I. The opportunity AI affords us to radically improve healthcare and create a trillion dollar company
In the US it is generally recognized that (1) costs are high and (2) outcomes could be much better. I contend that AI will materially help with both.
In my 2016 piece, “20% Doctor Included,” I imagined a twenty five year transition in medicine driven by AI. The change I approximately envisioned is happening faster and in a bigger way, creating a massive opportunity for healthcare players, not dissimilar to the almost trillion dollar market cap Tesla has created in a deeply entrenched industry. Tesla pioneered the electric car world when incumbents wouldn’t. This pioneering created more market value than the value of the ten largest traditional incumbent automobile companies (as of Dec’24). A similar scale opportunity exists in healthcare to create trillions in value.
The practice of medicine has gotten better every decade for the last century. But it has mostly been the practice of medicine with a smattering of the science of medicine. It has been relatively consistent for simple cases and fairly inconsistent in complex cases that matter most. The National Academy of Medicine long ago called for a learning health system but we are far from it.
Because of the cost of expertise in medicine, be it primary care doctors, endocrinologists, oncologists, cardiologists, physical therapists, or cardiac MRI technicians, expertise is hierarchically rationed and hard to access. Primary care is often limited to shorter term, not longer term patient needs. In general for healthcare payors, prioritizing immediate cost control comes at the expense of addressing more impactful lifestyle changes, disease prevention, chronic disease management, early intervention, and ultimately long term cost control. In today’s medical system, primary care that actually champions prevention, behavioral change for better health, healthy habits, nutrition and supplementation, and early intervention is expensive in a world where patients move. Here I mostly address outpatient care, but AI has equal implications in drug discovery and drug costs, diagnostics and inpatient and interventional care.
In the very short term, AI will provide us with near free expertise at the rapidly declining cost of computation in outpatient care. AI conversations are able to engage, diagnose, and monitor patients, all while patiently understanding their issues. Applying advanced diagnostic reasoning, a deep understanding of the latest medical research (including its limitations of applicability), and superior analysis of interactions between conditions, events, lifestyles, and social determinants is now economically feasible. With near infinite expert resources, how should we design the optimal healthcare system? And can we get to such an advantaged system starting from today’s system?
Healthcare is broken today in many ways from lack of adequate care, too often preventable disease burden, lack of accessibility, and physician dissatisfaction. AI can help materially with health care as it is practiced today by physicians, just much better. We can affect healthcare cost, quality and servicing of patient needs, simultaneously with fewer tradeoffs. Can “near free expertise” of AI systems help solve these issues even within the current healthcare paradigm?
II. Addressing skepticism and establishing benchmarks
To those who believe AI expertise will never match that of physicians, an often reflexive, human reaction, it is easy to assuage these concerns by establishing "standard of clinical care" benchmarks for human physicians across multiple areas and allowing AI to meet or exceed these standards, with or without physician assistance. (Furthermore, ample data now shows AI alone exceeds physician diagnostic reasoning.) The National Academy of Sciences' planning committee on reducing per capita healthcare costs and improving clinical outcomes could serve as an appropriate body to set and assess these benchmarks. The predicate device for an AI practicing medicine would be the human diagnostician. Only once the AI exceeds the quality standard and error rates of the median physician in each area should it receive FDA approval and be allowed to “practice medicine,” and even then, initially, under human physician supervision. The concept of requiring new technology to meet or exceed recognized benchmarks before broader deployment is already embedded in our regulatory processes—agencies like the FDA and analogous bodies have long required stringent proof of efficacy and safety before granting approvals. And well beyond science and medicine yet still within the world of AI, the same existed for self-driving cars and the approval thereof. (Historically, we have seen a similar skepticism that proved unfounded with self-driving cars. Early critics doubted that autonomous vehicles could match or exceed human drivers in safety and reliability, yet many real-world and controlled studies have shown they can perform at or above human-level performance in certain conditions.)
III. The “AI INTERN” model: AI practicing under supervision of a human physician
Even after AI meets these benchmarks, a prudent approach would involve treating it as a newly graduated MD intern for 5-7 years, requiring full supervision by an experienced physician to ensure safety and refine its performance. For the following decade, the AI would operate under the supervising physician's license. Instances of this model on a small scale already exist in diabetic retinopathy diagnosis by an AI.
IV. Toward a learning health system & beyond with AI as a foundation
Beyond that, the path to an optimal system made possible by AI progress is clear: embrace AI as the foundation of future medical practice and turn the practice of medicine into the science of medicine. We would, because of aggregation of cases and the varied interventions physicians use and the patients’ medical records, actually get a learning and ever-evolving health system from naturally-run “experiments” i.e., everyday interactions between patient and medical system. The value of observational data, n of 1 trials, mendelian randomization, etc. all can enhance what evidence based medicine looks like with a large, retrospective dataset to apply to evolution of the best practice of medicine.
Beyond a learning health system that improves with use, and free expertise, with AI, each cancer patient would get the benefit of the last thousand peer reviewed publications on their cancer, filtered through AI and human judgement. Each cardiac patient's type and dose of blood thinner can be modulated by the deluge of their genomic or transcriptomic data appropriately applied and simplified by the AI, and titrated by the AI through ongoing patient conversations, appropriately adjusted for the patient's lifestyle and physician approved care plan. With AI, all of medical literature and all of data intensive medicine which would overwhelm human physicians, would be available to apply to every patient care plan.
As to health equity, I claim, given the greater accessibility and lower cost of AI-based care, we now have the greatest opportunity we have had in my lifetime to make equity affordable. As to ethical, legal, and cultural concerns surrounding AI in healthcare, I believe it is unethical to not begin adopting it. Legal and regulatory needs to be worked on over the next few years. Cultural concerns of trust need to be worked out but there is early evidence patients may prefer AI interactions, especially in areas like mental health.
V. The tools are starting to exist in usable form
Here’s the exciting reality: many of the tools to radically improve healthcare already exist, especially in their early forms. These tools can mature rapidly with more frequent and diverse usage and a healthcare company can help make it happen and gain a significant competitive advantage. AI can make diagnostics faster, more accurate, and cheaper. It can automate primary care for a fraction of the cost and a hundred times the patient engagement. It can predict and prevent catastrophic conditions like cardiac events long before they happen.
One company has detected 30,000 cancers early from just the existing EHR data. This could be an indication for an actual blood based or imaging-based cancer test. The obstacles have nothing to do with technology and everything to do with human systems refusing to evolve. And most excitingly, a few years of concentrated effort can validate these new technologies and dramatically improve them while increasing care capacity today under the supervision of a human physician.
Mental health AI: As initial proof, in the NHS, around forty percent of their talking therapies services are using AI to handle intake assessments for mental health patients, but with human supervision, with nearly 400,000 patients processed as of December 2024. In one study using the same AI, recovery rates nearly doubled. Diagnostic accuracy across 100,000 patients was 93% across ten of the most common mental disorders. Can human therapists perform at this level? The answer is no. From their medical device submission form 18 months ago..."After training our ML-model, we conducted an additional evaluation study using a prospective dataset of 773 newly collected patients. Here the model was deployed in a "shadow" mode , i.e. it made predictions however these did not influence anything in the product. We then compared this prediction against the clinician diagnosis both at clinical assessment and their final “ground truth” diagnosis after treatment completion (they used this as the most reliable indicator of the true diagnosis). In these prospective data, the ML-model’s predictions agreed with the final diagnoses in 93.7% of cases. By contrast, the initial human clinical assessments aligned with the final diagnoses only 85.1% of the time. Statistical analysis demonstrated that the model’s diagnostic reliability significantly surpassed that of the clinicians at the initial assessment stage (Odds-ratio=2.08, CI=[1.57, 2.77], χ²=25.96, p<0.00001). These findings underscore the model’s potential to improve diagnostic accuracy and guide patients more reliably into appropriate treatment pathways." As for diversity and inclusion access increased by 29% for minority demographics overall and 179% for extremely marginalized groups like nonbinary patients. The cost? Minimal. Yet we’re still debating whether AI is “ready” to help. And this same mental health AI therapist is starting to deliver therapy today in the UK.
Multi-specialty primary care & engagement through virtual AI: The cost of a full patient visit, including human oversight for very broad primary care in one instance is roughly 5x cheaper than a telehealth visit and 10x cheaper than a typical clinical visit. And the AI can offer more than just primary care or urgent care. It can offer multi-specialty primary care as AI provides much of the expertise to the primary care human physician. Chronic disease primary care, from diabetes to gout including coaching, clinical care, and drug titration will soon come for free. And patient engagement using principles from social media and conversational AI allow for near free expansion to a 100 touchpoints every 100 days for chronic disease (and post acute patients). This makes behavior modification, nutrition, etc pragmatically possible.
Musculoskeletal (MSK) home care: Similarly, post surgical MSK therapy with AI has been so effective in terms of clinical outcomes and patient compliance that one of the leading companies in this space changed its pricing model to ensure it only makes a profit if the patient recovers. This is part of a shift we’re already seeing to broader “AI Care,” where patients recover at home - or engage in prophylactic exercise and care - by directly interfacing with an AI Care agent—removing all barriers to access while keeping the clinician in the loop. Beyond post surgical MSK therapy, the company is now offering therapy for general conditions like “pain” and specialty conditions like "pelvic pain in women" and broadening to other types of AI care.
Cardiac home care: Across almost 300,000 cardiac patients, customers of one cardiology monitoring service take on average six ECG’s per month at home, self administered and initiated at almost zero cost, reducing emergency room visits, and offering early intervention potential for cardiac events. Up to 37 cardiac conditions, soon to be 84, including the most severe ones, can be diagnosed at home by the patient under supervision, with an AI for less than $10 per month. AFIB detection at home is more frequent than a Holter monitor because of patient compliance. These technologies allow for significant patient engagement.
Conflicting tradeoffs between diseases in polychronic patients, who fall across specialties with sometimes conflicting advice, can be reasoned through to decide on the best course of action to recommend to the supervising physician. The most recent AI, OpenAI’s O3 model, has exceeded human multi-step reasoning ability and knowledge in most tests, even against doctorates operating in their speciality. That is an indicator of technology to come.
These are just a few, mostly illustrative but not comprehensive, examples of what is possible. The possibilities are endless and don’t stop here. The future is here but as the saying goes, neither evenly distributed nor universally visible. These are examples in direct care provision, the usually expensive use of human expert labor. Administrative functions from revenue cycle management, prior authorization, scribing and many others are already well in process. In addition to AI tools, many non-AI tools, like algorithmic drug titration, genomic based dosing and drug interaction, also become much easier to use diligently in this AI driven care management environment.
VI. Initial opportunities in risk bearing models of outpatient care provision
On the other hand, certain AI solutions are trying to train physicians and running into resistance. A structural problem in US healthcare is that the average physician doesn't have the time to care about making a 5% more optimal clinical decision or learning some tool to make a materially better one, but an AI can just incorporate this. The reduction of disease burden or provision of near free physician services is thought to reduce the income of some person or some institution because of billing codes anchored in the pre-AI world where many AI services replace human services but are not billable. Value-based care or risk-bearing care is the first place where AI can demonstrate its utility with unlimited and early primary care and near unlimited mental health therapy. Such care, when provided by an AI, is appropriately called “multi-specialty” primary care, a term I am trying to use, as the AI has deep specialty knowledge in each and every specialty and can apply it through integrative primary care.
AI-driven care can reduce disease burden through frequent patient engagement, behavioral change and better nutrition, leading to significant savings on the part of payors. However, these benefits materialize only (1) if patients remain with a payor for several years or (2) if the cost of care is so low that short-term savings outweigh the investment in chronic disease management. AI enables the latter, making it effective even when patients switch providers frequently. Early discovery and intervention for serious conditions would be another major advantage of AI driven engagement, reducing downstream costs.
Eventually, AI based care will provide advantage in fee for service care too as physicians realize they can significantly improve their patients’ care while doubling (or more) their patient panels and not reducing their net income with no increase in workload.
VII. Billing codes: the invisible handcuffs to better care provision
The single biggest structural obstacle to AI adoption is billing and incentive structures. AI can transform primary care, allowing doctors to handle many times more patients while providing 10–100 times more touchpoints. In a specific instance we have seen AI drop a patient’s blood pressure by 30 points with 100 interactions over 90 days—only one of which involved a live text exchange with a doctor. The cost was $15 a month for the patient under a direct primary care model. Though this is not a statistically valid trial, this kind of engagement should be routine for most chronic disease patients.
But under today’s fee-for-service model, this kind of care is penalized because it avoids billable visits. The resultant stroke induced by unmanaged hypertension might increase revenue. If doctors can’t bill for supervising an AI, they won’t use it. Today most billing codes are based on time spent by a physician, setting up the wrong incentives. We need billing codes that treat AI interactions like human ones and based on their impact on patient care. To facilitate adoption we need to do this in a way to maintain stability in physician compensation and bring doctors along. Without keeping this in mind, we’re actively blocking better, cheaper, and more frequent care. And the right system can increase physician income by increasing patient panels with less work by the physician, increase physician satisfaction at knowing the patient is getting better care and frequent touchpoints, decrease overhead, alleviate physician shortages, while reducing downstream lifetime patient costs to the payor. Health systems like Australia’s have significantly more primary care interactions per year. AI interns could immediately provide hundreds of percent more patient interactions without reducing physician income. This surge in capacity could quickly eliminate primary care shortages without increasing physician workload.
VIII. Cultural resistance, the incumbent problem: risk aversion with plausible excuses
AI as an “intern” aligns with medicine’s mentorship culture. Over the next 5–10 years, it can handle routine tasks—patient intake, preliminary diagnosis, drug titration, chronic disease monitoring, and follow-ups—while doctors supervise, make high-level decisions, and approve care plans. Why not provide near-free personalized follow-up after every clinic visit to confirm adherence, titrate doses, incorporate genomics, and offer mental health support?
Resistance persists over fears of “AI hallucination” based on naive testing, yet data show standalone AI diagnostic accuracy at 89%, above the average clinician’s 74%. Human-AI teams drop to 76%, possibly due to human intuition overshadowing AI. In contrast, a custom-built physician application using standard AI models, with guardrails and safety checks, can reduce errors even further. If this trend continues, not using AI may soon be malpractice—violating the Hippocratic oath. Indeed, a more recent study on diagnostic performance confirmed that LLMs - when used alone - significantly outperformed physicians as well as physicians + LLM combined.
Though many claims here lack peer-reviewed studies, waiting risks losing competitive edge and delaying better, low risk care (with human supervision for safety). Deploying AI with human oversight now, while pursuing studies in parallel, will refine and validate these technologies in just a few concentrated years. Systems built for the past, incentives designed to protect fee-for-service, and cultural resistance masquerading as caution, are all reasons that might slow down adoption of AI.
Incumbents resist change because they’re built to protect the status quo and avoid risk. AI governance, legal concerns, and bureaucratic inertia are good excuses to not take any risks. But in the process we do much more patient harm at increased costs and withhold due care by omission instead of commission. We need to take some initiative, with appropriate cautions, to reap the benefits of innovation and this once in a lifetime opportunity to dramatically change healthcare for the better.
As for “excuses” not to use these technologies, we can use robust general purpose AI models, or, we can use poorly implemented “AI” solutions that fail to meet basic clinical standards, like with the now-defunct Babylon Health. Despite marketing itself as an AI-based triage and diagnostic tool, Babylon faced significant clinical accuracy concerns and ultimately crumbled financially, fueling the narrative that “AI in healthcare doesn’t work.” These are lazy, press-headline driven examples. Babylon was never quite a modern AI system and was poorly implemented and aggressively promotional. That is exactly why we must set up standards of performance above those of human physicians that AI systems (and humans) should be required to meet. And guardrails for safety, and technology to triage up to humans, should be built into such systems before they are allowed to practice, and even then, under supervision. Using a standard large LLM is necessary, but nowhere near sufficient.
If you think the big players in healthcare, hospitals, insurers, device manufacturers, will lead the charge on AI, you’re wrong. Innovation doesn’t come from incumbents; it comes from startups. Moderna and BioNTech, not Pfizer, created the mRNA vaccine. Genentech when a startup, not pharma, created the biotechnology field. Tesla innovated electric cars, not the incumbent automakers. Amazon innovated retail, not Walmart. Google Youtube, Netflix and Meta innovated media, not the traditional media players. Uber innovated transport, not Hertz, Avis or taxi companies. Sun and Amazon AWS innovated computing, not IBM.
IX. The real opportunity: using unlimited near free expertise & patient engagement
Let’s talk about what’s possible if we get this right and introduce these technologies with due care. Well before at least 2030, every doctor could have five or more AI interns. These systems won’t just assist—they’ll transform care. Imagine an AI that follows up with every patient three days after a visit, ensuring they’re taking their meds and feeling better. Imagine weekly check-ins for chronic diseases, or continuous monitoring for conditions like heart failure. With frequent engagement with the patient even behavioral change and hence behavioral and nutrition therapy would be possible. And better use of medical literature and genomic medicine will be possible. Many errors, like drug interaction errors and errors of accidental omission or commission, that humans being humans often commit, would be dramatically reduced. And easier and cheaper diagnostic care would be available.
This isn’t hypothetical. One company is using AI to perform cardiac ultrasound without traditional cardiac ultrasound technicians in an FDA-approved manner, making at-home ultrasounds cost effective for screening high risk patients. Patients get better care. The payor wins with a 600% ROI in this case, though hospital systems would see less business.
AI isn’t limited to diagnostics or monitoring. It can deliver personalized physical therapy at home, with compliance rates that no human therapist could match. In one company, 50% of patients did their physical therapy on Christmas Day. Try getting that level of engagement with a human-only in-clinic system with appointments weeks or months out instead of being able to do it impromptu when time permits at 11pm, at the patient's convenience! This impacts outcomes. Other examples of AI for “multi-specialty primary care” and mental health are covered above.
X. What needs to change: starting opportunities & changing structural impediments
Adopting AI in healthcare isn’t just about new tools; it’s about a new mindset. The question isn’t, “what does this technology replace?” it’s, “what does this technology make possible?” The question isn’t “is the technology perfect” but rather “where can it make a difference safely”. The opportunity will be harvested by those who start the learning journey for AI in healthcare provision and for those who participate in improving these AI systems through usage in their environments. Those who wait will fall behind and lose competitive advantage. They will also fail to train their personnel in this new technology which will be the core of healthcare provision in five years. Both the technology and internal processes with healthcare providers and payors need to change and for process and habit change, five years is a short time.
The barriers are real: billing codes, safety and liability concerns, and entrenched cultures. But they’re solvable. Here’s how:
- Create new billing codes for AI-supervised care. Let physicians bill for oversight, under their license, not just face-to-face visits in fee for service models.
- Redefine risk. Stop penalizing innovation. Encourage providers to experiment with AI pilots. No different than training interns and residents, AI should practice under a physician's insurance and under their supervision. It should reduce insurance premiums for providers by reducing physician errors of omissions and commissions, though legal liability is a complex issue that affects physician behaviour in complex ways.
- Educate the workforce. Show clinicians how AI works, not as a competitor, but as a tool that makes them better at providing patient care. Much better and frequent care that allows the human to do 5x more!
- Push for value-based or risk bearing care. In capitated models, AI could thrive because it reduces costs and improves outcomes with aligned financial incentives.
A word on CMS and medicare programs: Providers’ interests today are not well-served by reducing the cost of care. Reducing US healthcare costs more than anything reduces the market size for providers, but there is no shortage of ways to innovate on providers' profitability. It is a shame that we let Medicare Advantage companies risk-score patients when an AI can do the same (with appeal pathways to human physicians). But we let the fox guard the henhouse and the MA plans spend very little time on optimizing care. They mostly pass their risk onto other subcontract providers for primary care.
Other regulatory hurdles also exist but can be overcome over time through FDA approval. Much can be done without regulatory approval because of direct involvement of physicians for oversight.
XI. The humongous opportunity and the cost of doing less/waiting
The biggest risk isn’t adopting AI too quickly; it’s moving too slowly. Every year we delay, patients suffer and costs balloon. Quality can improve with better access while costs decline and the current health professionals supervise much better care. Those that don’t adapt will find that startups and nimble healthcare systems will outpace them. The opportunity is massive: better care, lower costs, more access for everyone, and strong competitive advantages for adopters.
AI isn’t the future of healthcare—it’s the present. The question is whether we have the courage to use AI and who will pioneer this world, just as Tesla pioneered the electric car world when incumbents wouldn’t. This pioneering created more market value than the value of the ten largest traditional incumbent automobile companies combined. The opportunity for an innovator in healthcare is humongous. The evolution, safety and triage features and the full characterization of these AI systems and setting the standards they must meet will take a few years but can be done in far less than five years if we set our minds to it. Process and habit change within organizations will take longer but I believe a 50% lower cost of care provision is possible for both payors and providers and will be a source of great competitive advantage.