AI Labor Tracker

AI Codes 70% to 90% of Radiology Cases at Some Hospitals Without a Human Touching Them. Those Easy Cases Were How New Medical Coders Learned the Job.

A RAND study of seven hospitals that adopted AI coding found it handles the simple, high-volume cases and leaves people the hard ones. One emergency department coding team went from 32 positions to 11. The bigger problem may come later: the cases AI took were the training ground for the next generation of coders, the ones hospitals say they'll need as auditors.

At a hospital in the South Atlantic states, the assistant vice president of health information management told a RAND researcher what two years of AI coding in the emergency department had done to the team’s size.

“Although we went from 32 FTEs down to 11 that are required now, we didn’t necessarily eliminate all of those positions,” the leader said. “We did close all of our open positions, and we did do a strategic reduction in force, for some low performing staff. We immediately eliminated all of our contractor support.”

That account is one of dozens in “Three Essays on Artificial Intelligence and Work in the U.S.,” a dissertation by Jonah Kushner that RAND published on Monday. Kushner interviewed hospital leaders, medical coders and people from AI vendors, 27 in all, at seven hospitals that had brought in AI coding or were bringing it in. RAND kept every participant anonymous, so I’m describing them the way the study does.

Hospital 1 was the only one in the sample that laid people off. The others shrank the way most employers do now, by not filling seats.

And the study’s most important finding isn’t about either kind of cut. It’s about what happens to the people who were supposed to come next.

What a medical coder does

If you’ve never seen the job, here’s the short version. A doctor documents a visit.

A coder turns that documentation into standardized diagnosis and procedure codes, the ones insurers use to decide what to pay. One of the coders in the study put the standard simply: “The insurance person should be able to look at all the codes and charges that I’ve put on the patient and get a full picture without actually seeing the patient and reading any of the documentation.”

It’s a job with a ladder. New coders start on the simplest cases, like standalone X-rays and other imaging.

With experience they move up to emergency visits and outpatient surgeries, then to inpatient stays, where one patient can carry dozens of diagnoses. Above that are auditors, who check other coders’ work, and trainers.

It’s also one of the better jobs you can get without a four-year degree. Hospitals usually want a certification from AAPC or AHIMA, the two main industry groups, and the study notes they don’t typically require college.

Nationally, medical records specialists, the federal category that includes coders, earned a median of about $51,000 in May 2025, slightly above the median for all workers, according to the Bureau of Labor Statistics. In hospitals the median was about $59,000.

One coder in the study explained the appeal: “A lot of people have come to this career because it doesn’t require a four-year degree. It’s fairly well-paid. You work from home.”

So what happens if that’s your job, and software starts doing the bottom of it?

The AI takes the easy cases

The earlier generation of coding software suggested codes and let a person accept or reject each one. The new AI coding modules work on their own.

When one codes a case successfully, no coder ever sees it. “The goal is for it to go straight to bill, which means nobody else touches it,” a revenue cycle leader said.

The modules are trained on years of a hospital’s own history, matching past documentation to what was actually billed. That only works where there’s a mountain of similar cases, which is why vendors started with radiology.

“A patient can be in the hospital for a week and have four to five chest x-rays a day,” one coding director said. “They do those well.”

Across the hospitals in the sample, leaders reported the modules were coding 70% to 90% of radiology cases and 35% to 85% of emergency department cases. [Chart: share of cases AI coding modules handled at sample hospitals, by service.] The big inpatient cases stayed human.

One leader described cases with 85 diagnosis codes that have to be sequenced correctly before they can go on a claim.

The software also gets things wrong in ways a person wouldn’t. A coder described radiology reports that say something like “question mark, finger fracture,” meaning the doctor is asking whether there’s a fracture.

“At least right now, AI thinks, oh, there’s a fracture,” the coder said. “If something’s not definite, we can’t code it.”

Most hospitals cut by not hiring

Most leaders were emphatic that they hadn’t laid anyone off. What they described instead was quieter.

“I didn’t have to lay anybody off, you know? I just didn’t fill my positions that were there,” a leader at Hospital 6 said. “And then I had two people retire, I didn’t fill those positions.”

It’s a pattern that keeps turning up in other data, too. The cut shows up as a job that never gets posted, not as a layoff notice.

Leaders also said they’d adopted AI because they couldn’t hire enough coders, and one vendor representative told Kushner, “The average age of a coder in the United States, I think, is 56 or 57 years old.”

Here’s where the study earns its keep. Kushner checked that claim against federal survey data and it didn’t hold up. In 2023, medical records specialists in hospitals averaged about 48 years old, against about 42 for the whole workforce, and the occupation has grown steadily for two decades.

“Overall, these data suggest that adoption of AI coding was a deliberate choice by hospitals rather than an unavoidable response to workforce trends,” he wrote. That’s a researcher declining to repeat a convenient story from the people he interviewed, and it’s the kind of check I wish more AI coverage did.

The government still expects the field to grow. BLS projects about 8% more jobs for medical records specialists between 2025 and 2035, much faster than average, while noting that AI tools “may affect the demand for these workers.”

The people who stayed got the hard work

Here’s the part that rarely makes it into vendor case studies. When the easy cases go to the machine, the human queue gets harder.

“There’s less things that are easy to code, quick and easy,” a coder at Hospital 3 said. “Oh, that’s a hand x-ray with a wrist fracture. I’m done. That would be taken care of by [the AI coding module].”

Whether that felt like a promotion or a grind depended on the hospital. At Hospital 1, coders said the quotas never changed.

“They never changed our productivity. No one makes productivity anymore,” one said. “We’re coding difficult charts now. They didn’t even acknowledge that the charts are more difficult.”

At Hospital 4, a leader said coders reached burnout faster “because of the complex cases and there’s no opportunity to breathe and do an easy one.”

Some coders also became the AI’s unpaid quality department. One described reviewing everything the module produced for at least six months and feeding corrections back to the vendor, “but even now, there’s still stuff that I have to review that they haven’t gotten right.”

At Hospital 3, though, a coder who had asked to be involved before the hospital committed said the easy X-rays were “kind of boring” after years on the job, and that working with the vendor’s engineers was fun. Kushner’s conclusion is that management choices, like adjusting quotas and giving coders a voice, decided which version people got.

The hospital leaders and vendors agreed on where the job is going. “AI is the production coder, and the human role is now going to be auditing,” one vendor representative said. “So, I think there’ll be less coders, but more auditors going forward.”

The bottom rung

Now the problem. Auditors have to be grown, and they grow out of years of coding simple cases under a trainer’s eye.

“Once you’re, you know, hired into a position, you work with a trainer for weeks or months or whatever it takes. And they teach you the bulk of it,” a coder said. Those weeks are spent on exactly the high-volume, low-dollar cases the AI now takes.

The pipeline was already weak before the software arrived. “I can find me a green coder every day of the week,” a leader at an academic hospital said, “but given the complexities… we really don’t like to hire people without experience.” A coder said breaking in was “probably one of the hardest fields I’ve ever broken into.”

And the leaders know what AI does to it. A coding manager at Hospital 2 put it in one line: “What are we going to do with the baby coders?” A leader at Hospital 7 was more specific.

“In the past, you know, you could bring new coders in, so you could start them on the easier cases and then work to the harder cases,” the leader said. “It will eliminate most of the easier cases. So, I won’t have that luxury of training coders as much as I used to.”

Even a vendor saw it. A coding and auditing manager at one AI company reflected that the hands-on experience that got that manager into management “isn’t necessarily going to be there going forward.”

Kushner calls this an irony of automation. The software raises the expertise the job requires while “removing the bottom rung of the ladder that produces it.”

Why the training disappears with the work

There’s a first-principles reason this keeps happening, and it isn’t unique to hospitals.

Most on-the-job training is a byproduct of production. Nobody budgets for it as training.

The new person codes the X-rays because the X-rays need coding, and the trainer checks the work because it’s going to an insurer. The learning rides along for free.

Automate the production and you automate away the free ride. The hospital doesn’t decide to stop training. It just finds that the work it used to train people on is gone, and replacing it costs real money and senior time.

That’s the same shape as the worry about junior lawyers and junior programmers, and it’s why I wrote last week that verification is becoming the scarce job. The difference in coding is that someone wrote down the mechanism in the words of the people living it.

Back at Hospital 1

The 32-to-11 hospital is also the one with the most aggressive retraining. The same leader who described the cut said the hospital promoted seven people “to a supervisor role, to an auditor role, to an edit coder role,” and moved the remaining four into coder three positions, with a mix of in-house training and outside courses.

So the people who stayed went up. The people who would have been hired next didn’t get a door.

Hospital 2 tried to build one on purpose. It created a formal coding career ladder with a “coder one” entry position, mentoring from auditors and a learning platform. “It allowed us to create a coder one position. We want to grow the team from within,” a supervisor there said.

But the hospital’s own staff told Kushner the ladder ate a lot of senior time, couldn’t speed anyone past a rung, and might be out of reach for smaller hospitals. “When we bring AI in, we’re going to have to revise the career ladder again,” the coding manager said.

My read

In my opinion, hospitals are making a reasonable short-term trade and an unpriced long-term one. The software clears backlogs, codes all weekend and doesn’t need overtime, and the study has coders who honestly preferred life without the boring cases. That’s real.

But the trade borrows from the future. Every hospital that stops hiring entry-level coders saves money this year and assumes someone else will train the auditors it’ll need later. If every hospital makes the same assumption, nobody trains them.

The honest limits matter here. This is seven hospitals, recruited from AI vendors’ customer lists, and 27 interviews.

It shows how the change works inside hospitals. It doesn’t tell us how many coding jobs AI has removed, and the BLS projections still point up.

What I’d do

The study ends at Hospital 1 and Hospital 2 for a reason: one cut deep and promoted from within, the other is trying to build the bottom rung back by hand. Neither has a finished answer, and the study says so.

My recommendation, for anyone running a team where software is taking the simple work: keep some of the simple work for people on purpose. Route a set share of the easy cases to trainees and count it as training cost. It’s cheaper than discovering in five years that nobody in the building knows how to check the machine.

Chart note: Share of cases coded autonomously by AI coding modules at the hospitals in RAND’s sample, by service, low and high end of the range leaders reported. Universe: seven U.S. hospitals that adopted AI coding. Values are ranges as reported in interviews, not audited rates. Source: Jonah Kushner, “Three Essays on Artificial Intelligence and Work in the U.S.,” RAND dissertation RGSDA5507-1, Nov 9, 2026, Chapter 2. CSV: medical-coding-ai-2026.csv (also holds the Hospital 1 FTE figures and the BLS employment and projection figures with their own units and sources).

Sources

1. Jonah Kushner, “Three Essays on Artificial Intelligence and Work in the U.S.,” RAND Corporation, RGSDA5507-1, published Nov 9, 2026. Summary: https://www.rand.org/pubs/rgs_dissertations/RGSDA5507-1.html ; PDF: https://www.rand.org/content/dam/rand/pubs/rgs_dissertations/RGSDA5500/RGSDA5507-1/RAND_RGSDA5507-1.pdf (Chapter 2: sample of 13 leaders, 10 coders, four vendor representatives at seven hospitals; Tables 1 and 2 for titles and Census divisions; all participant quotes verbatim from the interview excerpts printed in the dissertation; 70 to 90 percent radiology and 35 to 85 percent ED automation ranges; Hospital 1 “32 FTEs down to 11” and seven promotions; average age 48.3 vs 41.6 in 2023 from ACS; “deliberate choice” quote; “bottom rung” quote; Hospital 2 career ladder).
2. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, Medical Records Specialists (last modified Aug 27, 2026). https://www.bls.gov/ooh/healthcare/medical-records-and-health-information-technicians.htm (median $51,140 May 2025 vs $50,980 all occupations; hospitals median $59,120; 200,700 jobs in 2025; projected 216,300 in 2035, +8%; AI “may affect the demand” line).
3. Post Labor Economics, “Verification Is the Scarce Job.” https://postlaboreconomics.com/posts/verification-is-the-scarce-job/

Quote notes: participant quotes are printed in the dissertation from interview transcripts and notes; we did not have audio. Ellipses mark our cuts. Bracketed words are RAND’s.

Derived values (our arithmetic): 32 to 11 is a 66% reduction in required ED coding FTEs at Hospital 1 (21 positions), achieved through closed openings, contractor removal and a small layoff, per the leader; we do not know the split. 216,300 / 200,700 = +7.8% (BLS rounds to 8%).