Illustration: a 3D model of a hospital that opens up floor by floor as you scroll. It shows an operating theatre on level 2, a coding team on level 1, and an MRI room and the billing office on the ground floor. A marker follows one case down through the building, labelled first as an op note, then as codes, then as a claim, until it reaches the billing office, where a staff member has been on hold with an insurer for forty minutes.
A point of view · Autonomous operations for healthcare
The administrative half of medicine has always been a machine. Until now, it ran on people.
What follows is an argument about arithmetic as much as technology: why that machine will begin to run itself, why almost no one will consciously decide that it should, and why the shape of the outcome is still ours to choose.
A woman there is on hold with an insurance company. She has been on hold for forty minutes. The question she is waiting to ask takes eleven seconds to say out loud: was this MRI approved, or not?
She does that eight hours a day, and she is good at it. That is the strange part. She is not slow, and she is not careless. The job itself is built out of waiting.
Upstairs, a doctor is saving a life. Downstairs, two hundred people are pushing paper so the hospital gets paid for it. Nobody designed this. It grew.
01
The mechanism
Here is the whole argument in one line. The back office exists because healthcare's data is fragmented; the human labor that patches it cannot scale; and once the data is connected, the work can run itself. Everything below is that sentence, slowed down and put next to the numbers.
The easy explanation is that insurers are slow and phone systems are old. That is not the reason. The reason is that the answer to her eleven-second question does not live in any one place.
To know whether the MRI is approved, four separate things have to agree: the clinical record that says why it was ordered, the payer's current policy for that procedure, the patient's specific benefit design, and the history of what has already been tried. Each of those lives in a different system. None of them was built to talk to the others.
The clinical record
Why the scan was ordered, and what the physician already knows.
The payer policy
The rules that decide approval, revised more often than anyone tracks.
The benefit design
What this specific patient's plan covers, and under what conditions.
The prior history
What has already been tried, billed, denied, or appealed.
She is not waiting on a person. She is the connective tissue between four systems that were never wired together. She is doing by hand, on the phone, what no software was ever built to do on its own.
$11
What a practice spends working one prior authorization manually, by phone and fax. The identical request handled electronically costs about half that.
14 hrs
The time a physician and their staff lose to prior authorizations in a single week, close to two full working days.
40+
Prior authorizations the average physician's practice files every week, most of them routine and rule-bound.
Rounded industry estimates, shown to convey scale rather than to forecast any one organization.
The back office is large because the data is fragmented. Every administrative job is, underneath, the same job: reconciling systems that were never designed to agree.
This is why healthcare resisted automation while banking and logistics did not. The work here was never mainly about typing or dialing. It was about holding four sources of truth in one head long enough to make a judgment. Fragmentation does not add a little friction to that. Industry estimates put the administrative overhead it creates at roughly ten times what a connected system would need, which is why close to thirty cents of every dollar the country spends on healthcare now goes to coordination rather than to care.
The fragmentation tax
Most of the administrative effort a fragmented system spends would not be needed if its data were connected.
About 10× the necessary work
needed
overhead created by fragmentation
of every unit of administrative effort a fragmented system spends
02
The deep cause
There is an older, deeper reason the back office kept growing, and it is not incompetence. In the 1960s the economist William Baumol noticed that in work where productivity cannot easily rise, cost rises anyway, because wages across the economy rise together. A string quartet needs the same four players it needed two centuries ago, so as everything else gets cheaper to produce, live music gets relatively more expensive. Healthcare has lived under this rule for decades.
Care delivery is partly like the quartet, and mostly should stay that way. Healthcare's administrative coordination is a worse case. Its volume grows with the complexity of the system, and it resisted every prior wave of productivity, because the work does not sit inside any one system where software could reach it. It sits in the seams between systems that do not share data. For fifty years, coordination was the one kind of work no technology could touch, so its share of the healthcare dollar rose, and kept rising.
The scissors
As technology drove down the cost of what it could automate, the cost of coordination it could not touch kept climbing.
Cost of healthcare coordination
Cost of what technology automates
Illustrative and indexed to a common starting point. The directions reflect the long-run pattern Baumol described; the values are simplified to make the divergence legible.
This is the part most discussions of AI in healthcare miss. The prize is not making a physician read faster, which you cannot do and should not want to. The prize is that coordination, the connective labor between every record and every decision, can for the first time approach the cost of software rather than the cost of people. That is a different lever, and a far larger one, than helping a clinician at the point of care.
Autonomy is the first productivity shock in fifty years that reaches the work between the systems, not the work inside them.
03
Anatomy of one approval
Follow a single prior authorization, the permission slip a hospital needs before it can scan you. Below are the twelve steps it actually takes, spread across four disconnected places: the EHR, the payer portal, the phone, and a tracking spreadsheet nobody owns. Step through to see which the machine takes on, which stay human, and the new oversight work that appears on the other side.
Stays human
Automated
New oversight
Twelve steps, every one of them done by a person today. Ten involve no medical judgment at all. They are checking eligibility, gathering documents, sitting on hold, and watching a portal for a status that changes without warning.
A single elective imaging authorization, decomposed. Ten of the twelve steps carry no clinical judgment; they are lookups, assembly, submission, and waiting. Two are judgment calls that were never paperwork. Automation does not shrink the job to nothing. It moves the weight from waiting to oversight.
04
The anatomy of the spend
The number is abstract until you break it apart. The roughly $1.1 trillion the country spends every year on administration is not one thing. It is five large functions, and each of them is a version of the same reconciliation problem the billing office was solving by phone: taking facts that live in separate systems and forcing them to agree.
$1.1 trillion, five functions
The administrative spend is five large functions, each a version of the same reconciliation problem.
Illustrative split of roughly $1.1 trillion in annual administrative spend. Tile area is proportional to spend. Shares are rounded and directional, meant to show composition rather than precise accounting.
Two things stand out once it is laid out this way. The largest slice, the revenue cycle, is also the most adversarial: getting paid means coding a visit, submitting a claim, absorbing a denial, and appealing it, a loop that exists mostly because two institutions disagree about what happened and what it is worth. And every one of these functions shares a profile. Each is high volume, rule-bound, repetitive, and almost never touching a clinical decision. That profile is not a coincidence. It is the exact description of the work that automates first.
None of these functions is small, and none is optional. A hospital that stops verifying eligibility does not get paid. One that stops credentialing cannot let its physicians practice. One that stops reporting quality loses its contracts. The back office is not bloat waiting to be cut. It is load-bearing, which is exactly why it has been so expensive to carry by hand, and why moving it onto a system that can carry it is worth so much.
05
A short history
Health systems have not underinvested in technology. They have spent enormous sums, through three distinct eras, and the back office grew through all three. The pattern is worth seeing clearly, because it explains why the next era is different in kind, not in degree.
| Era | What the software does | What a person still does |
|---|---|---|
| Era oneSystems of record | Digitizes the chart and captures the data. | Works around silos that were never built to talk. |
| Era twoSystems of insight | Finds the care gap and names the denial. | Closes it by hand, across the same disconnected systems. |
| Era three · nowSystems of action | Inherits the full context of a case and does the work across systems. | Governs the system and handles the exceptions. |
Each of the first two eras added a system. Every system added a seam, and every seam added coordination work for a person to do by hand. That is the quiet mechanism by which more software produced a larger back office rather than a smaller one. Digitizing the record did not connect it. Layering analytics on top produced advice, not action, and left a human to carry the advice across the same disconnected systems. The third era is the first that operates across the seams instead of adding another one.
06
The forces
For forty years, the fix for more paperwork was more people. More coders, more schedulers, more prior-authorization staff, more denial specialists. That worked because people were available and the volume was survivable. Neither is true now, and three lines have crossed that will not cross back.
10,000
People age into Medicare in the US on a typical day, while the number who want a back-office job in a hospital keeps falling.
1/3
Of a nurse's shift now goes to administrative work rather than to patients. The clinical workforce is doing clerical work it was never hired for.
30¢
Of every US healthcare dollar goes to coordination rather than to care, at margins thin enough that one bad quarter is dangerous.
Three forces, two of them crossing
The work to be done and the people to do it stopped moving in the same direction.
Capacity
the gap
work with no one left to do it
Work
Capacity index: 22
Work index: 90
Illustrative index of administrative work against available workforce capacity. Both directions reflect widely reported industry trends; the values are simplified to make the divergence legible.
Demand is rising, the workforce is shrinking, and the money is tightening, all at once. The country now spends on the order of $1.1 trillion a year on administration, somewhere between a fifth and a third of all health spending depending on how it is counted, and the coding and billing workforce behind it is aging out faster than it is being replaced. The paperwork grows faster than the medicine, because every new treatment and every new payment rule adds coordination rather than removing it. Labor scaling was never a solution to fragmentation. It was a way of paying humans to absorb it, and the bill for that decision has come due.
The only door
When a system needs more work done and has fewer hands to do it, only one thing has ever happened.
The work gets automated. Not because anyone decided it should. Because there is no other door left open.
07
The shift
The three eras explain the shape of the problem. They also explain why the last wave of tools produced recommendations and left the doing to people: each one could only ever see a fragment. The record era captured the data and locked it away. The insight era read a slice of it and gave advice. Neither could act, because acting requires seeing the whole case at once, and no layer had ever assembled the whole case.
The judgment was never automated, because the context required to make the judgment was never assembled in one place. The screen was faster than paper. The thinking was still done by hand.
The real constraint
Models are becoming a commodity. Context is the constraint.
Every health system will soon have access to capable AI. That is not the advantage. The advantage is whether the model can see the clinical record, the payer policy, the benefit design, and the history at the same time. That is not a model problem. It is a data problem, and it takes years to solve.
The current wave can take the next step, because for the first time the context underneath it is whole. Read messy human language, hold the facts of a case, and act on them. Not perfectly. Well enough for work that is repetitive and rule shaped, which is most of what the back office does. The change is not that the model got smart. It is that the ground beneath the model finally got connected.
08
Where value accrues
If autonomy is inevitable, the serious question is where the value settles, and whether this is a durable business or a passing feature. Every platform shift answers the same way: value accrues to the layer that is hardest to replicate and closest to the point of action. In this shift, that is neither the model nor the record.
Where value settles
Only one layer is both close to the work and hard to copy.
The four layers health systems and payers build on. Placement is a qualitative judgement drawn from the argument in this section, not a measurement.
Take the layers one at a time. Compute is a scale game owned by a handful of hyperscalers. Foundation models are commoditizing in real time, with capable options multiplying and prices falling, which is precisely why they confer no lasting advantage. The record systems are entrenched, but they are systems of record by design, built to hold data in rather than unify it across the enterprise and across payers, and not built to act on it. Point tools automate a single task but cannot be trusted with end-to-end action, because they can only ever see a fragment of the case.
That leaves the one layer that passes both tests: the unified context spanning clinical, financial, operational, and payer data, with the ability to act across systems on top of it. It is the hardest thing in the stack to build, because it takes years, across millions of lives, in production. It compounds, because every workflow it runs deepens the context that makes the next one better. And it sits closest to the action, where trust and value are decided. That is where the durable position is, and it is why this is infrastructure rather than a feature.
09
A framework
Autonomy is not a switch that flips. It arrives in levels, the way it did for aviation and for driving. Naming the levels matters, because most of the argument about "AI in healthcare" is really people standing on different rungs, describing different things and assuming they disagree.
Cost figures are illustrative, expressed as administrative cost per healthcare dollar, and are produced by the simple model below.
Most systems will not jump to the top. They move one rung at a time, and they notice the difference first at L2, the first level where a patient, not just a staffer, feels it.
10
The economics
Three things decide how far this goes: what machines can actually handle, how many organizations switch them on, and how much runs without a person in the loop. Move the three, and watch what falls out.
A deliberately simple model, meant for intuition rather than forecasting. It assumes capability only matters where it is adopted, and that work finished end to end is worth roughly twice work that stops for a signature.
The scale of it
Per dollar spent, reaching L2 frees several times a typical operating margin.
6×
a typical operating margin
Illustrative and per dollar spent. Operating margin varies widely by organization; the comparison is directional. The freed capital at scale is larger than most health systems' entire annual margin, which is the quiet reason this happens whether or not anyone is enthusiastic about it.
11
Three futures, not a forecast
Everything to this point argues a direction. It would be dishonest to argue a destination. The forces are real, but the future is not settled, and the responsible way to talk about it is not to predict a single outcome. It is to lay out the futures the same forces allow, say plainly what separates them, and be clear about what the picture leaves out.
Here are three, drawn from the same levers as the model above: how much machines can handle, how many organizations connect their data and switch it on, and how much runs without a person. They are not predictions, and no probability is attached to any of them. Their only purpose is to make the consequences of different assumptions comparable.
Cost and wait figures are illustrative, produced by the same model as the explorer above and the same three end values the fan above diverges to. The scenarios are conditional, not ranked by likelihood.
Notice what decides between them. It is almost never the capability of the model, which all three assume is abundant and cheap. It is the slow, unglamorous, institutional work of connecting data and earning trust, and the willingness of organizations that currently profit from friction to give it up. The binding variable is human and organizational, not technical, which is why the honest answer to "when" is "it depends on choices that have not been made yet."
A picture this simple earns its keep only if it is honest about its edges. It models the administrative back office, not clinical care, and it assumes the line between them holds. It holds regulation roughly constant, when a single rule change can create or erase whole categories of this work overnight. It assumes no safety incident serious enough to reset public trust and freeze adoption for years, which is the kind of event that has stalled other automations before. It says nothing about who moves first, whether payers automate denial faster than providers can automate appeal, or the reverse. And it takes for granted the one step that is hardest and least certain of all, that the data actually gets connected. Change any of these and the futures above rearrange.
The size of the prize is not the interesting question. Which of these futures arrives, and who keeps what it frees, are choices, not outcomes.
That last point is the one most worth sitting with. Even in the future where the work runs itself, the money it releases does not have a natural owner. It can lower the price of care, widen a health system's thin margin, pad a payer's, or accrue to whoever comes to own the connected layer underneath. Nothing about the technology decides that. People do, and the argument over it has barely started.
12
Second-order effects
The first-order effect is less manual work in the back office. The more interesting effects are the ones that follow, and they are where most of the value, and most of the disruption, actually live.
The P&L changes shape.
Administrative cost moves from headcount that scaled with volume to a platform that scales sublinearly. Health systems regain the ability to grow volume without growing cost in lockstep, which they have lacked for a generation.
The vendor landscape consolidates.
Point solutions were a business model built on fragmentation. Trust in end-to-end action requires seeing the whole case, which a single-task tool never can, so the platforms that hold the context absorb the work the point tools used to split.
The org chart moves up.
The back office's work shifts upward. The people in it handle exceptions, govern the system, and audit its judgment. The roles that grow are oversight roles that did not exist before.
The friction between payers and providers thins.
When authorization and adjudication both run in near real time, the weeks of float and the denial-and-appeal war of attrition compress. The cost both sides paid to fight each other becomes cheap for both.
Capacity becomes the story, not cost.
The freed resources do not simply fall to margin. They meet the demand a shrinking workforce cannot, so the binding constraint shifts from what a system can afford to how much care it can deliver.
13
The line that doesn't move
"Autonomous healthcare" sounds like a machine deciding whether you get surgery. That is not what any of this describes, and saying so plainly matters more than any number on this page. The human in the loop is not a safety blanket bolted on at the end. It is the design.
1 of 4
Stays human
The waiting goes away
Stays human
Diagnosis, and deciding what to do next.
The waiting goes away
Forty minutes on hold.
Stays human
The conversation. Sitting with someone and explaining it.
The waiting goes away
Re-keying the same address four times.
Stays human
The judgment call at the edge, the case that matches no rule.
The waiting goes away
Chasing a denial through a queue.
Stays human
The decision to approve or deny care. A machine never makes that one on its own.
The waiting goes away
The referral that fell through a crack because the queue was too long.
In this world a machine never issues a denial of care by itself. It assembles the evidence, drafts the case, and routes the judgment to a person with the full picture already in front of them. That line does not move. What moves is everything around it: the waiting, the re-keying, the chasing, the forty minutes that were never care in the first place.
14
The hard questions
You do not get to inevitability by feeling good about it. Four objections are worth taking seriously, and each deserves an honest answer rather than a reassuring one. The hardest of them, about work and jobs, gets more than a sentence.
The question assumes the two hundred people in the billing office are labor, hired because there is work that needs human hands. They are not. They are an integration layer. Each of them sits between institutions that never agreed on a way to exchange information a machine could read, doing by phone and by hand the reconciliation a shared interface would have done for nothing. Their salaries are what the system pays, every year, for the interoperability it declined to build. Read that way, administrative headcount is not a workforce at all. It is technical debt, denominated in people.
That reframing also explains why the back office kept getting more expensive: it is the largest untreated case of the cost disease described earlier, several million workers and roughly $1.1 trillion a year. Forty years of adding staff was not a failure of imagination. It was the only lever a function trapped that way has ever had.
Administrative labor is the last great pool of work in the economy that never found a productivity path. That is the thing now changing, and almost everything else follows from it.
Once you see it as an integration layer stuck without productivity, the interesting question stops being how many jobs get cut. It becomes where the surplus goes when this pool of work gets a productivity path for the first time in its history. That is a question about who captures the value, not about who loses a seat, and the answer is less obvious than either side of the usual argument assumes.
Start with the effect that runs against intuition. A large share of this work is not productive at all, it is adversarial. One institution adds a rule to deny, another adds a person to appeal, the first adds a rule to counter the appeal. When only one side automates, its own costs fall. When both sides automate, the loop between them accelerates. The likely equilibrium is not less administrative activity but far more of it, authorizations and denials and appeals traded between systems at almost no marginal cost, with people moved off the repetitive middle of that loop and onto the exceptions and the governing of it. Volume and staffing, which have climbed together for a century, finally come apart, because the same people are no longer needed to carry every transaction by hand.
That is what actually changes for the people. The value stops accruing to doing the work and starts accruing to governing it: deciding what the rules should be, judging the cases no rule fits, and answering to the patient on the other end. The floor under pure clerical throughput gives way, and a premium appears for the judgment and oversight the old back office never rewarded. It is a real transition and not a painless one. But it runs upward, toward the part of the job that was always the point, and in a sector this short of hands it relieves a shortage instead of creating a surplus.
When the binding constraint is people, giving them their hours back is the point, not the casualty.
15
The honest part
The technology is now the easy part. That sentence would have been false three years ago and is true today, which is exactly why the remaining obstacles are the ones worth naming. None of them is about whether the models can do the work. All of them are about the conditions around the work.
Data access is political, not technical
Autonomy needs unified context, and the data is held by parties whose incentives do not always favor unifying it. The real contest is over access, and it will be fought in contracts and standards more than in code.
Trust is earned at the institution
Someone is accountable when an autonomous step is wrong. The governance model, human judgment on the decisions that matter, full audit trails, and clean escalation of anything uncertain, has to satisfy clinicians, regulators, and courts, not just engineers.
Institutional will is the rate limiter
The rungs are not only technical. Moving from assisted to supervised is an act of organizational courage as much as a deployment, and the systems that move first will be the ones whose leaders decide to.
Regulation can pull either way
Mandates for automated prior authorization and interoperability accelerate this. Liability regimes and limits on automated determinations can slow it. The direction of the rules will shape the speed more than any single product will.
None of this is a reason it will not happen. The forces underneath are too strong for that. It is a description of where the work actually is, now that the technology has stopped being the hard part.
How you'll know it happened
There will not be a day when this arrives. It is already happening in pieces, and most people will notice it only as an absence. A bill that was right the first time. An approval back in a day instead of two weeks. Nobody asking for your address again.
The reason to call it inevitable is not that the technology is impressive. It is that the alternative has quietly stopped being available. The work keeps growing, the workforce keeps shrinking, the money keeps tightening. Something has to give.
On the other side of it is not a smaller version of today. It is a system whose limiting factor is care itself again, not the paperwork around it. The people who care spend their hours on people. The machines spend theirs on the seams. That division of labor is the whole point, and it is the one healthcare should have had all along.
It will be the paperwork.
It should have been the paperwork all along.