex medical journeys.

DIY healthcare and new challenges. AI is enabling a new era of "participatory medicine," where patients can self-diagnose common ailments and receive guidance, potentially avoiding unnecessary doctor visits. Wearable devices and home monitors, coupled with AI, can provide "round-the-clock oversight" and "digital nudges" for healthy behaviors. However, this democratization raises concerns about:

• Misinformation: Patients using general AI platforms may struggle to discern truth from falsehoods.
• Gatekeeping: Intense debates are emerging over whether traditional credentials should still determine who can order tests and authorize treatments.
• Lack of research: There's a critical need for studies on the consequences of patients using AI for self-care.

The future of patient interaction. While AI can act as a powerful coach and information source, the human element remains crucial for profound life decisions and complex emotional conversations. The question of transparency—whether patients should be informed about AI's role in their care—is actively debated, with some studies suggesting patients prefer human-written responses when aware of the author. Ultimately, AI will likely lead to a tiered system where patients can choose between AI-first or human-first care, depending on their needs, preferences, and ability to pay, reshaping the patient-provider relationship.

11. A Golden Age for Healthcare: AI's Transformative Potential and Enduring Human Need

I’m convinced that AI will lift many of the bureaucratic burdens that weigh down today’s medical practice, allowing us to rediscover what makes medicine profoundly satisfying: the privilege of helping people when they’re most vulnerable.

Optimism for a golden age. Despite initial skepticism, the author concludes with "guardedly hopeful" optimism that AI will usher in a "golden age in healthcare." AI's rapid advancements, coupled with thoughtful implementation by clinicians and organizations, are already yielding tangible positive results in areas like scribing and billing. This early buy-in suggests a solid foundation for tackling more ambitious problems, with patient perceptions of AI in medicine being "positive—much more so than their views about AI in general."

Rewiring for transformation. Realizing AI's full potential requires healthcare organizations to "fundamentally rewire themselves." This involves establishing new governance structures, vetting tools, training staff, and addressing bias and privacy concerns. Academic health systems are integrating AI into education (assessment, coaching) and research (testing tools, responsible use). The "Productivity Paradox" reminds us that reimagining workflows is crucial, and AI is poised to create the "bandwidth to tackle the really hard stuff" beyond current tractable problems.

Enduring human element. While AI will undoubtedly transform tasks and roles, the author remains convinced that the core of clinical care—intricate decisions under uncertainty, delicate conversations, hands-on procedures, and the orchestration of complex teams—will always require a human guide. AI will enhance physician capabilities, but the "emotional intelligence to recognize and address their unspoken fears, the leadership skills to orchestrate their care across diverse teams, the patience and wisdom to navigate the inherent uncertainties of medicine and healthcare’s bureaucratic cul-de-sacs, the uniquely human capacity for deep compassion that transcends both the practical and the algorithmic" will remain irreplaceable.

Last updated: March 13, 2026