By: Akanksha
Study Scope: Researchers analyzed nearly $45 billion in 2023 Medicare Part B payments across 0.59 million physicians in 56 medical specialities.1
Overall Breakdown: 24% of Medicare physician payments go to hands-on physical work AI cannot do, 21% to non-physical work AI could potentially replace, and 55% go to office Evaluation and Management (E&M) visits.
Most Vulnerable to AI: Pathology (99% exposure) and Diagnostic Radiology (96% exposure) face the highest risk because their work relies on interpreting slides and scans rather than physical touch.
Most Protected from AI: Dermatology (68% hands-on revenue; only 7% AI-replaceable), Orthopaedic Surgery (53% hands-on; 10% AI-replaceable), and Ophthalmology (45% hands-on; 13% AI-replaceable) are heavily shielded by physical procedures.
Greatest Uncertainty: Primary care fields like Family Medicine (87% E&M visits) and Internal Medicine (90% E&M visits) could be either augmented or replaced depending on healthcare policy and reimbursement.
Also Read: AIIMS Delhi's AI-Generated Radiology Reports: Innovation or Controversy?
In 1986, a 62-year-old diabetic patient presented to their local Emergency Department (ED) with breathlessness. This patient was assessed by an internal medicine doctor working as a locum in the ED. They were admitted to a general medicine doctor who came to do their rounds each morning. In 2023, a 62-year-old diabetic patient presented to the Emergency Department with the same symptoms. They were assessed by an emergency medicine specialist, admitted to a hospitalist service, and their primary care doctor had no involvement. The hospitalist then called a cardiologist, who called in another cardiologist who specialises in structural heart disease. It is now reported that 3 out of 4 doctors are specialists. This has led to specialists accounting for the majority of healthcare expenditure and increasing outpatient attendance.
Between 1933 and 1941, there were 15 recognized specialist boards, including ophthalmology, radiology and pathology. This was due to the explosion in medical knowledge. In 2020, there were nearly 1 million new PubMed-indexed medical articles published. Although there have been amazing advances in medicine, such as joint replacements and targeted cancer treatments, specialization has caused fragmented care and long waiting times for specialist care: 84 days for gynecologist in Boston, and over 100 days for a dermatologist in Poland.
Now, Artificial Intelligence (AI) is changing the future of medical specialists. This is because AI is being used in medical imaging, pathology, and diagnosis.1 As AI can now read scan,1,2 create preliminary diagnoses,3,4 and communicate with patients, there is a huge debate in healthcare about whether AI will soon replace doctors or help to make medical expertise more accessible.
The researchers identified areas where AI cannot be used because software cannot physically assess patients, perform surgery, or perform a biopsy. This means that the amount of hands-on work in a medical specialty will determine how much AI can currently be used.5
In a new paper published in NEJM AI, Maria Polyakova, Ph.D. from Standard Health Policy; Dan Zeltzer from UC Berkeley; undergraduate students Tyler Tam and Benji Bhattacharya from Stanford; and Victoria Udalova from the U.S. Census Bureau identified how much AI can be used in 56 different medical specialties. This has shown that there is a huge difference in the amount of potential AI substitution in each specialty. Areas involving hands-on work using equipment are mostly protected from AI substitution.
However, areas involving data, imaging, and slides are completely vulnerable. It was also found that the longer the postgraduate training period,6 the more protected a specialty is.
The data used in this study was taken from Medicare Part B claims from 2023. This data was sourced from the U.S. Centers for Medicare and Medicaid Services and the National Plan and Provider Enumeration System.1 The researchers categorized the nearly $45 billion that was paid to 0.59 million physicians across all 56 specialties.
Overall, the data showed that:
24% of payments to physicians in Medicare were for hands-on work. This cannot be done by software-based AI.
21% of payments were for non-physical work. This means there is no physical barrier to AI substitution.
55% of payments were for Evaluation and Management (E&M) visits. The impact here depends on how the AI is developed, used, and paid for.
The methodology is very straightforward. Every time a doctor sees a Medicare patient, they bill for the service they provided. The research team can take all these billed services and divide them into categories based on whether they require physical contact.
Routine visits, Evaluation and Management (E&M) visits, were removed, as they involve both talking or thinking and physical examination. The remaining services were divided into ‘physical’ work or ‘non-physical’ work. Physical work included biopsies, injections, and surgery. Anything that did not require physical contact between the doctor and patient was categorized as non-physical work.
This shows the maximum amount of work that can be done by software alone. The figures were rounded to whole numbers to give a clean overview of the workload in each specialty.
According to Polyakova and her colleagues, diagnostic radiology and pathology are most likely to be replaced by AI. From the data in the study, diagnostic radiology has 96% exposure to AI, and pathology has 99% (100% excluding office visits). This is because almost all of their work involves looking at imaging, scans, and tissue slides.
The majority of the 21% of money spent by Medicare with no physical barrier to AI involves imaging, pathology, and lab work. This includes looking at echocardiograms, screening mammograms, CT scans, and biopsy tissue under a microscope.1 However, there are huge benefits to having an AI boom in these specialties. AI classifiers can now identify skin cancer and read scans at a specialist level.2 This helps increase diagnostic accuracy, reduce turnaround times, and improve workflow. Due to the lack of physical barriers in radiology and pathology, these are the two specialties where the idea that AI will replace doctors is a real possibility.
Also Read: Interventional Radiology Work has Become Increasingly Concentrated Among Fewer Radiologists
Surgeons and proceduralists are the least vulnerable to AI substitution. If a specialty involves a lot of physical examinations, surgery, direct patient interaction, and physical intervention, such as working in a burn unit, joint replacements, and bedside procedures, then there will be a physical barrier.
Even if AI can identify skin lesions using an image classifier,2 this only accounts for a small percentage of dermatologist’s revenue. The researchers found that the most hands-on specialties were highly protected. This is because they bill for physical work such as biopsies, excisions, and cryotherapy:
Dermatology bills 68% of its total revenue, 91% excluding routine visits, for physical work. Only 7% of its total work can be done by AI.
Orthopaedic Surgery bills 53% of its revenue, 85% excluding office visits, for physical work. Only 10% of its work is vulnerable to AI.
Ophthalmology bills 45% of its revenue, 78% excluding office visits, for physical work. Only 13% of its work is vulnerable to AI.
This does not necessarily mean that radiologists, pathologists, and other doctors will be completely replaced by AI. As the researchers wrote in their NEJM AI perspective:1
“Physical limits to AI implementation are not permanent: Robotic and hardware-integrated systems may, over time, weaken the barrier in procedural fields, and AI may end up augmenting rather than replacing physicians in every specialty. Yet progress in hardware technology moves more slowly than software, so the current differences between specialties are likely to persist for at least several years.”
Even where AI is strong, human oversight is still required. In specialties such as family practice and internal medicine, where 87-90% of Medicare is billed for E&M visits, doctors must manage multiple interplaying chronic illnesses and perform physical examinations. Across all specialties, 73% of office visits, 74% in family medicine and 77% in internal medicine, are billed at level 4 or 5. This requires moderate- or high-complexity medical decision-making.1 Even during procedures such as a colonoscopy, AI is used as an aid to identify polyps, while the doctor is still in control of the scope.
AI exposure means that some tasks performed by doctors could potentially be handled by AI. AI substitution occurs when AI takes over those tasks, while AI augmentation means AI assists doctors without replacing their clinical roles
The NEJM AI analysis identifies medical services that software could potentially automate. However, it does not measure actual physician job losses or predict when doctors might be replaced. AI's impact will depend on its clinical capabilities, regulations, healthcare infrastructure, reimbursement models and human oversight.
This can raise many questions for medical students. The specialties that are the quickest to enter and most needed to tackle primary care shortages, family and internal medicine, have the least physical protection and the most uncertainty.
As Polyakova told Stanford Health Policy: “Physical limits to AI implementation are not permanent: Robotic and hardware-integrated systems may, over time, weaken the barrier in procedural fields, and AI may end up augmenting rather than replacing physicians in every specialty. Yet progress in hardware technology moves more slowly than software, so the current differences between specialties are likely to persist for at least several years.”
This can affect medical students in two ways:1
They may be discouraged from going into primary care and radiology: If hospitals and insurance companies use centralized software to replace primary care tasks and do not reimburse individual doctors for these tasks, income in these specialties may decrease. This can cause uncertainty and may deter trainees from going into these specialties. This could cause more primary care shortages and encourage students to go into long-training surgical specialties.6
They may be encouraged to go into primary care: If AI used to create documentation, summarize patient histories, and complete routine follow-ups, this can reduce the huge administrative burden causing doctor burnout. If AI makes primary care more efficient and sustainable whilst keeping reimbursements with doctors. This may make primary care very appealing to trainees.
Whether AI will replace or help doctors is a policy decision. This is decided by educators, regulators, and healthcare payers.1
References:
Zeltzer, Dan, Tyler Tam, Benji Bhattacharya, Victoria Udalova, and Maria Polyakova. 2026. "Physical Limits to AI Substitution across Physician Specialties — Implications for the Physician Workforce." NEJM AI 0 (0): AIp2600957.
Esteva, Andre, Brett Kuprel, Roberto A. Novoa, Justin Ko, Susan M. Swetter, Helen M. Blau, and Sebastian Thrun. 2017. "Dermatologist-Level Classification of Skin Cancer with Deep Neural Networks." Nature 542 (7639): 115–18.
Singhal, Karan, Shekoofeh Azizi, Tao Tu, S. Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, et al. 2023. "Large Language Models Encode Clinical Knowledge." Nature 620 (7972): 172–80. https://pubmed.ncbi.nlm.nih.gov/37438534/.
Zeltzer, Dan, Ziv Kugler, Lee Hayat, Tomer Brufman, Ruth Ilan Ber, Keren Leibovich, Tal Beer, et al. 2025. "Comparison of Initial Artificial Intelligence (AI) and Final Physician Recommendations in AI-Assisted Virtual Urgent Care Visits." Annals of Internal Medicine 178 (4): 498–506.
Rose, Christopher, Eric Molins, Anthony Schoeffler, Ryan Kabeer, Morgan R. Frank, and Colin Preiksaitis. 2026. "Will Artificial Intelligence Replace Me? Automation Susceptibility of Emergency Physician Tasks." Annals of Emergency Medicine 88 (5): 551–61. https://pubmed.ncbi.nlm.nih.gov/42687470/.
Gottlieb, Joshua D., Maria Polyakova, Kevin Rinz, Hugh Shiplett, and Victoria Udalova. 2025. "The Earnings and Labor Supply of U.S. Physicians." The Quarterly Journal of Economics 140 (2): 1243–98. https://www.gottlieb.ca/papers/DoctorsIncomePaper.pdf.
(OG/APC/MSM)