The White-Collar Reckoning Nobody Predicted Correctly

In 2023, the consensus view among labor economists was that AI would first eat the jobs of radiologists, paralegals, and junior software developers. Goldman Sachs published a widely-cited report suggesting that 300 million full-time jobs were “exposed” to automation. McKinsey had its own version of the apocalypse, neatly organized into quartiles. The Financial Times ran a bracket tournament of professions, ranking which would survive. Everyone was predicting, almost nobody was right.

What actually happened is stranger and more instructive than any of the scenarios those reports outlined.

The radiologists are mostly fine. Paralegals are not. Junior software developers are doing something completely different from what they were doing in 2023, but they still exist in roughly similar numbers. And the occupations that genuinely cratered — medical coders, certain categories of financial analysts, a specific subset of marketing professionals — were barely mentioned in any of the big forecasts. The mismatch between prediction and outcome is large enough that it deserves a serious accounting, not just because of intellectual honesty, but because understanding why forecasters got it wrong tells you something important about how AI adoption actually works inside organizations.

The Bottleneck Problem

The core error in most 2023-era forecasts was treating “task exposure” as equivalent to “job elimination.” A radiologist’s job involves reading scans, yes. But it also involves patient interaction, liability management, consultations with referring physicians, participation in tumor boards, and about forty other things that aren’t scan-reading. When AI became genuinely capable at reading chest X-rays, it didn’t eliminate radiologists — it eliminated the backlog. Hospitals that had six-week waits for certain interpretations suddenly had six-day waits. The radiologists were freed up to do more complex cases, second opinions, teaching, research.

This pattern repeated across knowledge work. The limiting factor in most white-collar jobs isn’t the cognitive task that happens to be most legible to AI — it’s everything else. The relationship management, the judgment calls that require organizational context, the accountability structures, the regulatory compliance, the liability surface. AI ate the legible tasks and left the rest, which often turned out to be the majority of what people were actually paid for.

This is why the forecasts were so badly miscalibrated. They were built on task-level analysis of job descriptions, not on understanding how work actually flows through organizations. Job descriptions are legal documents as much as they are accurate accounts of daily activity. They describe what a person is formally responsible for, not how they spend their hours.

What Actually Got Disrupted

Medical coding is almost gone as a profession. This one the forecasters got right for the wrong reasons. Everyone cited it as an example of routine cognitive work that would be automated. What they missed is that medical coding is also a highly political act — it determines reimbursement rates, it creates audit trails, it interfaces with insurance company systems that have their own logic. The prediction that AI would handle it was correct. The reasons turned out to be partly different from what anyone thought.

By 2026, major hospital systems were running AI coding pipelines that had accuracy rates above human coders on straightforward cases. The remaining human coders were working as exception handlers and appeals specialists — a role that requires understanding both the clinical and billing sides deeply. The total number of coders dropped by roughly 60% over three years. The ones who remained were the most skilled, commanding better salaries. This is actually the optimistic version of disruption: the profession shrank but the survivors did better.

The story for a certain category of financial analyst is less clean. The analysts whose work consisted primarily of building models from public data — the kind of people who would spend three days building a DCF model that will be used once and then discarded — found their work directly competed. Not eliminated, but competed. Where one analyst used to do one model, now one analyst was supervising the output of a dozen AI-generated models. Headcount dropped. But the nature of the surviving work changed dramatically: it became more about question formulation than execution, more about knowing what to ask the model than knowing how to build the spreadsheet.

The Occupations Nobody Was Watching

Certain marketing roles got absolutely flattened, and almost no forecast mentioned them. The content marketing writer — the person hired to produce eight blog posts a month about industry trends — is largely gone as a category. Not because AI writes better than humans (it often doesn’t), but because the economic calculus changed. If a company can produce that content for $0.40 instead of $4,000, they will, even if the quality is somewhat lower. Volume matters in SEO. Good enough beats great when the price differential is 10,000%.

The entry-level legal associate, on the other hand, is still very much employed. The prediction was that AI would replace first-year associates doing document review and legal research. What actually happened is that document review got automated, but the total amount of legal work expanded because the cost of litigation dropped enough that cases which were previously uneconomical to pursue became viable. The large firms are running leaner on associates than they were in 2023. The total number of lawyers in the country is actually up slightly, distributed differently across firm sizes and specializations.

The same expansion-of-market dynamic that kept lawyers employed also kept certain software developers employed. When the cost of building software dropped, more software got built. The developers whose work survived were those who could specify clearly what needed to be built and verify whether what got built actually worked. The developers who struggled were those whose value was primarily in implementation speed on well-understood tasks.

The Forecast Failure Mode

There’s a philosophical error lurking beneath the task-exposure methodology that I think deserves naming directly. Most of the prominent 2023 forecasts implicitly assumed that the demand for outputs would remain roughly constant while the cost of producing those outputs fell. This is almost never true. Price affects demand. When legal research gets cheaper, more legal research happens. When content production gets cheaper, more content gets produced (and the filtering problem becomes more acute, but that’s a different issue). When software development gets cheaper, more software gets built.

The forecasters who got closest to the actual outcomes were those who modeled demand elasticity, who asked “if this task gets 10x cheaper, how much more of it will be purchased?” The ones who got furthest from reality were those who treated the total amount of work as a fixed pie that AI would simply take a larger slice of.

This is not a minor methodological quibble. It’s the central issue. And the fact that it was overlooked by otherwise sophisticated research teams suggests something about how hard it is to reason clearly about general-purpose technologies. We tend to think of AI tools as substitutes for human labor. They often turn out to be complements that change the shape of the market rather than simply contracting it.

What This Means for Policy

The mismatch between prediction and outcome has real consequences for the policy interventions that were designed based on those predictions. Retraining programs focused heavily on radiologists, who turned out to mostly not need retraining. They underinvested in support for medical coders and certain categories of marketing professionals, who did.

The workers who fell through the cracks were often in occupations with lower median wages and less political visibility. This is not a coincidence. The high-status professions generated more research attention and more policy concern. The occupations that actually contracted significantly were concentrated in the administrative and lower-to-middle white-collar tier, which is also the tier that had the least access to the kind of continuous education that would have allowed adaptation.

The six-year retrospective on AI labor disruption is not a story about technology moving faster than policy could respond. It’s a story about misidentifying which workers were at risk, designing interventions for the wrong group, and underestimating the extent to which demand expansion would offset displacement in many sectors. The framework for understanding what happened needs to be rebuilt before the next wave of disruption arrives — and it will arrive before the lessons from this one have been fully absorbed.