Primary Care Is Sickcare. In Ten Years It Won’t Be.

Ask a primary care physician what they do all day and you get an honest, unflattering answer. They renew prescriptions. They route referrals. They document. They see a patient for fifteen minutes, address the one or two complaints that fit in the slot, and defer the rest to a follow-up that half the time never happens.

This is not a failure of talent or vocation. It is what the payment model purchased. We built a system that pays for encounters with sick people, then acts surprised when it produces sickcare. The average PCP has become a highly trained gatekeeper for the pharmacy and the specialist, a job that requires a medical degree to hold and almost none of one to perform.

That job is going to disappear– not the physician, the job. And its disappearance recasts a workforce crisis we have spent a decade measuring as a headcount problem.

The annual visit is a sampling problem

Primary care observes a patient roughly once a year, for a few minutes, using instruments that have not changed much in decades. From that single observation we ask a physician to detect deviation, assess trajectory, and intervene early.

It does not work, and we have known it does not work. A Cochrane review of general health checks in adults found no reduction in morbidity or mortality. When researchers examined Medicare’s own annual wellness visit across 17.8 million beneficiary-years, the apparent benefits disappeared once they accounted for trends that predated the visit.

The reason is in the arithmetic. One observation a year gives you a level, never a trajectory. A single reading with nothing behind it cannot be interpreted, so it gets worked up. That is why about one in five Medicare beneficiaries received a routine test that guidelines advise against during their wellness visit, and why a meaningful share of those tests set off a cascade of follow-ups. The annual visit does not fail because we look at healthy people. It fails because we look at them once.

Everything downstream of that flaw is a reaction. Disease gets detected when it becomes symptomatic, which is to say late. Under those constraints, responding to what has already gone wrong is not a corruption of the job. It is the job.

The binding constraint is observation frequency. Biomarkers are becoming passive to collect and AI is making the resulting stream interpretable, which breaks that constraint from both ends.

Continuous observation collapses the interval

Three things are converging. Physiologic data from wearables and implantables is moving from consumer novelty toward clinical-grade signals. Molecular surveillance of asymptomatic populations has moved out of concept and into large randomized trials, with early evidence that it changes what gets found and when, though nothing yet that settles the question. And AI has crossed the threshold where pattern recognition across a longitudinal record exceeds what a physician can hold in working memory during a fifteen-minute visit.

The measurement site is moving too. Blood pressure, ECG, glucose, oxygen saturation, sleep, weight, and activity are already captured outside the clinic. Home blood collection, portable imaging, and AI-enabled diagnostic devices will extend that list. The home stops being where patients wait between encounters and becomes a sensing layer of the delivery system.

Put those together and the observation interval collapses from one year to continuous. That is not an improvement to the annual physical. It removes the reason the annual physical exists.

The consumer market got there first

The demand side is not waiting. Function Health, Neko Health, Prenuvo, and Fountain Life sell versions of the same product: deeper measurement, earlier detection, longitudinal tracking. Methods run from blood biomarker panels to whole-body MRI to full concierge diagnostic centers. All of it is consumer-paid, largely outside insurance and largely outside evidence-based guidelines.

That is proof of demand. People are paying cash, in volume, for something the reimbursed system does not offer, and that preference is not going to reverse.

It is also a live demonstration of the failure mode, though not evenly across the category. Whole-body imaging in asymptomatic adults is a textbook incidentaloma generator. A tight, targeted biomarker panel carries a different risk profile. A meaningful share of what the imaging-heavy platforms find will trigger a diagnostic odyssey and workups that harm patients and cost money without extending a single life. The category is running the experiment on the central risk of this thesis, in public, at the customer’s expense.

Which raises the convergence question, worth a position rather than a split. Does traditional primary care add prevention, or do the preventive platforms add longitudinal care delivery? The platforms have the consumer relationship but no risk contract, no reimbursement, and no accountability for outcomes. Acquiring those is far harder than acquiring the diagnostics they already have. Convergence happens from the delivery side, and most of today’s preventive platforms end up as measurement layers inside someone else’s system.

What this looks like in 2036

The components move at different speeds, so a single date would be dishonest. Autonomous handling of routine work is a three-to-five-year story and parts of it are shipping now. Panel expansion follows on a five-to-eight-year lag as organizations restructure around it. Payment reform is slowest and will trail the care model by years. The technology will be ready well before the contracts are, and that gap sets the pace.

The routine work goes to machines. Somewhere between 70% and 80% of what primary care does today (refill management, protocol-driven titration, screening scheduling, triage, results interpretation, chronic disease monitoring, and a large share of protocol-eligible acute care) will be handled by AI running against a continuous record, with no physician in the loop on the individual decision. Not physician-assisted. Physician-supervised at the protocol level, which is a different thing. The published evidence base for autonomous clinical action is near-empty today. This is the prediction with the least behind it and the most riding on it.

The panel grows three to five times, and the job inverts. Today a PCP holds 1,500 to 2,500 patients and sees the ones who show up. In 2036 they will hold 5,000 — 10,000 and see the ones the system flags: the diagnostic ambiguities, the multi-morbidity cases where protocols conflict, the patients whose data says one thing and whose life says another. The physician becomes the exception handler and the accountable party. They are there for the conversations no protocol can hold: the diagnosis that changes a life, the treatment that isn’t working, the point where the honest answer is that there is nothing left to try. They step in when the protocol runs out. That is a harder job than the current one. The easy cases were the recovery time.

The specialist referral threshold rises. Fragmentation exists partly because no generalist could hold the full picture or match specialist depth at the point of care. When synthesis gets cheap, the generalist reclaims the coordinating role and referral volume falls for the cases currently referred out of caution rather than necessity. Specialists do not disappear. They get a harder, more concentrated case mix and lose the routine volume that subsidizes it.

Prevention becomes a billable event. Reimbursement shifts from the encounter to the maintained state. You will not be paid to see a diabetic patient. You will be paid for the patient who did not become diabetic, measured against a risk-adjusted counterfactual we are currently bad at computing.

The relationship becomes scarce good. Everything else in the panel decomposes. The longitudinal bond does not. Where the routine is automated, the physician’s remaining monopoly is judgment under uncertainty and the trust required to act on it. That is what people will pay a premium for, and it is why concierge economics migrate downmarket rather than disappear.

The obvious objection to a fivefold panel is that care must get worse. It doesn’t, if the system is disciplined about what it notices. A system that automates the routine can attend to more people while making each one feel more attended to. A system that pings everyone recreates the denominator problem inside the inbox. Patients today experience a system that forgets them between visits. Patients in 2036 should experience one that notices a change worth acting on and reaches out first.

One caveat on the payment prediction, since it is the weakest link in the chain. Paying for prevention means paying for something that did not happen, which requires measuring a counterfactual. We are not good at that yet. Risk adjustment is the closest thing we have, and after two decades it remains contested enough that the industry still argues about what a legitimate score looks like. That does not make the destination wrong. It means the path runs through years of disputed measurement, and anyone modeling a clean transition should discount accordingly.

The shortage inverts

Now multiply, because this is where the argument lands.

The AAMC projects a shortfall of 20,200 to 40,400 primary care physicians by 2036. HRSA designates about 8,500 primary care shortage areas covering more than 90 million people. The crisis is real, it is concentrated in rural and underserved communities, and under the current model it is unfixable. A physician who starts training today reaches independent practice around 2034.

Those projections vary retirement, residency growth, and how much work shifts to nurse practitioners. What they never vary is panel size, because for decades it was bounded by how many patients one human could personally see. Break that bound and a workforce gap in the low tens of percent does not survive a productivity change measured in multiples. We do not have a shortage of primary care physicians. We have a shortage of primary care capacity, and we have been trying to solve it by making more physicians only because the ratio between the two was fixed.

Which makes graduate medical education expansion, the consensus answer to the access crisis, a fifteen-year lever pulled against a problem that may resolve sooner. The physicians will be needed. The job description will not be the one they trained toward.

There are two problems with wearing one name. National capacity is arithmetic, and arithmetic responds to productivity. Local access is geography, and geography does not. A fivefold panel in a well-staffed metro places no physician in a frontier county.

The objection that has to be answered

The strongest argument against all of this is economic, not technological. Prevention has never been the cost story its advocates keep selling. The most-cited modeling exercise found that raising use of twenty proven preventive services to 90% of the eligible population would save more than two million life-years a year and cut personal health spending by 0.2%. Read that twice. Enormous health gain, and a rounding error against a $5 trillion system. Broader analysis reaches the harder version: most preventive interventions add cost rather than save it.

But look at why prevention has been expensive. The cost driver is the denominator. Population-wide screening means testing enormous numbers of people who were never going to develop the condition, then absorbing the workup costs of every false positive. The math fails not because early detection lacks value but because undifferentiated screening spreads the cost across everyone and concentrates the benefit in a few.

Continuous longitudinal data attacks exactly that. When you know an individual’s trajectory rather than their demographic bucket, you screen the people whose data says to screen them. The denominator collapses, pre-test probability rises, the false positive burden falls, and the arithmetic that has doomed prevention for three decades starts to move.

Prevention has never been economical because it has never been targeted, and targeting is what longitudinal data buys.

It is also the part most likely to be wrong, and the consumer platforms are the leading indicator. If continuous surveillance generates more incidentalomas than it prevents late-stage disease, the cost curve gets worse rather than better. That is a scientific question, not a technological one, and it will be settled well inside this window.

The stack

Four layers, and today they barely speak to each other.

• Measurement. Labs, imaging, wearables, genomics, at-home diagnostics, molecular surveillance.
• Intelligence. The longitudinal patient model, risk prediction, clinical reasoning, the agents that execute against protocol.
• Intervention. Physicians, care teams, medications, specialists, navigation, behavior change.
• Economics. Payers, employers, Medicare, value-based contracts, and ultimately who holds the risk.

Nearly every company in this market owns one layer and rents the others badly. The measurement companies have no economics. The risk-bearing entities have thin measurement. The intelligence layer is being built inside both and shared by neither.

Value accrues to whoever spans Intelligence and Economics: the longitudinal patient model plus accountability for the outcome. That pairing is what makes preventive action underwritable, and it is the only combination the others cannot easily buy. Measurement commoditizes as sensors and assays get cheaper. Intervention is labor-constrained and hard to scale. The model and the risk position are where the durable margin sits.

So the investable question is not whether primary care becomes preventive. Demographics, physician supply, and the cost curve make that direction inevitable. The question is who owns the longitudinal record and who is accountable for acting on it. The likeliest winner is not another primary care provider. It is the infrastructure that lets every risk-bearing provider operate this way.

Founder takeaways

1. Treat liability as the product, not the compliance line. Autonomous clinical action at scale requires someone to hold the risk. Commercial carriers are writing AI exclusions into standard liability forms rather than writing coverage, so the gap is widening, not closing. The company that can warehouse that risk, with a protocol library, an audit trail, and a captive or stop-loss structure, does not exist yet.
2. Underwrite the counterfactual. Outcome-based payment collapses without credible risk-adjusted prediction of what would otherwise have happened. Whoever makes that measurable and defensible makes preventive contracting possible.
3. Target ruthlessly. Universal screening is a losing business and always has been. The wedge is precision about who to test, when, and why. Anything that raises pre-test probability is worth more than anything that raises test sensitivity.
4. Build for the exception. The valuable product decides which cases reach a human. Routing, escalation, and confidence calibration are the durable primitives, and they get harder as autonomy rises rather than easier.

None of these work if you automate the relationship. The trust layer is the last defensible position in medicine, and products that erode it to gain efficiency will win pilots and lose renewals.

Primary care spent two decades trying to make an obsolete unit of production more efficient. The next ten will retire it. What survives is the part that was always the point: someone who knows you, paying attention, and accountable when it matters.