Why open-ended visits stay with a doctor
Family medicine starts with a person who does not yet have a diagnosis. Someone books a visit for tiredness, mentions a new pain at the door, and has two chronic conditions already on file. Taking that history, examining the patient, and deciding what to rule out first is judgment under uncertainty. Language models are good at producing a list of possibilities. They are weaker at deciding which possibility deserves a test today and which can wait until the next appointment.
The second reason is physical. Listening to a chest, feeling an abdomen, checking a joint, giving an injection, and closing a small wound are hands-on tasks. The robotics panel above puts this job in the dexterous humanoid tier, which is the hardest class of machine to build and the furthest from routine clinic use. No software release changes that part of the day.
The third reason is continuity. Family physicians manage the same patients for years, weigh what a family can afford or follow through on, and carry legal responsibility for prescribing. Will AI replace family medicine physicians? The task split above says the paperwork moves first and the clinical relationship moves last. Work in the US bears that out on scale: about 107,510 family medicine physicians were employed at a median wage of $244,180, with projected employment growth of 3.3% from 2025 to 2035 (BLS, 2025).
What AI does, assists with, and leaves alone
The work AI can take on by itself is clerical. Ambient tools draft a visit note from the recorded conversation, and they draft replies to routine patient messages before a clinician signs them. Coding and chart summarizing sit in the same group. Of the exposed share of this job, the part AI can handle on its own comes to 0%.
The assist group is larger in clinical weight. Ordering and interpreting tests, building a differential list, checking a medication plan against interactions, and tracking which screenings a patient is overdue for all run faster with a model in the loop, but each one is reviewed and signed off by the doctor. That assist share is 44%. Coverage, printed above as 26, is our estimate of the task time AI can handle today; the coverage method explains what counts.
The rest stays with people. Physical examination and in-office procedures are the clearest cases. So is counseling a patient on diet, exercise, and medication they have not been taking, and explaining a serious result face to face. On the split above, the work that needs a person is 56% of task time, and the headline Still needs a human score is 73 out of 100 (higher is safer).
What the evidence actually covers
The evidence grade for this job is D. No study has yet tested an AI system against family physicians across a real panel of their own patients, over time, with outcomes measured. That is why no parity number appears above. Benchmarks on exam-style questions and on scripted single-visit cases exist, but a licensing-style question is not a Tuesday clinic with sixteen patients, incomplete records, and three people who did not fill their prescriptions.
What would settle it is narrow and measurable: a prospective trial in primary care comparing physician-led care against AI-led care with physician oversight, reporting diagnostic accuracy, missed serious diagnoses, referral rates, and patient follow-through over months rather than minutes. Until something like that is published and repeated, the honest answer is that the clinical half of this job is untested, not proven either way. Our quality parity method sets out what a graded test has to include.
When the picture could change
Most likely between 2041 and 2058 (8 in 10 of our scenarios). The replacement-year method sets out how that window is built.
Two things could pull it earlier. First, documentation tools are already cheap next to a clinician’s time, as the cost panel above shows, so practices adopt them fast and the clerical share of the job keeps shrinking. Second, if regulators and malpractice insurers accept AI-led triage with light physician review, a single doctor could oversee far more visits, which thins entry-level and locum demand before it touches senior roles.
Two things hold it back. Liability and licensure still put a named physician behind every prescription and every missed diagnosis. And the hands-on share of the job would need a dexterous robot in every exam room, which is not close to clinic reality or clinic budgets.
Good to know: the pressure in primary care shows up first as fewer hours spent on notes and more patients per session, not as fewer doctors.
How to stay needed
Lean into the parts of the week that sit in the needs-a-human group: the physical exam and office procedures, behavior-change counseling with patients who are not following a plan, and the difficult conversations about serious results and end-of-life choices. Those are the tasks that hold the visit together when the summary in the chart is wrong.
Two skills are worth adding. One is reviewing AI output well: knowing how a drafted note or a suggested differential goes wrong, and documenting your own reasoning when you overrule it. The other is population management, using registries and risk flags to decide who to call in before they get sick. Both make you the person the tools report to.
If you want to compare the nearby options, look at General Internal Medicine Physicians, Pediatricians, General, and Preventive Medicine Physicians. The diagnosing and treating practitioners family and the healthcare sector page show how this job sits against its neighbors, and the jobs that mostly need a person puts it in wider context. You can also put two jobs side by side on the compare tool, or read how all three questions are scored in our methodology.