The secretary who became fifteen executives’ assistant
A patient of mine spent his career as an executive at a large consumer-products company. When he started in the early 1980s, every executive had a secretary. Letters were typed and mailed. Travel went through an agency. Over the next twenty years, email replaced the mail, voice recognition replaced the typist, and a website replaced the travel agent. By the time he retired, one administrative assistant supported fifteen executives, and there was no travel agency.
Nobody announced a layoff. Positions became obsolete and were not refilled. The executives did not work harder; the tool made each task quicker than explaining it to someone else. The ratio of people doing the work to people managing the work moved from one-to-one to fifteen-to-one, and productivity moved with it.
That happened in nearly every industry. It did not happen in the hospital. Administrative staffing in American hospitals grew through the same decades that shrank it everywhere else, and the share of hospital spending that goes to administration is now a quarter or more by most estimates. We added layers. We did not remove any.
Real physician voices, twice a week
Free, and one click to unsubscribe.
What medicine did instead
Medicine got the same technology and used it backwards. The electronic record removed the transcriptionist and the chart room, and then handed their work to the physician. We type what a secretary once typed. We click through a dozen screens to find what a ward clerk once pulled. We finish the day’s notes at the kitchen table. The clerical work was never eliminated; it was moved onto the most expensive person in the building and renamed documentation.
The same thing happened on the business side. The record generated more data than anyone could read, so hospitals hired people to read it: coders to read every chart, billers to review every claim and remittance, analysts to answer the questions executives could not get out of the system themselves, and managers to supervise all of them. Every layer was a rational response to a tool that produced work rather than absorbing it.
That is why I say administrative bloat is a design choice. It is not a law of hospitals. It is what you get when software is built to record and the people are left to do.
What AI-native actually means
Every vendor now says its system has AI. Most mean a model has been attached to the old record: a button that drafts a note you still have to assemble, a summary of a chart you still have to search, a suggestion a coder still has to check against a chart she still has to read. The person still does the work. The model watches.
An AI-native system inverts that. The system does the work, and the person reviews it. The test is simple: After the system acts, is a human reading the whole chart again, or is a human looking at what the system did and deciding whether it is right? If the former, you have an assistant bolted onto a record. If the latter, you have changed the ratio.
The secretary disappeared because the executive stopped needing an intermediary between himself and the task. The same has to be true here. The physician should not need an intermediary between herself and the chart; the system should know the chart and answer. The biller should not need to stand between the claim and the payer; the system should defend the claim and the biller should audit the system. Supervision scales. Doing does not.
Where the ratio can move
Start with the note. Ambient listening already captures the encounter. The missing half is a system that knows the chart well enough to answer the physician’s questions in conversation, pull what the note needs from the record, and write the note in the form the physician would have written it. The physician reviews and signs. No searching, no clicking, no deciding how to work a potassium into an assessment at nine at night. The nurse gets the same deal: The documentation is written from what happened in the room, and the nurse reviews it instead of sitting at a workstation for a third of the shift.
Then the questions. A chief financial officer who wants to know why length of stay rose on one unit last month should ask and get an answer, with the supporting figures, in the time it takes to ask. Today that question becomes a request to an analyst, who becomes a position, who needs a manager. An assistant that never sleeps, never takes lunch, and is never on vacation removes the position and the manager with it.
Then the revenue cycle, which is where the arithmetic is largest and the clock is now against us. Payers have begun reviewing and denying claims with automated systems, at a volume and speed that no billing department can answer by hand. Denials that would be overturned are written off because the appeal deadline passes before a human reaches the file. The answer is not more billers. It is a system that reads the denial in the context of the chart and the payer’s own contract, assembles the justification from what the clinicians actually documented, drafts the appeal, files it, and tracks it, with the billing staff reviewing a sample and the exceptions rather than every case. The same system should do that work before the claim leaves, so fewer denials arrive at all. Machine speed on their side has to be met with machine speed on ours.
Finally the layers themselves. Each of these changes removes a kind of work, and with it the coordinators, reviewers, and managers that work required. That is where the hospital goes back to the administrative staffing it had twenty or thirty years ago: not by cutting clinical staff, who are the people doing, but by letting administrative positions become obsolete the way the secretary’s did, through attrition, redeployment into the clinical and support jobs we cannot fill, and retirement.
The conditions that make it safe
None of this works without rules, and the rules are what separate an AI-native system from an unsupervised one. A human still owns every consequential act: the signed note, the certified claim batch, the adverse credentialing decision, the write-off above a threshold. The difference is that the human exercises that ownership as review, over a sample and an exception queue, not as authorship of every item.
The system’s accuracy has to be measured continuously against what the reviewers find, and a function whose accuracy drops below its floor has to revert to full human review on its own, without anyone deciding to flip a switch. Coding suggestions must be blind to reimbursement; a system that learns to code toward payment is a compliance problem, not a productivity tool. Every automated act needs to carry its reasoning and its confidence, so the reviewer sees why it is in front of her. And every access to a patient’s record by a model has to be logged as the access of the person who asked, so the audit trail does not go dark behind the assistant.
These are not constraints on the idea. They are the idea. The secretary’s work could be automated safely because the executive still signed the letter.
What to ask of any system that claims the label
When a vendor tells you its system is AI-native, ask three questions. After it acts, does a person review the work or redo it? Can it show you, for your hospital, the minutes it removed from each role against a measured baseline, and the work it added back? And when it is wrong, what happens by itself?
If the answers are redo, no, and nothing, you are looking at the electronic record with a model attached, and the layers will stay.
I am a gastroenterologist who has spent thirty years in the hospital and the last several months designing a system built on these principles, so I have an interest in the answer. But the argument does not depend on any one product. It depends on a choice every hospital is about to make again, the same one the rest of the economy made in the 1990s: whether the technology does the work, or whether we hire people to do the work the technology creates.
Brian Hudes is a board-certified gastroenterologist and hepatologist with more than thirty years of clinical experience and a recipient of his specialty board’s thirty-year certification award. He built and ran a private gastroenterology practice for twenty years, then spent over a decade in hospital-based gastroenterology, most recently as chief of gastroenterology and medical director of GI and endoscopy at a 550-bed Level I trauma center in Pensacola, Florida. He holds a faculty appointment as assistant professor of medicine at Florida State University College of Medicine.
Those two halves of a career are the reason he writes about clinical software the way he does. The outpatient practice and the inpatient service fail their clinicians differently, and most people building for one have never worked in the other.
Dr. Hudes has been building the tools alongside the practice for just as long. In 1995, during his GI fellowship, he co-developed one of the first Windows-based endoscopy reporting systems in the United States, written because nothing on the market had been designed by anyone who had performed a procedure. He is now founder and chief executive officer of AiMOS Systems Corporation, where he is building an AI-native clinical platform for both settings, intended to replace the encounter-based electronic record rather than layer intelligence on top of it. AiMOS has been accepted into NVIDIA’s Inception program.
He writes on the architecture of clinical information, administrative cost growth, physician workforce shortages, board certification policy, and the widening gap between what clinicians need and what the industry builds. Professional updates are available on LinkedIn.
