AI’s Most Practical Role in Health Care May Be Cutting the Bureaucracy

America does not lack medical technology. It has accumulated an enormous administrative system around medicine. Artificial intelligence may be most valuable not as a substitute for doctors, but as a way to shrink the machinery surrounding them.

America spent $5.3 trillion on health care in 2024—about $15,474 for every person in the country. That was 18 percent of the entire U.S. economy. Centers for Medicare & Medicaid Services

Those numbers usually trigger familiar arguments about prescription drugs, hospital prices, insurance premiums and physician fees. All matter.

But there is another cost hiding in plain sight: the enormous administrative apparatus required to make American health care function.

Consider the accompanying graph.

A 2025 review in The Lancet compared the growth of physicians with the growth of people classified as health-care managers from 1970 through 2024. Physician numbers increased substantially. The managerial category increased vastly faster—by roughly 4,800 percent from its 1970 level according to the figure. PubMed

The chart deserves an important qualification. “Health-care managers” are not synonymous with useless bureaucrats, and the figure is based on an unpublished analysis by researchers David Himmelstein and Steffie Woolhandler using federal workforce data. Hospitals need managers, accountants, cybersecurity specialists, compliance officers, schedulers and quality-control personnel. Modern medicine could not operate without them.

The graph therefore does not prove that thousands of administrative jobs are unnecessary.

It asks a more important question:

Why has the machinery surrounding the delivery of medical care grown so much faster than the people delivering the care?

And now that artificial intelligence can read documents, understand ordinary language, retrieve information, reconcile data, complete forms and communicate between computer systems, do we still need humans performing so many of those transactions?

Increasingly, the answer may be no.

The first job for AI in health care should not be deciding what treatment a patient receives. It should be eliminating as much as possible of the paperwork surrounding that treatment.

The trillion-dollar paperwork problem

There is no universally accepted figure for the cost of health-care administration because researchers define “administration” differently.

That disagreement should not obscure the scale of the problem.

A 2020 study in Annals of Internal Medicine estimated that U.S. insurers and health-care providers spent $812 billion on administration in 2017, or $2,497 per person. Administration represented 34.2 percent of the health expenditures included in the researchers’ analysis. The comparable figure in Canada was 17 percent. PubMed

A different analysis published in Health Affairs Scholar estimated U.S. health-care administrative spending at approximately $1 trillion annually, including roughly $200 billion associated just with financial transactions. PMC

And a widely cited 2019 JAMA review classified $265.6 billion per year as waste attributable specifically to “administrative complexity”—the largest single category of waste the researchers identified. JAMA Network

The estimates differ because the definitions differ.

The conclusion does not.

America is spending hundreds of billions of dollars every year administering health care rather than delivering it.

That makes administration one of the most logical places to deploy artificial intelligence.

Why AI is unusually well suited to this problem

Computers have processed health-care transactions for decades. If conventional software could have eliminated the administrative problem, it probably would have done so already.

The problem is that much of health-care bureaucracy does not fit neatly into computer databases.

Information appears in physician notes, insurance policies, laboratory reports, PDFs, referral letters, scanned documents, billing codes, emails and attachments. Different insurers impose different rules. One organization may describe the same procedure differently from another. Someone frequently has to read something, interpret it, find information somewhere else and decide what to do next.

That “someone” has historically been a person.

Generative AI changes the economics because modern systems can increasingly work with unstructured information—the messy words and documents that older computer systems could not easily understand.

An AI system can potentially read a physician’s note, identify the requested procedure, determine which insurer covers the patient, retrieve that insurer’s authorization criteria, locate the relevant evidence in the medical record, populate a form, attach the required documents and flag the case for a human only when something does not fit the rules.

That is qualitatively different from simply replacing a paper form with an electronic form.

The objective is not merely to digitize bureaucracy.

It is to remove human beings from routine transactions altogether.

Prior authorization is almost designed for AI

Few processes illustrate the opportunity better than prior authorization.

Before a patient can receive certain medications, procedures or diagnostic tests, the physician often must obtain permission from the patient’s insurer.

That may involve determining whether authorization is required, looking up the insurer’s criteria, extracting information from the medical record, completing forms, submitting documentation, tracking the request, answering follow-up questions and appealing a denial.

According to the American Medical Association’s 2025 physician survey, practices complete an average of 40 prior authorizations per physician each week. Physicians and their staffs reported spending about 13 hours a week on them, and 40 percent of physicians said they had employees who worked exclusively on prior authorization. The survey involved 1,000 practicing U.S. physicians and, as a physician survey, reflects their reported experience rather than an independent audit of insurer records. American Medical Association

Now consider the transaction itself.

A 2023 analysis estimated that more than 90 percent of private-payer prior authorizations are ultimately approved, yet fewer than 25 percent are automatically determined. PMC

That is an extraordinary mismatch.

If more than nine out of ten requests are ultimately approved, why should so many of them require repeated human handling?

The logical future is an exception-based system.

Straightforward requests meeting clearly defined criteria should move automatically from the physician’s electronic record to the insurer and back again, potentially in seconds.

Humans should concentrate on the exceptions.

CMS has already created some of the digital plumbing necessary to make this possible. Under its Interoperability and Prior Authorization Final Rule, affected payers generally must implement standardized application programming interfaces—APIs—for electronic prior authorization beginning January 1, 2027. Centers for Medicare & Medicaid Services

An API is essentially a standardized way for computer systems to communicate directly with one another.

Combine those connections with AI capable of reading medical records and coverage rules, and much of today’s administrative relay race becomes technically unnecessary.

Nine billion claims create another enormous target

Prior authorization is only part of the opportunity.

The United States processes roughly 9 billion medical claims annually.

Researchers estimate that the combined administrative cost to providers and private insurers averages roughly $12 to $19 for each claim. Simple claims cost less; complicated claims can cost $35 to $40 to process. Labor accounts for more than half of payer processing costs and as much as 90 percent of provider costs. PMC

Multiply small inefficiencies by billions of transactions and they stop being small.

Leading insurers demonstrate what is already possible. Best-in-class operations can adjudicate roughly 95 percent of claims without manual intervention. PMC

AI can push the model further.

It can examine clinical documentation before a claim is submitted, identify missing information, suggest billing codes, compare the claim against payer requirements and recognize problems likely to result in rejection.

If a claim is denied, AI can read the denial, retrieve the relevant records, determine whether the denial is appealable and prepare the supporting documentation.

The desired workflow looks something like this:

This is not replacing judgment.

It is separating transactions that require judgment from transactions that do not.

We already know automation saves money

This is not entirely theoretical.

The 2025 CAQH Index, released in 2026, examined administrative transactions across more than 600 provider organizations and health plans representing about 63 percent of insured lives.

CAQH estimated that electronic transactions and improved data exchange avoided $258 billion in administrative costs in 2024. It identified another $21 billion in potential annual savings from further automation of transactions that remain manual or only partially electronic. More than half of health plans and more than one-quarter of provider organizations surveyed were already using AI in administrative workflows. GlobeNewswire

That $258 billion should not be described as savings produced specifically by artificial intelligence. Much of it comes from older forms of electronic automation.

But that is precisely why the number is important.

It demonstrates the underlying economic principle: removing human labor from repetitive administrative transactions can produce very large savings.

AI expands the universe of transactions that can be automated.

The next step is moving beyond structured electronic transactions into the far messier world of medical notes, insurer requirements, correspondence, attachments and exceptions—the information that previously required someone to read and interpret it.

AI is already giving doctors some of their time back

One of the clearest early examples is the AI medical scribe.

Rather than forcing a physician to spend part of an appointment typing into a computer, an ambient AI system can—with appropriate consent and privacy protections—listen to the conversation and prepare a draft clinical note.

A 2025 JAMA Network Open study involving 46 clinicians across 17 specialties found that ambient AI use was associated with 20.4 percent less time spent on notes per appointment and 30 percent less after-hours documentation time. JAMA Network

Another multicenter study of 263 clinicians found that after 30 days using an ambient AI scribe, the proportion meeting the study’s burnout threshold fell from 51.9 percent to 38.8 percent, alongside improvements in documentation burden and attention available for patients. JAMA Network

A much larger 2026 JAMA study across five academic medical centers subsequently found more modest but measurable effects: AI-scribe adoption was associated with reductions of 13.4 minutes in total EHR time and 16 minutes in documentation time per eight scheduled patient-care hours. JAMA Network

The point is not that AI scribes alone will meaningfully bend the national health-care cost curve.

The point is that they demonstrate the mechanism.

A scarce and expensive human being is spending time performing a task a machine can increasingly perform.

Give that time back.

Then repeat the exercise thousands of times across the health-care system.

How large could the savings become?

Here the estimates become more speculative, but they are large enough to warrant attention.

A 2026 analysis published in NEJM Catalyst Innovations in Care Delivery examined the potential value of machine learning, natural-language processing and generative AI across U.S. health care.

The researchers estimated that full national adoption across the administrative use cases they studied could reduce annual administrative expenses by approximately $120.1 billion to $252 billion—about 9.4 to 19.8 percent of administrative expense in their model.

Those figures should not be mistaken for a forecast.

The model assumes broad adoption, coordinated organizational change and successful implementation across the country. It also excludes one-time implementation costs. Hospitals do not simply install a piece of AI software on Monday and discover billions of dollars in the bank on Tuesday.

But even the bottom of that range is consequential.

And importantly, the largest sources of value identified by the researchers included labor productivity and administrative automation.

That is precisely where AI’s economic case is strongest.

There is one uncomfortable requirement: the workforce eventually has to shrink

This is where many predictions about technology become intellectually dishonest.

Suppose a hospital spends $10 million installing AI systems that allow its administrative employees to accomplish twice as much work—but it keeps every employee, every department and every existing process.

The hospital has not reduced its costs.

It has increased them by $10 million.

Productivity becomes cost reduction only when the organization actually requires fewer labor hours to produce the same amount of work.

That does not necessarily require mass layoffs.

Health-care organizations can capture some of the savings through attrition, leaving vacant positions unfilled, consolidating departments and redeploying employees from paperwork toward work that genuinely benefits patients.

But arithmetic eventually asserts itself.

If AI is to make health care less expensive, administrative labor per patient must decline.

Otherwise, AI simply becomes one more expensive layer added to an already expensive system.

The danger: an AI bureaucracy fighting another AI bureaucracy

There is also a plausible scenario in which AI makes the problem worse.

Imagine an insurance company using AI to scrutinize every authorization request and generate denials nearly instantaneously.

Providers respond by deploying their own AI systems to generate appeals.

The insurer’s AI reviews the provider’s AI-generated appeal.

The provider’s AI responds again.

We could automate the bureaucracy without eliminating it—creating what researchers have described as an AI “arms race” in utilization review. Health Affairs

That would be technological progress with little economic progress.

A separate 2026 NEJM Catalyst analysis raises an even broader warning. Under America’s fee-for-service system, researchers Bob Kocher, Brian Zhao and Erin Duffy argue that AI could actually increase health-care spending in the short and medium term by making it easier to deliver additional billable services. AI improves productivity, but if every productivity gain creates more reimbursable activity, overall spending can rise rather than fall. PubMed

That criticism is important because it identifies the difference between using AI and using AI specifically to lower costs.

The technology does not determine the outcome.

The incentives do.

A better model: automate the routine, escalate the exception

The appropriate architecture is relatively simple.

Routine administrative transactions should be handled automatically.

Ambiguous transactions should be referred to people.

Clinical decisions that could deny patients access to necessary treatment should retain meaningful human oversight.

Every significant automated decision should be auditable.

And organizations should be measured not by how much AI they purchase, but by whether administrative cost per patient, claim or encounter actually declines.

In practical terms:

100 routine transactions should no longer require 100 human touches.

Perhaps 85 or 95 should move through automatically.

People should spend their time on the remaining five or fifteen—the cases where experience, judgment, empathy or accountability actually matter.

That is how other information-intensive industries use automation.

There is little reason health care should be exempt.

Don’t replace the doctor. Replace the paperwork around the doctor.

The public discussion about artificial intelligence in medicine understandably gravitates toward dramatic possibilities.

Can AI diagnose cancer?

Can it interpret an X-ray?

Will it someday replace a physician?

Those are consequential questions, but they may distract us from a far more immediate opportunity.

America does not have an excess supply of physicians and nurses. It has shortages in many specialties and communities. Using artificial intelligence primarily to replace scarce clinicians would solve the wrong problem.

Meanwhile, an enormous administrative infrastructure has accumulated around those clinicians.

People are paid to transfer information between systems. To enter data already entered elsewhere. To determine which form an insurer requires. To check claim status. To obtain signatures. To assemble documentation. To review routine records. To call another organization and ask whether something has been approved.

For most of modern medical history, there was no alternative.

Someone had to do it.

There increasingly is an alternative now.

Artificial intelligence can read the form.

It can read the medical record.

It can read the insurance policy.

It can reconcile the three.

It can prepare the transaction.

And when something genuinely unusual occurs, it can hand the problem to a person.

That does not eliminate the need for administrators any more than ATMs eliminated banks or accounting software eliminated accountants.

It changes what we should be willing to pay humans to do.

The most valuable application of artificial intelligence in American medicine may therefore turn out to be far less glamorous than an AI physician.

It may be an AI clerk.

And in a $5.3 trillion health-care system burdened by extraordinary administrative complexity, eliminating millions of hours of clerical work may ultimately do more to make medical care affordable than teaching a computer to wear a white coat.