Why this work stays with a physician
An internist’s day is built around uncertainty. A patient arrives with fatigue, a cough, three long-term conditions and six medications. Someone has to decide what matters most, what can wait, and what to do today. That judgment carries clinical and legal responsibility, and responsibility does not transfer cleanly to software.
Two parts of the job show why. Examining a patient in the room — listening to the chest, pressing on an abdomen, watching how a person moves and answers — produces information that never reaches the chart. Setting and adjusting a long-term treatment plan is the other: weighing risk against side effects, cost, home circumstances and what the patient is actually willing to do.
AI is strongest on the work that wraps around those moments. It drafts the note, pulls the relevant history out of years of records, and flags results that sit outside the normal range. That is real time saved. The honest version of the story for internal medicine is task erosion inside a stable job, not the job going away. The Bureau of Labor Statistics counts about 67,150 general internal medicine physicians in the United States, with median pay of $256,560 and projected employment growth of 4% from 2025 to 2035 (BLS, 2025).
What AI does, what it helps with, what it leaves alone
Start with the share AI can run on its own: 7%. These are the contained, text-shaped pieces of the job — turning a recorded visit into a structured note, and assembling a discharge or referral summary from material that already exists in the record. The output still gets read and signed by the physician, but the first draft no longer has to be typed.
The bigger slice is assistance: 52%. Here the tools sit beside the clinician. They surface a differential the internist may not have considered, check a medication list for interactions, and sort incoming labs and imaging reports by urgency. Our coverage score — can AI do it? — is 31, and how coverage is measured explains why assistance counts differently from full automation.
Then there is the work that stays human: 41%. Physical examination, telling a patient and a family that the news is bad, and negotiating a plan a person will stick to. Supervising residents and other clinicians belongs here too. The headline figure above is 70 out of 100 (higher is safer), and our method sets out how the three questions combine.
What the evidence can and cannot settle
Evidence grade for this job: D. Grade D means no direct, like-for-like test of AI against general internal medicine physicians doing this job has been measured in our evidence set. So no parity number is given here, and any figure claiming one should be treated with care.
Plenty of studies score models on exam questions or written vignettes. Those are useful, but they are not the job. A vignette arrives tidy, with the relevant facts already selected. A clinic patient does not. What would settle the question is a prospective study in real practice: AI-led assessment compared with physician-led assessment on the same undifferentiated patients, measured on diagnostic accuracy, downstream testing, harm and patient outcomes over months, not minutes. Until that exists, the grade stays where it is. You can see how other jobs are graded in the full rankings.
When the picture could shift
Most likely between 2041 and 2055 (8 in 10 of our scenarios). The replacement-year method explains exactly what that window measures and how the scenarios are built.
Two things could pull it earlier. Ambient documentation and decision support are already being bought at scale by health systems, and the running cost of the software sits far below the cost of clinician time, as the cost panel above shows. Each wave of adoption also produces training data from real encounters, which is the one input models have been short of.
Two things hold it back. A meaningful part of this job is physical and bedside, and the robotics tier it would take to cover that work is a dexterous humanoid — nothing close to deployable in a clinic today. The second brake is accountability: liability, licensing and malpractice rules all attach to a named person. Those change slowly, and they are the reason an internist signs the chart even when a model wrote the first version of it.
How to stay needed in internal medicine
Lean into the parts of the job that resist handover. The hands-on physical examination. The difficult conversation — prognosis, goals of care, a plan a patient will actually follow. And the supervision and teaching of residents, nurse practitioners and physician assistants, which expands as more of the routine work gets delegated.
Two skills are worth building. First, reading AI output critically: knowing where a model drifts, when its confidence is unearned, and how to document your own reasoning when you overrule it. Second, running complex, multi-condition care across a panel — the coordination work that no single tool owns.
What to do: spend one week logging which parts of your day a drafting or triage tool could have handled, and which parts needed you in the room.
Nearby work scores on similar lines. Compare the task mix with Family Medicine Physicians, Hospitalists and Emergency Medicine Physicians, or put any two of them side by side with the comparison tool. For the wider picture, see the diagnosing and treating practitioners family, the healthcare sector page, or the list of jobs that mostly need a person.