Essay 1: Living With AI in Healthcare
August 3, 2026
A note before reading: when artificial intelligence in medicine makes the news, the stories almost always focus on the futuristic edge — robotic surgeries, AI algorithms making complex diagnoses, or software attempting to act as a digital doctor. Those stories are real, and this series will get to them. This first essay starts somewhere quieter: the administrative machinery — the back-office engines that determine how records move, how insurance claims get processed, and why navigating U.S. medicine often feels like being stuck in a different century.
The Universal Paper Clipboard
Almost everyone who has visited a medical specialist in the United States shares the same frustrating experience: sitting in a waiting room with a pen and a clipboard, manually filling out five or six pages of paper forms. You list your basic contact details, your insurance policies, and your medical history from memory.
Then you walk into the exam room to discover the specialist — sitting in front of a modern computer — has virtually none of your actual medical history.
This relies on a glaring structural flaw: the U.S. healthcare system regularly expects the patient to act as a manual “data mule.” Data isn’t missing because computer storage is expensive or difficult; it is missing because thousands of competing hospital networks, clinics, and private practices operate on isolated, proprietary electronic health record (EHR) platforms that often cannot — or aren’t configured to — share information easily with one another.
An Architecture Choice, Not a Tech Limit
It is easy to assume this friction is simply an unavoidable byproduct of modern medicine. It isn’t. It is an architectural and policy choice.
The Netherlands makes the point well. Its national exchange network, the LSP, lets a patient’s GP or pharmacist share their record with other providers — not automatically, but through a one-time consent folded into routine registration, with roughly two-thirds of the population currently opted in. A specialist there doesn’t hand over a clipboard; the system already knows who’s walked in the door. The lesson isn’t that better technology solved this — it’s that the system started from a different assumption: a record should follow the patient by default, with consent handled once, rather than every new provider starting from zero. The technology to move medical data has existed for decades. The reason the U.S. still leans on the paper clipboard is that healthcare here was built around fragmented, fee-for-service billing silos, not a centralized patient journey.
The $250 Billion Paperwork Engine
Inside this fragmented environment sits a massive, invisible industry known as Revenue Cycle Management (RCM) — the administrative machinery that translates clinical visits into standardized billing codes, processes insurance claims, and manages payment denials.
Estimates vary depending on what’s being counted, but even conservative studies put the scale of administrative bloat in U.S. healthcare at well over $250 billion annually. Medical practices do not employ vast armies of administrative staff to improve patient health; they employ them to fight their way through complex billing codes and insurance hurdles.
Take the prior authorization process — the requirement that a doctor receive advance approval from an insurance company before ordering a specific test, procedure, or medication. According to the American Medical Association’s most recent physician survey, a typical practice completes roughly forty prior authorization requests per physician, per week, consuming an average of 13 hours of physician and staff time. That’s more than a full working day spent every single week just getting permission to deliver care — and a follow-up survey a year later found the number had crept up, not down.
Enter the Software: What Actually Changed
Some of what gets marketed today as “AI” in this corner of healthcare isn’t new. Software that auto-fills a form, flags a missing field, or routes a claim to the right desk has existed for years, running on straightforward rules: if this box is empty, stop and flag it. That kind of automation deserves its own credit — a real share of the efficiency gains in medical billing over the last decade came from exactly this, with no machine learning involved at all.
AI is what actually adds something fixed rules can’t do: prediction and reading. Machine learning models can look at a claim before it’s submitted and estimate, based on thousands of past claims, whether a particular insurer is likely to reject it, and why. AI systems can read a doctor’s loosely written consult notes and extract the specific details a rigid coding system requires — something a static rule struggles with the moment the wording changes.
Where hospitals have published results, the numbers are real, but worth reading carefully — most of what’s publicly available comes from the vendors selling these tools or from healthcare finance trade groups, not independent audits. With that caveat: Texas Children’s Hospital reported a 30 percent reduction in claim denials after adopting AI-assisted pre-submission review, along with a 50 percent cut in new coder training time. Auburn Community Hospital, a 99-bed rural hospital in New York, cut its backlog of discharged-but-unbilled cases in half without adding staff. A broader academic review pooling several hospital case studies found denial rates typically falling somewhere between 10 and 30 percent, with claims processing time cut by as much as half in some deployments.
None of this has solved the underlying problem — it’s changed its shape. And health system leaders aren’t uniformly convinced yet. McKinsey’s own healthcare research notes that hospital executives are often skeptical going in, specifically because years of past automation and analytics projects didn’t deliver what they promised. That caution is worth taking seriously rather than waving away.
A small piece of this shift is also starting to show up somewhere much more personal: a growing number of waiting rooms now hand over a tablet instead of a clipboard, walking through the same intake questions one at a time. It’s a modest change, but it quietly fixes something paper never could — the illegible handwriting and skipped fields that come with filling out five pages by hand in a rush.
The Automated “Arms Race”
However, this transition is creating a strange new dynamic. As health systems deploy algorithms to auto-generate clean claims and fight denials, insurance companies are simultaneously deploying their own algorithms to audit incoming claims, identify technicalities, and issue automated denials at scale.
What used to be a slow, manual argument between a hospital billing clerk and an insurance representative is rapidly becoming an automated “arms race” between algorithms operating behind closed doors. The irony is hard to miss: software is increasingly being used to create paperwork faster, while other software is being used to reject that paperwork faster. The people caught in the middle are still physicians, office staff, and patients waiting for care.
In Plain Terms
Here, in the back office, AI isn’t replacing the doctor’s intellect or diagnosing rare diseases. Its job is acting as a software bridge across a broken, fragmented administrative system that was never designed to communicate with itself — and it’s doing that job unevenly, with real gains in some places and real skepticism in others.
By taking on the unglamorous burden of paperwork, prior authorizations, and billing friction, these tools aren’t trying to revolutionize medicine — they are simply trying to clean up the back-office noise so human doctors and patients can finally focus on care.
The next essay in this series turns to the exam room — AI’s growing role there, and what it means for the time doctors spend looking at a screen instead of a patient.
State of Play, as of August 2026
The prior authorization figures come from the AMA’s 2024 physician survey, with a follow-up survey a year later showing the burden holding steady or slightly increasing. The Dutch consent numbers reflect the most recent national reporting available. Estimates of U.S. administrative waste vary by study and methodology — $250 billion sits on the conservative end. Most of the RCM improvement figures cited here come from vendors or industry trade groups rather than independent audits; the academic review figures (10–30% denial reduction range) are the most independently aggregated data currently available, and are used here for that reason.