Which physician tasks have a physical barrier to AI? Medicare payments split 24% hands-on, 21% exposed, 55% uncertain
A US analysis classified nearly $45 billion in 2023 Medicare payments to 590,000 physicians. It maps tasks that software cannot physically perform today, but billing shares are not job-loss forecasts and office visits remain the largest unresolved category.
The short version
- What happened
- Researchers classified nearly $45 billion in 2023 Medicare payments to 590,000 physicians in 56 specialties, setting office visits aside as dependent on design, deployment and payment choices.
- Who it affects
- Trainees should distinguish changing task bundles from predictions that an entire specialty will vanish.
- What remains unproven
- Payment shares are not physician time, jobs, wages, encounters or patient outcomes.
Keep in mind
Payment shares are not physician time, jobs, wages, encounters or patient outcomes.
One year of US fee-for-service Medicare does not represent all patients or health systems.
Office visits—55% of the payment pool—remain explicitly uncertain.
The accessible source lacks the full code-level appendix, DOI and disclosure details.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The finding: physical work limits substitution, while most payments remain unresolved
A new US analysis asks a narrower and more answerable question than whether artificial intelligence will replace doctors: which billed services require a clinician to touch a patient or manipulate an instrument? Researchers classified nearly $45 billion in 2023 Medicare payments to 590,000 physicians across 56 specialties. Across that payment pool, 24% went to hands-on services that software alone cannot currently perform, 21% to services the authors considered potentially performable by AI, and 55% to office visits whose exposure depends on how systems are designed, deployed and reimbursed.
The 55% uncertain share is the most important result. It prevents the analysis from supporting a simple claim that one fifth of physician work will disappear or that one quarter is permanently safe. A visit can include history-taking, examination, judgement, communication, documentation and procedures. AI might automate one component, assist another or add new verification work without replacing the encounter. Payment categories do not reveal that internal division. The study maps a present physical boundary around some tasks; it does not forecast employment, wages, access, productivity or patient outcomes.[1]
What was counted: one year of Medicare payments, not time or jobs
The researchers used 2023 Medicare claims covering nearly $45 billion paid to about 590,000 physicians. They grouped physicians into 56 specialties and classified billed services according to whether physical contact was required. Office visits were deliberately set aside because their future depends on implementation choices. This service-level approach is more grounded than asking a language model to rank whole occupations: it begins with actual billed activity and recognises that a specialty is a bundle of different tasks. The denominator, however, is payment value rather than hours, encounters, clinicians or patients.
That distinction changes interpretation. A highly reimbursed procedure can account for a large payment share without consuming the same share of working time. Conversely, counselling, coordination and documentation may absorb substantial time while appearing differently in claims. Medicare payment also reflects policy, coding and relative values, not only clinical effort. The analysis covers US fee-for-service Medicare in one year, so it does not represent paediatrics, privately insured populations, other national health systems or work that is not separately billed. It should be read as a financial map of coded services, not a census of everything physicians do.[1]
The distinction that matters
Three payment categories, three different claims
The analysis classified billed services by today's physical barrier; it did not predict how many jobs will remain.
Hands-on
Requires physical contact
24% of payments
Potentially exposed
Could be performed without touch
21% of payments
Uncertain visits
Depends on design and payment
55% of payments
Not measured
Jobs, hours and patient outcomes
No displacement forecast
Procedural specialties had larger physical barriers in the 15-specialty comparison
In the university's account of the 15 large specialties shown in the paper, dermatology received 68% of its tracked revenue from physical procedures, orthopaedic surgery 53% and ophthalmology 45%. Only 7% to 13% of payments in those specialties were assigned to work that AI could potentially perform outright. The classification captures an intuitive constraint: software can analyse a digital image or draft a note, but it cannot on its own palpate tissue, take a biopsy, set a bone or operate an instrument at the bedside. Current robotics do not erase that distinction at scale.
Even here, 'protected' would overstate the evidence. AI can change referral patterns, pre-procedure planning, image interpretation, documentation, scheduling and the number of cases a clinician handles. A procedure may remain physical while surrounding work is reorganised. Hardware and robotic systems could also move the boundary over time, although the authors argue that hardware progresses more slowly than software. The study therefore identifies a near-term physical limit to direct software substitution, not immunity from technological change. Training and workforce plans should consider task redesign around the procedure as well as the procedure itself.[1]
Radiology and pathology appeared exposed because their billable work is already digital
At the other end of the comparison, diagnostic radiology was classified as 96% exposed and pathology as 99% because much of the coded service consists of interpreting images or slides rather than physically interacting with a patient. Those figures describe the payment categories under the authors' framework. They do not show that current AI can independently perform 96% or 99% of the work to an acceptable clinical standard, obtain regulatory approval, manage unusual cases, communicate uncertainty, integrate history or accept legal responsibility. Technical possibility and deployable substitution are different thresholds.
These specialties also illustrate why task exposure need not equal job loss. If an AI system handles a routine first pass, specialists may review more cases, focus on complex findings, oversee quality, communicate with care teams or expand access where staff are scarce. Alternatively, organisations could reduce staffing or shift bargaining power if payment and liability rules allow. Claims data cannot choose between those futures. The direction depends on accuracy across real populations, workflow integration, reimbursement, professional standards, capital costs, regulation and the response of demand when interpretation becomes faster or cheaper.[1]
Primary care sits in the uncertain middle, where policy may matter most
Primary care fields have relatively little separately coded procedural work, but that does not make the entire role a text-generation task. Family and internal-medicine visits combine symptom history, physical examination, longitudinal knowledge, medication management, prevention, emotional context and coordination across services. The researchers placed office visits in the uncertain category precisely because AI could be a substitute for some components or a complement that gives clinicians more time with patients. Payment rules could reward either outcome: reimbursement for human oversight may preserve a team model, while incentives focused only on throughput could encourage more substitution.
This uncertainty matters for access and training. Primary care is both a common entry route into medicine and a field facing shortages in many communities. Cutting training places on the assumption of rapid substitution could worsen access if the technology disappoints or increases demand. Ignoring automation could leave clinicians unprepared to supervise systems that draft notes, triage messages or propose plans. A prudent response is task-level education: verification, escalation, communication of uncertainty, data governance and recognising when a digital recommendation conflicts with examination findings or knowledge of the patient.[1]
What workforce planners can use—and what would change the assessment
The analysis gives workforce planners a transparent first filter: identify tasks that require embodiment before projecting what software can replace. It also warns employers and trainees against treating a specialty-level label as destiny. Organisations should measure changes in time per task, case volume, error rates, access, wages and staffing after deployment. They should separate a tool's technical capability from its authorised scope and from whether a clinician remains accountable. Workers need visibility into which tasks are being redesigned and whether productivity gains are shared through lower workload, more patient time or simply fewer positions.
Confidence in substitution forecasts would require code-level methods and sensitivity analyses, data from multiple payers and countries, direct measures of clinician time, and longitudinal evidence after actual AI adoption. Researchers should test whether exposure predicts hiring, hours, productivity or outcomes, and whether the effect differs by rurality, patient complexity and organisational ownership. The accessible Stanford account does not provide the article's DOI, detailed appendix, funding statement or author disclosures, limiting a full independent methods audit. For now, the percentages are a useful description of 2023 Medicare payments, not a deadline for the disappearance of any profession.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring automation: compare the finding and its evidence limits.
If robots can do much physical work, why are so few tasks economical to automate?
Anthropic researchers rate 7,594 physical job tasks and estimate that robots can perform 74% in some setting, yet are currently cheaper than people for only 0.3% of all work. The gap between technical exposure and economic adoption is the result that matters for workers.
Read the next reportExplore the full source and evidence record
What this means for people
- Trainees should distinguish changing task bundles from predictions that an entire specialty will vanish.
- Clinicians in exposed fields may see more verification and exception handling rather than simple replacement.
- Patients could gain access or lose continuity depending on deployment, reimbursement and accountability choices.
Global context
Occupational AI studies often assign one exposure score to a job title. This analysis instead treats medicine as a bundle of billed tasks and adds an embodiment constraint. That is a useful correction, but countries with different payment systems, scopes of practice and workforce shortages may experience different outcomes even if the technical capability is identical.
What the evidence does not yet show
- Payment shares are not physician time, jobs, wages, encounters or patient outcomes.
- One year of US fee-for-service Medicare does not represent all patients or health systems.
- Office visits—55% of the payment pool—remain explicitly uncertain.
- The accessible source lacks the full code-level appendix, DOI and disclosure details.
What to watch next
- Publication of the full methods and code classification.
- Longitudinal hiring and workload data after adoption.
- Reimbursement rules for AI-assisted care.
- Robotics that change today's physical barrier.
Living evidence record
Impact record IAI-0E8ZDJF
Evidence stage
Observed
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
5 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Stanford Health Policy published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
How much physician work has a present-day physical barrier to software substitution, using service-level Medicare payments across specialties as a task map
Topics: Physicians · Health workforce · Task exposure · Medicare · Automation · Professional skills
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.