The Unit Problem

On what happens when the symbol outlives the function it was supposed to represent

In the early hours of a Thursday morning in April 2026, I could not walk from my chair to bed. I live in Brisbane and have an artificial hip. I opened my laptop and described my symptoms to an LLM.

It did not tell me to wait and see. It was quite strict about that. The serious concerns it laid out were late-stage post-operative infection, fracture, damage to the prosthesis, or loosening — each capable of becoming significantly worse overnight. The less alarming possibilities were bursitis, synovitis, or a torn muscle. It recommended emergency department presentation and was clear that going to bed to see if things improved was not the right call. I went.

At the ED I waited. An X-ray was taken. Blood was drawn. A GP consult occurred. The GP explained, carefully, that their role was not to diagnose my problem — it was to determine whether I could safely go home. After three hours, exhausted and in significant pain, with no one having formally discharged me, I left at five in the morning. The results were not explained. I had no diagnosis.

The next day a GP teleconsult also produced no diagnosis.

Two days after the original presentation I reviewed the trajectory of my condition with the LLM — the acute onset, the three-day resolution pattern, the absence of fever, the prosthetic joint context. It identified a probable bursitis flare. I had been almost pain-free for twenty-four hours by then. The specialist appointment was still a week away.

The LLM was not given a stethoscope. It was given a conversation. It did more diagnostic work than the institution.

This is not a story about a bad hospital or a bad GP. Everyone in that ED was doing their job correctly by the standard their system defines. That is exactly the problem.

The bundle

The GP is not a profession. It is a bundle.

In the late 19th century, the functions now assigned to a GP — triage, history-taking, differential diagnosis, examination, prescription authority, referral, and chronic disease management — were wrapped into a single credentialled role. Not because one person performs all of those functions best. Because in 1890 a single itinerant physician was the only affordable delivery mechanism available. The bundle was a resource constraint. It was never a design principle.

A hundred and thirty-five years later, we are still running the constraint as if it were architecture.

The bundle now has a liability structure, professional identity, regulatory framework, union representation, and a political lobby. Medical schools produce it at controlled scarcity. Insurers bill against it as the unit of account. Proposing disaggregation is no longer an operational argument — it is a political challenge to a profession. The symbol has absorbed the substance. Challenging the bundle reads as attacking the doctor.

This is how most institutional bottlenecks survive. Not through conspiracy, but through the accumulated weight of interests that benefit from the unit remaining intact — none of whom are the patient or the taxpayer.

What the LLM already does better, faster, and for almost nothing

In the five years before that very painful and frightening Thursday morning, roughly 80% of my GP visits were for two things: repeat prescriptions and explanation of blood test results. Both are tasks an LLM performs instantly, at any hour, without requiring me to sit in a waiting room with sick people for twenty minutes longer than the agreed appointment time before paying $100 for the privilege.

That 80% is only valid as my personal experience and is probably not completely generalisable. I am health-literate, comfortable with technology, and managing defined chronic conditions — a profile that skews toward the replaceable end of the distribution. The research-based floor for the general population is more conservative. Australia's most comprehensive study of GP encounter content, the BEACH program — nearly 1.8 million GP-patient encounter records collected over 18 years — found that the five most frequent reasons patients visited a GP were: to obtain a prescription, for review or follow-up, for upper respiratory tract infection, for immunisation, and to discuss results. The top two reasons by volume are directly and completely LLM-replaceable today. The third is largely so. That is the empirical basis for a floor estimate of 30–40% of consultations amenable to LLM-first triage without physical examination. It is a floor, not a ceiling, and it is conservative by design.

In 2023–24, Australia spent $14.6 billion on unreferred medical services, primarily general practice, across 167.2 million GP attendances — approximately $1.2 billion per month, or roughly 0.54% of GDP in direct consultation costs alone. Applying the conservative 30–40% floor produces an addressable fraction of $360–480 million per month. These figures are illustrative order-of-magnitude estimates, not precise calculations, and should be read as such — but their direction is not in doubt and the true figure is likely higher, not lower.

The indirect costs compound the picture considerably. Avoidable emergency department presentations — cases that accessible after-hours LLM reasoning would have managed at home or redirected to a pharmacist — cost the hospital system $800 to $1,200 per presentation. Diagnostic delays converting manageable conditions into acute presentations carry their own downstream treatment costs. Productivity losses from mandatory in-person attendance during business hours, for consultations that require neither physical examination nor a credentialled professional, add further. The combined direct and indirect cost of running the 19th-century bundle as 21st-century architecture plausibly exceeds 1% of Australian GDP annually. That figure compounds every year the system does not move. It is back-of-envelope arithmetic, and the envelope is not small.

The comparison standard is not perfection. It is the GP. Studies of diagnostic accuracy in primary care consistently find error rates of 10–15%. These errors are distributed across individual practitioners, not systematically detectable in advance, and produce litigation years after the fact if at all. By 2024, LLMs were matching or exceeding GP-level diagnostic accuracy on structured clinical vignettes. By 2025, that finding was replicating across real-world triage scenarios. A Pakistan trial in early 2025 found physicians with LLM assistance produced a 27.5% improvement in diagnostic performance. A 2,069-patient trial published in Nature Medicine found LLM pre-assessment reduced specialist consultation duration by more than 28%. A 2,113-patient randomised controlled trial ran across eleven Chinese provinces, including rural areas with constrained infrastructure, testing LLM-first primary care delivery. These are published, peer-reviewed results.

The trials are real. They are simply not happening in Australia, the UK, the United States, or Canada. Every deployment-scale trial of LLM-first primary care has been conducted in countries where the GP guild has less institutional political power. That is probably not a coincidence.

The alternative is a disaggregated system

Six components. None of them new technology.

A primary reasoning layer — an LLM handling triage, differential diagnosis, treatment recommendation, and test interpretation at near-zero marginal cost per interaction. A second LLM with genuine architectural diversity as an error-checking layer, flagging out-of-distribution presentations for human escalation. A nurse practitioner or paramedic performing the physical examination the roughly 15% of cases require, directed by the LLM's examination brief. A pharmacist verifying the proposed treatment before dispensing, replacing GP script authorisation. The specialist layer, unchanged, now receiving better-prepared referrals from a system that has already synthesised the patient's full history. A public accountability layer: outcome dashboards by system, patient choice between competing providers, accuracy as the feedback mechanism.

The GP role that handles genuine complexity — multimorbidity management, longitudinal care, the patient presenting with something other than what they say they're presenting with — is preserved. What is retired is the GP as the default atomic unit for every interaction. The triage function. The repeat script. The test result interpretation at midnight. The determination, at five in the morning, that a man with a painful prosthetic hip can safely go home without a diagnosis.

That last function cost the Australian health system approximately $400 and produced nothing. The LLM did it in four minutes.

The steelman, given its due

The case for the current system is not stupid. Three arguments deserve honest treatment.

The first concerns accountability. The legal system around medical negligence was not designed to punish doctors for not knowing things. It was designed to punish them for not taking enough care — for being tired, rushed, inattentive, financially conflicted, or simply jaded after twenty years of twelve-minute consultations. The threat of litigation is a motivational instrument for a workforce constitutionally capable of the failure modes it deters. An LLM is not that workforce. It cannot be exhausted or conflicted or jaded. It brings the same attention to its ten-thousandth consultation as to its first. Every failure mode that medical negligence law was built to prevent is a failure mode the LLM is architecturally incapable of. The residual risks — wrong knowledge, hallucination — are knowledge and quality-control problems, managed through licensing, architectural oversight, and the dual-LLM error-checking layer already in the proposed system. Importing the accountability framework from one system into the other and claiming the fit is wrong is like demanding a calculator carry malpractice insurance in case of arithmetic errors.

The data governance question is different and should not be conflated with the litigation argument. Health information privacy as currently constructed is primarily three things: a price regulation mechanism in insurance markets, an employment protection instrument in labour markets, and a stigma maintenance system in social contexts. None of these are well served by information suppression, and all three would be better handled through transparency with access controls rather than blanket concealment. The insurance case is better resolved by opt-in disclosure for premium reduction — share your LLM health profile and pay a lower premium; decline and pay population average plus an information loading. The employment discrimination case is better resolved by anti-discrimination law than by hiding the information. The stigma case runs backwards: you do not destigmatise depression by concealing it.

The residual legitimate privacy concern is narrower than the general framework claims. It concerns people for whom disclosure of health information to a third party creates a direct physical safety risk — a person in a coercive relationship whose partner monitors their communications, or a person in a jurisdiction where certain diagnoses carry legal consequences. That population deserves targeted protection through access controls and audit trails on who can see what, not a population-wide suppression framework that costs everyone else the benefits of information transparency. The problem is the same class of problem as social media privacy, and it has the same class of solution: granular access controls, audit trails, and prosecution of misuse.

The second steelman is the deskilling risk, and it is the most technically serious argument the current system's defenders have. In the proposed system, the pharmacist verifies the LLM's recommendation before dispensing. The nurse practitioner conducts the examination on the LLM's direction. If those roles become sufficiently routine, the humans in the loop may gradually lose the independent clinical judgment needed to catch the cases where the LLM is wrong. Oversight becomes performative. This is the documented failure mode of autopilot dependency in aviation and algorithmic monitoring in finance. The mechanism for preventing it needs to be designed in, not assumed. Anyone proposing deployment at scale is already thinking about how that works.

The third argument — that a distributed GP network constitutes an early warning system for emerging disease clusters — collapses on examination. The claim is that GPs noticing unusual symptom patterns across a small local sample provides surveillance value that would be lost in a disaggregated system. But software is not merely better at this in principle — it is already doing it better in practice, and has been for years. Syndromic surveillance systems, wastewater monitoring, and real-time aggregation of structured clinical data already outperform informal clinical observation networks by every measurable metric. During COVID, wastewater detection preceded clinical network detection consistently. The LLM-first system would generate more structured, more consistent, and more geographically tagged symptom data than the current system produces — not less. Arguing that replacing GPs weakens epidemic surveillance mistakes the worst signal source in the stack for an irreplaceable one.

Two of these three arguments are genuine design requirements for the replacement system. One is not an argument at all.

The obvious deployment environment that nobody is deploying in

Remote and regional Australia has had a documented GP shortage for decades. Rural communities routinely operate without a resident GP. The Flying Doctor service exists because the alternative is nothing. People in those communities die from conditions trivially manageable in Brisbane because the triage function — the first conversation, the initial differential — is not available.

The LLM closes that gap. It requires a smartphone and a data connection. The examination layer requires a nurse or paramedic, both more available in regional areas than GPs. The pharmacist layer requires a chemist, which most regional towns have.

A controlled deployment in remote communities was the obvious Phase 1 in 2024. It would have generated outcome data, refined the escalation protocol, identified the failure modes in this specific population, and built the accountability architecture needed for broader rollout. It would have saved lives in communities currently using nothing as their primary care system.

It was not done. It is not being done.

The Australian Government's response to the primary care access problem in its 2025–26 Budget was a $7.9 billion investment to expand bulk billing — making existing GP consultations free to more people. This is not a criticism of the intent. Free GP access is better than expensive GP access. But it is a large public investment in the perpetuation of the existing bundle at the exact moment disaggregation became technically feasible. The policy moved in the wrong direction, at scale, with bipartisan support, because the people who made that decision were responding rationally to the incentive structure they operate within. Health ministers face the medical lobby, not the patient lobby. Regulators face accountability for action, not for inaction. One AI misdiagnosis becomes a Senate inquiry. Ten thousand patients who never received adequate triage in regional Queensland do not.

The asymmetry is structural. Nobody needs to be malicious for it to cost lives quietly.

Why it will not happen soon

If you knew there was a way to get instant, better medical care for all Australians at significantly less total cost, wouldn't you do it?

The answer, apparently, is no. Or more precisely: not yet, not here, and not without a fight that no health minister currently wants to have.

A reasonable forecast for meaningful deployment of disaggregated LLM-first primary care in Australia: not within two years. Probably not within five. The evidence for that forecast is not about technology readiness — the technology passed the threshold more than a year ago. It is about the political economy of health system reform, and that economy has no forcing mechanism.

The aging population does not accelerate this timeline in the way intuition suggests. An aging population increases the political salience of the existing system for the people who use it most. The demographic pressure that should create urgency instead creates incentive to protect what is familiar. In a compulsory voting system, older voters who are satisfied with their GP are not a marginal constituency — they are a reliable, concentrated bloc, and no health minister voluntarily picks a fight with them.

If the change comes within a decade, the more probable driver is not public policy. It is private health insurers identifying the cost arbitrage and funding parallel systems outside the public framework — LLM-first for those with private cover, GP-bundle for everyone else. That path produces a two-tier primary care system stratified by income before it produces a better system for everyone. The wealthy get faster, cheaper, more accurate triage on demand. Everyone else waits. The irony is that the system most likely to democratise healthcare access — available on any smartphone, at any hour, at near-zero cost — arrives first as a premium product, because the public system that should deploy it universally is too captured by incumbent interests to move.

That is the darker and more accurate forecast, and it is worth sitting with rather than glossing over.

The technology is not the constraint. It has not been the constraint for at least a year.

What this means

The GP bundle is healthcare's most expensive sacred simulacra. Everyone working inside the system knows, at some level, that the bundling is not functional — that the repeat-script GP and the complex-multimorbidity GP are not performing the same role and should not be the same unit. The knowledge exists. The fix is available. The political will is absent.

The Paragentist question is not whether disaggregation would work. The research says it would, for most of the distribution. The question is who benefits from the unit, and what that tells you about why the unit persists. Liability structures benefit from a single responsible party. Regulators benefit from a simple accountability target. Medical schools benefit from scarcity of the credential. Insurers benefit from a legible billing unit. None of those beneficiaries are the patient or the taxpayer.

The patient needs the function. Everyone else needs the bundle.

In the meantime, people with painful prosthetic hips get triaged by LLMs at midnight and discharged by attrition at five in the morning. The LLM sends them in. The institution sends them home without a diagnosis. Nobody is responsible. The system worked exactly as designed.

That is the Unit Problem.