The prevailing narrative about AI and white-collar work is a displacement story: AI learns a cognitive task, the humans who perform that task lose their jobs, and the only question is how fast it happens. Software engineers are the canary. Lawyers and accountants are next. Eventually, the thinking goes, most knowledge work gets automated away.
There’s a problem with this narrative. It mistakes the death of a role for the death of a function — and the data increasingly shows these are very different things.
Take software. The Bureau of Labor Statistics tracks two distinct occupations. “Computer programmers” translate specifications into code — a narrow, structured task. “Software developers” do something broader: they design systems, gather requirements, make architectural decisions, and coordinate across teams. Programming is one thing they do, but it’s not the whole job. Since 2023, programmer employment has fallen 27.5% — headcount at its lowest since 1980. Software developer employment has declined just 0.3%, and the BLS projects it to grow 17.9% over the next decade.
Meanwhile, the output is exploding. GitHub code pushes are up 35% year over year. New iOS apps are up 50%. New websites, 40%. Software engineer job postings, after cratering through early 2025, are now rising faster than overall job postings. AI didn’t reduce the need for code or for people who work with it. It collapsed the narrow role and expanded the broader one. The Jevons Paradox — efficiency increasing total consumption, not reducing it — applied to knowledge work.
I’ve spent some time analyzing whether this pattern holds in healthcare, where the stakes are higher and the structural dynamics are different. I analyzed 515 medical scribe job postings from 301 employers across 44 states. The pattern is strikingly similar — and the differences reveal which workers AI will augment and which it will replace.
The Parallel
Medical scribes map onto “computer programmers” almost perfectly: a role defined by a single, highly automatable cognitive task. Scribes document patient encounters in real time — structured, repetitive, and increasingly within reach of AI ambient scribe tools from Abridge, Suki, and Microsoft’s DAX Copilot, now available for a few hundred dollars a month. Scribe job postings dropped 20% in 2025. The narrow role is contracting, just as it did in software.
But as with software, the output hasn’t contracted. It’s exploding. There are an estimated 80,000 medical scribes in the U.S. (there is no clean BLS occupational category, so any count is approximate). There are over a million physicians and more than 425,000 nurse practitioners and physician assistants. Even at peak, scribes covered a fraction of clinical encounters. Most AI scribe adoption is happening where no human scribe ever existed — a primary care doctor in rural Iowa going from nothing to a fully documented visit, not from a human scribe to an AI one.
A pre/post quality improvement study in JAMA Network Open (Olson et al., October 2025) found that across 263 clinicians at six health systems, self-reported burnout fell from 51.9% to 38.8% after 30 days with ambient AI. That’s the first-order Jevons effect: net new value where none existed before.
But there’s a second-order effect. When documentation becomes effectively free and instantaneous, health systems don’t just maintain the same output at lower cost. They demand more of it — richer notes, more aggressive coding and billing capture, structured data for population health analytics, quality metrics that were previously too expensive to extract from clinical encounters. This is why 36.5% of scribe and hybrid postings in my dataset now reference coding, billing, or revenue capture. The cost of documentation dropped, so the appetite for what documentation can do expanded. Jevons, applied to clinical notes.
The Rebundling
In software, the programmer is being absorbed into the developer — a role with a wider remit and higher expectations. The same rebundling is visible in my scribe data. In 515 job postings, 31.6% now bundle documentation duties with another clinical role, most commonly medical assistant.
The hybrid role isn’t a downgraded scribe. It’s an upgraded MA. Standalone scribe postings recruit from the pre-med pipeline: 23% mention pre-med experience and only 2.3% require clinical certification. Hybrid postings recruit credentialed clinical workers: 39% require MA or ophthalmic certification, 32% require BLS/CPR. The task profile shifts accordingly — standalone roles devote roughly 70% of responsibilities to documentation, while hybrid roles dedicate 40% to physical, in-the-room work: vitals, diagnostic tests, injections, referral coordination.
In both industries, AI automates the routine cognitive core and the remaining human role expands outward into judgment, coordination, and tasks that require a body or a relationship.
Where the Two Cases Diverge
The augmentation pattern looks structurally identical for coding and medical documentation, but the durability of augmentation differs dramatically — and the reason comes down to three variables.
Physical residual. Software development has near-zero physical residual. It’s screen work, performed anywhere, and AI can replicate much of it without a human in any room. Healthcare has a large physical residual — the part of the job that still requires a human body in a specific room. In ophthalmology, where scribes traditionally operate diagnostic equipment, 79% of postings are now hybrid. In primary care, where scribing is purely cognitive, the hybridization rate in my sample was zero. The physical residual determines whether the role has an anchor or not.
Labor shortage. The tech labor market has a deep pipeline of candidates. CS enrollment boomed for a decade; bootcamps and global talent pools keep the supply elastic. When AI boosts senior developer productivity, companies can be selective about junior hires. In a recent MGMA Stat poll, 47% of medical practice leaders cited medical assistants as the single hardest role to recruit, and 43% of practices hired alternative staff to cover MA vacancies. When you can’t hire enough MAs, absorbing scribe duties into MA roles is a survival strategy, not an optimization choice. The shortage pulls humans into expanded roles rather than pushing them out.
Regulatory stakes. Code that ships with a bug gets a patch Tuesday. A clinical note with a hallucinated medication can harm a patient. This is why 98% of scribe postings still describe documentation creation as the core function rather than QA over AI-generated notes — but the QA role is emerging. As ambient AI becomes standard, someone still needs to confirm the note is right. In software, this maps to the growing recognition that AI-generated code ships faster but can be buggier, shifting bottlenecks from writing to review. But in healthcare, the validation function carries legal and clinical weight that gives it unusual staying power.
The formula, roughly: augmentation durability = physical residual + labor shortage + regulatory stakes. Healthcare scores high on all three. Software scores low on at least two. Most white-collar knowledge work falls somewhere in between.
The Macro View
A February 2026 NBER working paper from Bloom, Barrero, Davis, and colleagues surveyed nearly 6,000 executives across the U.S., UK, Germany, and Australia. More than 90% reported no AI impact on employment over the past three years. But looking forward, firms expect AI to boost productivity by 1.4% and reduce employment by 0.7% over the next three years — implying roughly 1.75 million fewer jobs at existing firms across those four economies. Employees, meanwhile, predict AI will increase employment by 0.5%.
The scribe market and the programmer market may be early windows into what that coming transition actually looks like at the job-posting level: not mass layoffs, but quiet rebundling of who does what. Roles reorganizing around AI, with some functions absorbed, others expanded, and new ones (validation, orchestration, oversight) beginning to emerge.
What This Means
The zero-sum model says AI automates a task, a human loses a job. The evidence from both programmers and scribes suggests a more nuanced sequence: AI automates the routine cognitive portion of a role, the capability diffuses to millions who never had access, the humans who remain are expected to do more, and a new validation function begins to take shape.
This isn’t a permanent equilibrium. The physical tasks that protect hybrid scribe roles today are themselves subject to technological erosion. The junior developer pipeline that’s being squeezed now may never recover its pre-AI shape. Today’s augmentation may be tomorrow’s displacement. But in February 2026, the data across both industries shows augmentation — and it’s worth building workforce strategy around what’s actually happening rather than what we fear might happen next.
The question for any role facing AI exposure isn’t simply “can AI do this task?” It’s: what’s the physical residual? How tight is the labor market? How high are the stakes when AI gets it wrong? The answers determine whether you’re in healthcare’s position — where augmentation has structural anchors — or in programming’s position, where the ground is shifting faster than most people realize.
Software employment data from BLS Occupational Employment and Wage Statistics; chart via James Bessen, Technology and Policy Research Initiative at Boston University. Coding output data from FT analysis by John Burn-Murdoch (Domain Name Industry Brief, SensorTower, GitHub). Software engineer job posting rebound from Citadel Securities / Indeed. Medical scribe analysis based on 515 U.S. job postings from 301 employers across 44 states, collected February 2026 via JSearch API. Burnout findings from Olson et al., JAMA Network Open, October 2025 (pre/post quality improvement study, self-reported outcomes). MGMA staffing data from 2024 MGMA Stat polls. Firm expectations from Yotzov et al., “Firm Data on AI,” NBER Working Paper 34836, February 2026.
Leave a Reply