LO
Get told when a verified record lands on this occupation →
General practitioner / primary care doctor
The doctor a person sees first: works out what is wrong when the problem has not been sorted into a specialty yet, decides what to do about it, and carries the consequence of being wrong.
This is not a probability of losing your job. It combines how much of the role's task load is exposed to automation with how far adoption has actually gone — useful for comparing occupations on one consistent basis, and for nothing else.
Written for doctors in general and primary care — the consultation where the problem arrives undifferentiated. Hospital specialists, surgeons and radiologists have different task mixes and radiology has its own page here. Health systems differ enormously in who holds the prescribing pen and who controls the referral, and that is the single biggest thing this page cannot generalise across.
The evidence base holds verified records for other occupations, but not one for this one yet. Until it does, the analysis below is reasoning about task structure and known technical capability — for this job in particular it is not backed by traceable sources, and we would rather say so than cite things we have not verified. An empty section here is a gap in our coverage, not a finding about the work.
What is actually changing#
The unit of analysis is the task, not the job title. A role is not replaced — its task mix shifts.
Is this your job? Say so and this page narrows to your share of it.
A job title is a bundle of tasks bought together, and no two people hold the same bundle. Nothing is sent anywhere — it stays in this browser.
Writing the visit down
Automating≈ Platform inferenceThe note: what the patient said, what you found, what you thought, what you did — written to a standard that another clinician and, later, a lawyer can both read.
Ambient documentation has the shape that automation needs and almost nothing else in medicine has: the input is a conversation that just happened, the output is prose in a known format, and a clinician reviews and signs every one before it counts. The reviewer is the referee, which is why this moved while diagnosis did not — and it is the same tool being sold to nurses, therapists and assistants, so the adoption curve is one curve, not five.
Time saved on the note is not time given back to the doctor — in most systems it is absorbed by seeing more patients, and whether it lands as relief or as throughput is a management decision made after the tool arrives, not a property of the tool. Nor does it touch what the note is for: a record written to be defensible is a different document from one written to be useful, and automation has so far made the first cheaper without making the second better.
Working out what it is
New task≈ Platform inferenceTaking a story that does not fit a textbook, deciding which two or three things it could be, and choosing which one to act on before you can be sure.
Models score well on written vignettes, and a vignette is a case someone has already cleaned: the relevant facts are present, the irrelevant ones are absent, and somebody decided where the story starts. A real consultation is the opposite — the patient volunteers the wrong thing first, omits the thing that matters, and the doctor's actual skill is in what they ask next. That step is interactive and undertested, which is why this is emerging rather than automating.
A tool being good at the diagnosis does not put it in the room, and in most systems what reaches the consultation is decided by procurement, liability and the electronic record vendor rather than by accuracy. Read the direction as a statement about where the capability is pointing, not as a forecast about your clinic — and note that the same tool arriving can change the job without changing who does it, by turning the doctor into the person who overrides a suggestion and documents why.
Putting hands on the patient
Still human-led≈ Platform inferenceLooking, feeling, listening — and the part nobody writes down: noticing that this person looks unwell in a way the numbers do not yet show.
The examination is a physical act performed on a person who did not consent to a machine doing it, and its most valuable output is the least structured: the general impression that makes an experienced clinician escalate before any test justifies it. Devices can capture individual signals well and several do; what they cannot do is decide which signal to go looking for, which is the whole of the examination.
Hard to automate is not the same as valued: in several systems the examination is already the part squeezed hardest by appointment length, and a task can be eroded by the clock without any technology touching it. Remote consultation has removed it entirely from a growing share of visits, and that happened for reasons of cost and access rather than capability.
Deciding it together, and saying the hard thing
Still human-led≈ Platform inferenceExplaining a trade-off to someone frightened, finding out what they actually want, and telling people news they did not come in expecting.
What makes this work is not the information — the information is already online and the patient has usually read it. It is that a specific person takes responsibility in front of them for a recommendation that could be wrong, and stays in the room afterwards. Delegating the sentence to a machine removes the thing that made it bearable, which is why even systems that automate the note do not automate this.
This being human does not protect the headcount, because it is the part of the visit that scales worst and is therefore the first to be rationed — pushed to a nurse, a leaflet, a follow-up call that does not happen. Read it as a statement about who must do it, not about how many minutes anyone will be given to do it in.
Holding the prescribing pen
Still human-led≈ Platform inferenceSigning for a prescription, a referral, a sick note or a test — the acts where the doctor's name is what makes the thing valid.
This is not held by skill, it is held by law and by insurance: the signature is a legally attributed act, and in every major system the person who signs is the person who answers for it. That makes it the most durable task on this page and also the most contingent — it is durable exactly as long as the rule stays.
Protection by regulation is a policy choice, and policy changes: several jurisdictions have already widened who may prescribe, and each widening moved work without removing the pen. The safe thing here is the signature, not the hours behind it — and a system can keep the doctor's name on the prescription while moving the consultation that produced it to somebody cheaper.
Which technologies matter here#
Four separate signals. They are deliberately not added together — a job exposed to two technologies is not twice as exposed.
How it got here#
The index is not a static number. This is where it would have sat at each capability checkpoint since ChatGPT — reconstructed, and labelled as such.
A low start that reflects where medicine actually was: in late 2022 the tools in a consulting room were a template library and a drug-interaction checker, neither of which had moved in a decade. The climb through 2023-2025 is one task and one task only — the note — arriving in a form clinicians would tolerate, because a draft a doctor signs has a referee and a diagnosis suggestion does not. It goes flat early and stays flat, and the flattening is the finding: the tasks that carry the weight of this job are the examination, the decision made with the patient, and the signature, and none of them has a mechanism to move that does not require a change in law. Read the height as documentation, not as medicine.
A flat line is not a forecast of safety. It says which tasks automation has reached so far — the occupations that moved least here are the ones where the constraint is physical or regulatory, and both of those can change.
Recent changes#
No verified events recorded yet.
This section will fill from the monitoring pipeline as events are collected, de-duplicated, graded and linked to the tasks above. An empty list here means we have not verified anything — it does not mean nothing is happening.
"We found no news" is not the same as "you are safe."
What this means for you#
The entry ladder in medicine is unusually well protected by training requirements, so the first thing to change for you is not whether you get a job but what the job consists of — expect the documentation to be drafted for you from day one, and expect to be assessed on whether you caught what the draft got wrong. The skill that used to be built by writing a thousand notes has to be built some other way, and nobody has said how.
Your leverage is in the two tasks that do not compress: the examination and the conversation where something is decided. The risk is not that they get automated, it is that they get rationed while the parts that were automated free up minutes that management books as capacity. The thing worth defending in a negotiation is appointment length, not scope of practice.
Your options#
Four directions, each with its real constraints and one thing you can test this week. Continuing as you are is a legitimate choice — it just has to be a chosen one.
Become the person who audits what the tools wrote
Every system deploying ambient documentation acquires a new problem within a year: nobody knows how often the draft was wrong in a way the signer did not catch. Someone has to design that check, and it has to be a clinician.
It is unpaid in most job plans and invisible in most appraisals, so it has to be negotiated as sessions rather than absorbed.
Take ten of your own signed notes from the last month and read them as if you had not been there. Count how many you would have to ask a question about.
Move toward the undifferentiated end
The further a problem is from being labelled, the less any tool can do with it — and the more of the job is deciding what question to ask. That end of the work is also where the shortage is.
It is the harder end, it carries more uncertainty and more risk, and in fee-based systems it is often paid less than procedural work.
Look at your last twenty consultations and mark which arrived with a label already attached. That ratio is your exposure.
Clinical safety for the tools themselves
Regulations in several markets now require a named person with the competence and the authority to oversee a high-risk clinical system. Health systems are discovering they do not have that person.
It moves you away from patients, and the pay in safety roles is usually below clinical pay.
Find out who in your organisation signs off a clinical decision-support tool before it goes live. If the answer takes more than two calls, that is the gap.
Common questions#
Not as a single event, and the more useful question is which parts of the visit move. On the evidence we hold, the documentation is moving first and fastest because a clinician reviews every draft, which gives the automation a referee. The examination and the conversation where something is decided are not moving, and the prescribing signature is held in place by law rather than by skill. What that adds up to is a job with the same name and a different inside — which is a real change and not the one the question usually means.
We do not answer this with a date, and any source that does is guessing. The signal to watch in this occupation is specific and you can check it yourself: how much of your appointment is spent on things a machine now drafts, and whether the minutes that freed up were given back to the consultation or booked as extra patients. The second question is answered by your rota, not by the technology — and it is the one that will actually change your working life.
Careful with what the comparison measures. The published scores are mostly on written cases, and a written case has already been cleaned up by whoever wrote it: the relevant facts are in, the irrelevant ones are out, and someone chose where the story begins. A real consultation hands you none of that. Until a benchmark makes the model do the asking — decide what to enquire about next from an incomplete story — the scores tell you about a different task from the one you do.
It is not what the evidence on this site shows anywhere else: in the deployments we have verified, time saved on documentation has been absorbed by volume rather than by headcount. But the honest answer is that this is a management decision made after the tool lands, not a property of the tool — and it is made in a budget meeting you are probably not in. The thing worth watching is whether appointment length moves, because that is where the decision becomes visible.
Method and sources#
- Assessment date
- 2026-09-14
- Basis of the task judgements
- 0 evidence-backed · 5 platform inference · 0 not enough evidence
- Verified events
- 0