How AI Is Hollowing Out the White-Collar Career Ladder

The work has not disappeared. The ladder into it is starting to.

That is where the first durable pressure from AI is showing up in South Korea. Not at the top of the profession. Not, at least yet, in the complete removal of lawyers, accountants, programmers or analysts. The damage is lower down. It sits in the research memo, the first draft, the routine review, the low-risk reconciliation, the work once passed to juniors because it needed to be done and because that was how people learned. From July 2022 to July 2025, jobs held by Koreans aged 15 to 29 fell by 211,000. Almost all of that decline came in industries with high exposure to AI. Over the same period, employment among workers in their 50s rose by 209,000, much of it in those same sectors. The Bank of Korea has described the pattern plainly: in the early phase of AI diffusion, junior employment fell while senior employment rose, a form of “seniority-biased technological change.” 

That matters for a reason more serious than ordinary job churn. In structured professions, junior work was never only junior work. It was how institutions reproduced themselves. Law firms did not give young associates research, review and drafting simply because the tasks were billable. They did it because judgment had to be built somewhere. Accounting firms did not load early-career staff with documentation, testing and repetitive review merely out of tradition. Those were the lower rungs of a training model. Remove too much of that layer and the issue changes shape. The problem is no longer only who gets hired this year. It is whether the people a firm will need five or ten years from now are still being made at all.

Law and accounting bring the shift into focus because both fields depended on dense layers of junior labor. That arrangement is now under pressure. Thomson Reuters’ 2025 review of the U.S. legal market describes firms operating under changing client expectations, sharper competitive demands and advances in AI-driven technology. In accounting, CPA.com’s 2025 industry report argues that AI has moved from experiment to operating infrastructure. Neither development means the professions are about to vanish. It does suggest that the economics of beginner work are changing faster than the institutions built on it. A firm can narrow junior intake, buy software, lean more heavily on laterals and still look stronger in the short run. The weakness shows up later, when there are too few people in the middle who were once expected to grow into the next layer of responsibility. 

Why the first cut lands on juniors

The asymmetry is not accidental. Firms do not begin by removing the most expensive judgment in the room. They begin where work is easiest to standardize, easiest to supervise through software, and hardest to defend to a cost-conscious client. In white-collar professions, that usually means the bottom of the ladder. The first-pass review. The background memo. The reconciliation sheet. The document summary. The work that once justified a junior salary because it was necessary, because it was billable, and because it trained the person doing it.

That is why the early labor-market effect of AI does not look like clean professional extinction. It looks narrower. A firm can keep its partners, managers and senior specialists in place while quietly reducing the volume of work handed to beginners. The headline may still read “hiring slowdown” rather than “layoffs.” The organizational consequence, however, is larger than the language suggests. When entry-level work contracts, what shrinks is not only the number of newcomers on payroll. What shrinks is the space in which competence is formed. A profession may preserve its upper tiers for years while weakening the pipeline beneath them.

The Korean data fit that logic with unusual clarity. The Bank of Korea found that from July 2022 to July 2025, youth employment fell sharply in AI-exposed industries while employment among workers in their 50s rose in those same sectors. Its conclusion was not that AI had already erased entire occupations. It was that the early stage of diffusion was producing a seniority-biased pattern: younger workers were more exposed to displacement pressure, while older workers remained more complementary to the technology. OECD analysis points in the same direction. In Korea, the adverse employment effects of AI appear more likely to fall on younger, lower- or middle-skilled workers, while the gains from AI-related productivity are more likely to accrue to high-income, high-skilled groups.

There is a practical reason for that divide. Junior labor is cheaper than senior labor, but it is also costlier to carry than it looks. It requires onboarding, correction, supervision and time. In periods of uncertainty, firms can decide that the hidden cost of training is no longer justified by the immediate value of the work itself. AI sharpens that temptation. If software can compress the first round of research or documentation, management no longer sees only a junior employee. It sees an avoidable training expense. What disappears, then, is not merely a task. It is the economic rationale that once made firms willing to absorb inefficiency in exchange for future expertise.

Recent research suggests that this is exactly how the adjustment begins. A 2025 working paper using large-scale employment data found that firms adopting generative AI reduced junior hiring more than senior hiring, and that the change showed up less through dismissals than through weaker inflows. Stanford’s Digital Economy Lab reached a similar conclusion from a different angle: in AI-exposed occupations, employment among workers aged 22 to 25 declined after the release of ChatGPT, even as older workers in those same occupations saw gains. This is what makes the present moment easy to underestimate. Nothing dramatic has to happen all at once. The ladder can be weakened one hiring cycle at a time.

That shift also helps explain why the problem cannot be reduced to a simple clash between “young” and “old.” What protects senior workers is not age by itself. It is position. The further one moves up a profession, the less the job consists of standardized output and the more it depends on tacit knowledge, judgment under uncertainty, client management, internal coordination and institutional memory. Those are precisely the parts of white-collar work that remain difficult to automate cleanly. A machine can produce a draft. It cannot easily inherit responsibility for why that draft was accepted, how it was framed for a client, or what risks were judged tolerable in the first place. That gap is where senior labor continues to hold its ground.

The danger lies in assuming that such an arrangement can sustain itself indefinitely. It cannot. Senior judgment does not reproduce itself. It comes from years spent inside exactly the kinds of workflows now being compressed. That is why the first cut matters so much. It does not simply reduce the number of junior workers in the present. It begins to thin the cohort from which the next layer of experienced professionals would have emerged. By the time the shortage is visible in the middle of the organization, the decision that produced it may already be years old.

The pattern becomes clearer in professions built around structured junior work. Law and accounting are not the only cases. They are simply the ones in which the internal logic is hardest to miss.

Law and accounting: where the ladder is easiest to see

The pressure is easiest to see in professions that were built on layered work. Law and accounting are not unique because they face AI. They are revealing because they long depended on dense bands of junior labor that served two purposes at once. The work generated revenue, and the work trained the people doing it. That dual function is what now looks unstable.

In law, the lower end of the hierarchy was never incidental. Junior lawyers were handed research, review, drafting, citation checks, factual synthesis and first-pass analysis because those tasks needed to be done, but also because that was how legal judgment was formed. A profession that presents itself as argument and strategy is built, in practice, on repetition. Younger lawyers learned by tracing how a matter moved from fact pattern to memo, from memo to draft, from draft to advice. What AI alters is not only the speed of that process. It changes the economics of paying a junior to go through it at all. Thomson Reuters’ 2025 review of the U.S. legal market describes a sector under pressure from client demands, competition and advancing AI capability. LexisNexis, in a 2024 survey, found that a majority of law-firm respondents believed generative AI would affect the apprenticeship model of large firms. That is a precise way of naming the risk. The issue is not that legal practice disappears. The issue is that the mechanism through which firms turned beginners into lawyers begins to lose its economic footing. 

The shift is already visible in legal education. In California, regulators are considering a proposal that would require students at the state’s accredited and unaccredited law schools to complete practice-based training on the competent use, capabilities and limits of AI. That move matters less as a curriculum story than as a professional signal. When a licensing system starts treating AI fluency as part of basic legal competence, it is acknowledging that the structure of entry-level work has changed. Students are being asked to prepare not for the older sequence of junior legal labor, but for a profession in which machines will handle more of the first pass and humans will be pushed earlier toward verification, framing and responsibility. 

Accounting presents the same structural problem in a different register. The profession has always relied on junior-heavy work in audit preparation, reconciliation, documentation, testing and repetitive review. Those tasks were not glamorous, but they gave early-career staff a way into the system. They taught pattern recognition, materiality, error detection and procedural discipline. CPA.com’s 2025 report makes clear how exposed that layer has become. Workflow automation, intelligent audit assistants, AI tools for research and autonomous systems for repetitive tasks are no longer peripheral experiments. They are moving into the operating core of the profession. That does not eliminate the need for accountants. It does, however, narrow the amount of work that can still justify hiring large numbers of beginners whose value used to lie partly in doing exactly that kind of structured repetition. 

This is where the standard language of “augmentation” becomes too soft. In both law and accounting, AI may well amplify the output of the professionals who remain. But it also redistributes what kind of labor remains worth buying. The first tasks to be compressed are usually the tasks that were easiest to assign to juniors. The profession survives. The lower rung weakens. A firm can tell itself that nothing fundamental has changed because the title still exists, the clients are still there and the senior people are still billing. Yet the internal architecture is no longer the same. A thinner entry layer means a thinner training layer. Over time, that is not a staffing adjustment. It is a change in how the profession reproduces itself. 

That is why the debate should not be framed too narrowly as one of substitution. The more serious question is whether these professions can keep their institutional depth after stripping away so much of the work that once produced it. A law firm can still hire laterals. An accounting firm can still buy software and retain senior expertise. For a while, both may even look more efficient than before. The harder question arrives later, when the market wants experienced five-year lawyers, dependable mid-level auditors and managers who understand not only how to read an output, but how to judge it. By then, the missing layer will not be a theory. It will be a cohort that was never fully built.

The firm that trains less becomes a different kind of firm

A firm does not need to stop hiring altogether to change what kind of institution it is. It only needs to narrow the layer through which younger workers once entered, learned, and moved upward. That is how a staffing decision becomes a structural one. The organization may look intact from the outside. The same partners remain in place. The same managers run teams. Clients still see familiar titles. What changes is the composition beneath them. The base grows thinner. The middle stops thickening. More of the missing capacity is supplied from outside: software, contractors, laterals, offshore teams, specialist vendors. The result is not a smaller version of the old firm. It is a different one.

This is where the logic of efficiency begins to diverge from the logic of institutional depth. In the short run, reducing junior intake can appear rational. It lowers supervision costs. It cuts time spent correcting beginner mistakes. It reduces the burden of carrying staff whose immediate billable value is limited. If AI can absorb part of the first-round work, the choice looks cleaner still. A firm can keep expensive judgment at the top, trim the training layer beneath it, and tell itself it has become leaner. In narrow accounting terms, that may even be true. The difficulty is that organizations do not run only on current output. They also run on delayed formation. The people who hold the middle of a firm together are not purchased at the moment they are needed. They are produced years earlier.

Recent research suggests that this change is already underway in precisely that incremental form. A 2025 working paper on firms adopting generative AI found that junior hiring fell more sharply than senior hiring, and that the shift showed up less through dismissals than through weaker inflows. Stanford’s Digital Economy Lab reported a similar pattern at the occupational level: in AI-exposed work, early-career employment fell after the arrival of ChatGPT even as older workers in the same occupations saw gains. The signal is easy to miss because it lacks drama. Nothing collapses at once. The lower layer is simply replenished less fully than before. Over time, that is enough.

The likely destination is a more polarized organization. At the top sit senior professionals, partners, managers and specialists whose value lies in judgment, client handling, coordination and responsibility. At the bottom remains a smaller intake of juniors, more heavily screened and expected to become useful faster. Around them sits an expanding ring of technical tools and externalized capability: AI systems, contract workers, offshore labor, data specialists, implementation consultants. What thins is the broad middle—the layer once built by years of accumulated junior experience. It is this shape, more than any single layoff round, that may define the white-collar firm in the AI era.

There is a reason that structure carries long-term risk even when short-term performance looks strong. Middle layers do more than execute. They stabilize organizations. They supervise juniors, translate senior demands into workable process, catch weak analysis before it becomes costly, and carry institutional memory across turnover at the top. A firm that weakens this layer may not feel the loss immediately because the senior cohort can absorb the strain for a time. But that arrangement is self-consuming. If too few people are moving through the system, the organization begins drawing down future capability to preserve present margins. What looks efficient in year one may become a staffing problem by year five and a leadership problem by year ten.

That is why the real danger is not simply a “jobless future,” a phrase that usually says less than it seems. The more plausible risk is thinner institutional memory, weaker internal promotion pipelines, and a labor market that becomes increasingly dependent on buying finished expertise from elsewhere. OECD work on skill gaps has already pointed to the broad menu firms use when internal capabilities fall short: external hiring, task reallocation, automation, domestic outsourcing, international outsourcing. Those responses can solve immediate problems. They do not solve the problem of how a profession renews itself from within.

The pressure, then, is not only on workers trying to get in. It is on firms deciding what kind of institution they intend to remain. A company that stops training at scale is not merely cutting labor costs. It is changing its dependence structure. It will rely more on the market, more on outside capability, more on imported judgment, and less on the slow internal accumulation through which expertise once became native to the organization. For a while, that may look like modernization. The question is what remains when the market tightens, laterals become scarce, and the people who were supposed to have grown into the middle were never there in sufficient numbers to begin with.

A labor market that produces less middle

The organizational shift does not remain inside the firm. Repeated often enough, it begins to alter the shape of the labor market itself.

That is the larger danger in the present moment. A company can decide to hire fewer juniors and rely more heavily on software, laterals, and outsourced capacity. One firm doing so is a staffing choice. Hundreds of firms doing so becomes a labor-market pattern. The immediate result is not always visible in unemployment alone. It appears instead in the narrowing of credible entry paths, the thinning of mid-career cohorts, and the growing distance between those who are already positioned to benefit from AI and those who are trying to enter the system under new and harsher conditions.

That divide is already visible in the data. OECD analysis suggests that the gains associated with AI are more likely to accrue to workers who are already higher-skilled and higher-paid, while younger and lower- or middle-skilled workers face greater adjustment risks. IMF work on skill demand points in a similar direction. New AI-related skills can improve wages and employability, but the returns are uneven. The result is not a single labor market moving forward together. It is a sorting process. Some workers gain leverage because they can use AI to amplify existing expertise. Others face a different reality: the work through which they might once have built that expertise is no longer as available, as paid, or as institutionally protected as it used to be.

That is what makes the erosion of entry-level work more serious than a temporary hiring cycle. The old white-collar labor market did not simply distribute jobs. It distributed trajectories. A junior analyst did not remain a junior analyst. A first-year associate was not meant to stay on the first rung. The system was valuable because it offered a route—imperfect, unequal, often exhausting, but real—through which large numbers of workers could move into stable, skilled, mid-career positions. If those routes narrow too far, the economy does not just lose openings at the bottom. It loses one of the main mechanisms through which professional stability used to be built in the middle.

This is where the language of “upskilling” can become misleadingly thin. A labor market cannot sustain itself on abstract training alone. Courses, credentials, and AI fluency matter, but they do not substitute for structured exposure to actual work. Judgment is not downloaded. It is acquired under conditions of repetition, error, correction, supervision, and escalating responsibility. Once the jobs that provided those conditions begin to thin out, the problem is no longer only whether workers can learn the new tools. It is whether they still have enough places in which learning leads to durable professional standing.

A more polarized labor market follows almost naturally from that shift. At one end sit workers whose experience, credentials, institutional access, or existing client-facing role make them complementary to AI. They become faster, more productive, and often more valuable. At the other end are those competing for a smaller number of entry points, under conditions that demand immediate polish and demonstrable usefulness. What weakens in between is the broad layer of workers who once had time to become competent before they were expected to be exceptional. The risk is not simply more inequality in pay. It is a thinner social and professional middle.

That matters beyond economics in the narrow sense. Mid-career professionals are not only units of labor. They are the people who hold institutions together in ordinary time. They manage teams without running the entire organization. They translate expertise across functions. They train younger staff, absorb shocks, and keep complex systems legible to the people inside them. A labor market that produces fewer such people may still generate elite stars and efficient firms. It may still look technologically advanced. What it loses is depth.

This is why the question raised by AI is larger than whether some occupations will shrink. The deeper question is whether advanced economies are drifting toward a model in which a smaller number of workers capture more of the productivity gains while a larger number struggle to find durable paths into skilled professional life. If that happens, the problem will not be confined to one profession or one hiring cycle. It will be built into the structure of the labor market itself.

Who is still responsible for making professionals?

Public policy has started to respond to AI. It has not yet fully responded to what AI is doing to the structure of professional formation.

That distinction matters. Governments are not ignoring the technology. South Korea has already moved to build a formal policy framework around AI, and public discussion has expanded around competitiveness, infrastructure, workforce preparation, and the strategic need to keep pace with other advanced economies. OECD analysis of Korea reflects that wider shift: the policy conversation is increasingly organized around productivity, technological adoption, and the distribution of gains and adjustment costs across different groups of workers. But that still leaves a narrower and more difficult question underexamined. Who, in an AI-era labor market, remains responsible for producing the next generation of competent professionals?

The usual answer is some combination of universities, training programs, and workers themselves. That answer is no longer sufficient. Formal education can adapt to new technical demands. Professional schools can add AI literacy. Regulators can revise standards of competence. California’s discussion of AI training in legal education is one example of that institutional adjustment. But none of those measures can fully replace the function once performed by the workplace itself. A labor market does not produce experienced professionals through coursework alone. It produces them by giving beginners room to do real work under conditions of supervision, correction, repetition, and escalating responsibility. Once that layer starts to thin, the question is not simply whether education is modern enough. It is whether the economy still contains enough training-bearing jobs to turn instruction into expertise.

This is where much of the current policy language feels one step removed from the actual problem. Terms such as reskilling, upskilling, and AI readiness imply that the central task is to help workers catch up with new tools. In one sense, that is plainly necessary. Workers do need new technical fluency. But the larger institutional issue is not fluency alone. It is capacity. A society can expand AI training while still allowing the number of jobs that build judgment to contract. It can produce more certificates while producing fewer durable routes into professional maturity. Training policy, in other words, is not the same thing as training capacity.

The same applies to the increasingly attractive language of efficiency. Firms under pressure to adopt AI are often encouraged to become more productive, more flexible, and more competitive. That logic is not wrong. Yet taken too narrowly, it encourages a model in which the costs of formation are treated as inefficiencies to be removed rather than investments to be sustained. The beginner who requires supervision looks expensive. The task that can be automated looks wasteful. The workflow that once took time looks ripe for compression. In the short run, each judgment can be defended. Taken together, they begin to hollow out the institutional base from which the next layer of professionals would have emerged.

The policy difficulty is that no single actor has a clear incentive to solve this problem alone. Individual firms may prefer to hire fewer beginners and buy finished expertise later. Workers cannot create apprenticeship structures by themselves. Universities can teach more, but they cannot substitute for the slow transfer of responsibility that happens inside actual organizations. Governments, meanwhile, are often best positioned to see the long-term supply problem but least equipped to redesign firm behavior without making choices that are politically and economically contentious. That is why the issue remains underdefined. It sits awkwardly between labor policy, industrial policy, education policy, and professional regulation, while belonging fully to none of them.

Yet the costs of ignoring it are likely to arrive in delayed form rather than immediate crisis. There may be no dramatic moment when the system announces that it has stopped producing enough mid-career professionals. The shortage will first appear as something else: a scramble for laterals, rising dependence on contractors, thinner internal promotion pools, overloaded managers, weaker mentoring, and a widening gap between elite entrants and everyone else. By the time it is recognized as a pipeline problem, the cohorts that should have filled the middle may already be missing.

That is the blind spot. Much of the policy response to AI still assumes that the main challenge lies in helping workers adjust to technological change. A more serious challenge may be whether institutions can still absorb workers in ways that allow them to become the kind of people advanced economies will later depend on. The question is no longer only how to make labor more AI-ready. It is how to preserve, or rebuild, the conditions under which labor becomes capable in the first place.

The deeper crisis is the erosion of professional time

The most serious danger in the age of AI may not be that work disappears too quickly. It may be that the time once required to turn inexperienced workers into capable professionals is no longer being protected by firms, rewarded by markets, or even fully recognized by policy.

That is a different kind of labor-market crisis from the one most public debate still imagines. The familiar argument assumes that technology destroys some jobs, creates others, and forces workers to adapt. There is truth in that account, but it remains too flat for what is now unfolding inside white-collar professions. In law, accounting, and other structured fields, the first thing being compressed is not expertise at the top. It is the lower layer of work through which expertise used to be made. Korea’s own labor data already suggest that the early burden has fallen disproportionately on younger workers in AI-exposed sectors, while older workers in those same sectors have remained more protected. That is not simply a generational imbalance. It is evidence that the labor market is beginning to economize on formation itself.

That shift carries a consequence that standard debates about “job displacement” do not fully capture. A profession does not renew itself automatically. It requires years of tolerated inefficiency, supervised repetition, minor error, correction, observation, and gradual exposure to consequence. In other words, it requires time that is not yet fully productive. For decades, institutions absorbed that time because they understood it as part of the cost of reproducing competence. AI changes that calculation. Once software can perform part of the beginner’s work faster and more cheaply, the market stops seeing the lower rung as an investment and starts seeing it as drag. The danger begins there.

That is why the long-term risk is not only a smaller number of junior hires. It is a thinner moral and institutional understanding of what professional development actually requires. A society that forgets how expertise is formed may continue, for a while, to consume expertise that was built under older conditions. Senior professionals remain. Firms still function. Clients still receive service. Productivity may even rise. But the appearance of continuity can be misleading. If fewer workers are allowed to move through the slower, lower, formative stages of a profession, then what disappears is not simply a job category. What disappears is a social mechanism for producing reliability, judgment, and institutional memory at scale.

That is the point at which the problem becomes larger than labor economics in the narrow sense. It becomes a question about the kind of society advanced economies are choosing to become. One possibility is a labor market that grows more technically capable and more socially thin: a small number of highly productive, AI-complementary professionals at the top; a narrow and highly selective stream of entrants beneath them; and a much wider outer ring of contingent, outsourced, or permanently preparatory labor. Such a system may be efficient. It may even be innovative. But it would also be more brittle, more unequal, and less able to reproduce the dense middle layers on which institutions depend. OECD and IMF work already points to this broader risk of uneven gains, higher barriers for vulnerable groups, and a more polarized labor market if the benefits of AI continue to flow disproportionately to already advantaged workers.

The unresolved question, then, is not whether AI can do more of the work. It can. Nor is it whether firms will continue to adopt it. They will. The harder question is whether institutions that no longer spend enough time on beginners can still expect to inherit enough capable people later. That question is economic, but it is also civilizational. It goes to how a society thinks about continuity: whether competence is something it can keep buying on demand, or something it must keep cultivating even when cultivation looks inefficient.

That is where the policy conversation remains too shallow. The dominant language is still one of reskilling, AI readiness, and competitiveness. Those things matter, but they do not reach the center of the problem. A labor market can become more AI-literate while growing worse at making professionals. It can teach workers how to use tools while offering them fewer places in which to acquire judgment. It can celebrate adaptability while quietly dismantling the institutions that once turned adaptation into durable adulthood. The real challenge is not only to help workers adjust to AI. It is to decide whether the economy will still bear the cost of forming people.

In the end, that may be the clearest way to state what is at stake. The central issue is no longer whether AI will replace professionals. It is whether a system that stops paying for beginner work can still reproduce professionalism itself.