For years, entrepreneurs have pursued a familiar path in healthcare: create complex AI tools for medicine, design them for doctors and sell them to hospitals and health systems at high prices.
The model mirrors the way the industry commercialized MRI machines, CT scanners and surgical robots. It remains the dominant approach today, from narrow AI tools built for specific clinical tasks to generative AI products designed solely for physician use.
But a new study challenges the assumption that entrepreneurs today should follow this traditional approach.
Published June 12 in Nature Medicine, the study compared two specialized, physician-facing AI tools, OpenEvidence and UpToDate Expert AI (from WoltersKluwer) with three widely available, general-purpose large language models: Claude Opus 4.6, GPT-5.2 and Gemini 3.1 Pro.
The researchers tested the five tools on 100 medical knowledge questions and multiple real-world clinical scenarios. Then they had practicing clinicians grade the answers without knowing which system had produced them.
The results were striking. On the MedQA portion of the study, the three consumer models (currently available for $20 per month or less) outperformed the physician-facing tools built specifically for clinical use, including one that requires a clinical subscription costing up to $600.
The MedQA numbers: Gemini scored 97.4%, ChatGPT 94.2%, Claude 90.2%, OpenEvidence 89.6% and UpToDate 88.4%.
These findings raise an important question for entrepreneurs: If doctors and patients can obtain comparable expertise from LLMs like ChatGPT, Claude and Gemini for $20 a month or less, why build expensive AI systems only for physicians?
Entrepreneurs should instead consider the larger opportunity: helping 330 million Americans use the tools already available to manage and improve their health.
Monetizing Medical AI: Past And Future
Before ChatGPT launched publicly in November 2022, most AI tools used in medicine were task-specific machine-learning applications called “narrow AI.”
These tools are trained on large sets of “labeled data” (examples where the correct answer is already known). In mammography, for example, developers train an algorithm on thousands of images, some labeled as cancer and others labeled as normal or benign. From those examples, the model identifies dozens of differences between images labeled cancer and those labeled normal or benign. It then uses statistical analysis to provide a risk score, or likelihood, that cancer is present in a mammogram it has never seen.
In a large international evaluation published in Nature, an AI system outperformed the average radiologist and reduced false positives by 5.7% and false negatives by 9.4% in the U.S. dataset.
Still, narrow AI has major limitations. A tool built to evaluate mammograms cannot interpret chest X-rays or identify a rib fracture, even though the lung and rib cage are inches away. Although the FDA’s list of authorized AI-enabled medical devices now includes more than 1,500 entries, adoption in day-to-day medical practice remains uneven, limited by narrow application, high price, reimbursement models and liability concerns.
Generative AI tools, by contrast, are trained on enormous bodies of text and data, including the majority of medical information from textbooks and research studies. With that foundation, they can respond to a much wider range of medical questions. They can help explain imaging reports, assess symptoms, support chronic-disease management, explain medication side effects and help patients understand their chronic diseases. Most important, these powerful applications can be used not only by clinicians, but by patients as well.
Research from OpenAI shows that 55% of U.S. adult users already turn to ChatGPT to check or explore health symptoms.
Despite GenAI’s growing popularity among patients, its commercial use in medical settings has been limited mostly to administrative tasks. Ambient listening tools like Microsoft’s Dragon Copilot are built to capture doctor-patient conversations and convert them into clinical notes. AI coding and revenue-cycle tools are designed to help doctors select billing codes, reduce claims denials and accelerate payment.
But almost no generative AI application has been built for independent patient use.
Looking back, it didn’t make economic sense to create new tools for patients at a time when medical AI mostly consisted of narrow AI models. Obtaining the data sets, validating performance and securing FDA approval is time-consuming and expensive. That forces companies to charge prices beyond what nearly all patients can afford. Further, selling narrow AI tools designed to address a single, specific medical problem directly to patients would also create legal risk if someone were harmed. By selling to hospitals and health systems, manufacturers can place physicians in the role of ultimate decision maker.
This is why the NYU study matters so much. In comparing the five alternatives, the specialized clinical tools carried the medical branding, the professional positioning and the implied promise of greater reliability. Yet the inexpensive, widely available LLMs performed just as well or better.
A general-purpose LLM, unlike a medical device built for a defined clinical task, is not marketed to diagnose or treat a specific condition. Its answers depend on the information entered, the prompt used and the follow-up questions asked.
For those reasons, general-purpose LLMs have not required FDA review and, so far, have encountered limited medico-legal exposure. That flexibility and broad availability make these applications ripe for entrepreneurial innovation.
Patients Are The Larger Market
In the healthcare strategy course that I teach at the Stanford Graduate School of Business, I warn future entrepreneurs that 90% of start-ups fail, not because their products are bad or their leaders weak, but because nascent companies fail to monetize their ideas successfully.
When they graduate, some students will likely pursue the traditional technology model. But I predict that many will decide to figure out how they can create inexpensive educational or assistive tools that allow people to leverage inexpensive LLMs to obtain expertise and improve their health, prevent disease and make the best medical decisions possible.
Two opportunities stand out:
- Combatting chronic disease. Conditions such as diabetes, hypertension, obesity, heart disease, kidney disease, depression and asthma affect 75% of Americans. They are also inadequately managed. In the United States, less than half of patients with hypertension have their blood pressure under control, and an even smaller percentage of patients with diabetes have their blood sugar effectively controlled. Generative AI is well suited to help patients take greater control. It can explain diagnoses and treatments in language people understand. It can remind patients to take medications. More important, it can analyze daily clinical data and alert individuals when their blood pressure or blood glucose is not responding as expected. In today’s model, patients return to their physicians months later for a follow-up visit only to discover that their condition has not improved. With AI support, patients could notify their physicians earlier, allowing medication changes or other interventions to happen before the next scheduled appointment. The same tools could also reassure patients when everything is going according to plan and help avoid unnecessary visits.
- Improving patient access. Patients often need help at night, on weekends or during the long wait for a primary care or specialty appointment. When the doctor’s office is closed, the only available option may be the emergency room, even for problems that could be handled safely in a doctor’s office were it open. Patients are already using large language models this way. The big opportunity is to help them do it better, with safeguards that improve the quality of the information they receive and guidance that helps them know when a symptom requires urgent medical attention. That support would be valuable for parents worried about a child with a fever, adults trying to interpret new symptoms or families caring for an aging parent with worsening pain.
The Next Healthcare AI Unicorn
If inexpensive LLMs are as good as expensive medical AI, then entrepreneurs have an opportunity to create the next healthcare unicorn by helping more Americans apply current GenAI models for medical care rather than by building high-priced specialized applications.
Within this opportunity, many difficult decisions await. The next generation of entrepreneurs will have to decide if the product will be an application, a set of agents or an educational tool. They must figure out whether to sell directly to consumers, charge a subscription fee, contract with employers and insurers, or be rewarded based on measurable clinical improvement? They will need to find a way to test a product for safety and reliability, when accuracy depends on how patients enter information and ask follow-up questions? Finally, they must decide what services or facilities they should invest in to provide clinical assistance for patients 24/7 when the LLM identifies a medical problem.
But one thing is clear: inexpensive generative AI solutions designed for patients remain a wide-open opportunity. The financial and medical logic for pursuing it is simple. Treat patients in a doctor’s office and you help them once. Teach patients to use generative AI effectively, and you improve their ability to manage their health for the rest of their lives.
