1. The Paralegal at the Approval Button
March 2025, Seoul, Gangnam.
A law firm on Teheran-ro. The fourteenth floor. 9:07 a.m. Lee Jin-hee, age thirty, sits down at her monitor.
Three years of law school. ₩40 million in student loans. She has been at this firm for fourteen months. In law school she memorized contract law precedents, analyzed the articles of the Civil Code's general provisions, argued in moot court. The desk she occupies after graduation sits in front of an AI contract-analysis system.
On the screen, 57 contracts processed overnight are lined up in a queue. A color indicator appears beside each entry. Green — "No issues": 49 items. Yellow — "Review recommended": 6 items. Red — "Attention required": 2 items.
Lee Jin-hee clicks a green entry. An AI-generated analysis summary appears. Contract type, key clauses, risk-assessment score. She reads the summary and presses the approval button. Next entry. Click, read, approve. Click, read, approve.
Processing all 49 takes 12 minutes.
The 8 yellow and red items are forwarded to a partner attorney. Forwarding takes 3 minutes.
Fifteen minutes. Lee Jin-hee's core morning work is done.
Six months ago this work was different. She read the contracts herself. She tracked each clause with her eyes, identifying risk elements. "This indemnity clause is disadvantageous to our side" — rendering that judgment was her job. One contract took 40 minutes. Fifty-seven contracts meant three days' work.
Now it takes 12 minutes. Efficiency has improved. The firm is satisfied.
Lee Jin-hee is not satisfied. More precisely, satisfaction and unease coexist. What does she fill the remaining hours with, after work that finishes in 12 minutes? What she does in that time is review the outputs of other AI tools. She checks summaries extracted by a case-search AI, proofreads naeyong-jeungmyeong — formal legal demand letters — drafted by a legal-document AI. All day, she confirms what AI has produced.
In Chapter 3, Lee Jung-hoon returned his employee badge. A man who had spent 28 years listening to sounds and sensing the tremors of a mold. The moment his expertise became unnecessary, the badge became a piece of plastic.
Lee Jin-hee has not returned her badge. She still commutes. She still receives a salary. But the substance of what she does has changed. From someone who "reads" contracts to someone who "confirms" what AI has read. From judgment to approval.
The verb has changed.
It is the same position as Lee Jung-hoon's confirmation button. A different industry, the same structure.
2. The Logic of Task Decomposition — Jobs Are Not Replaced; Tasks Are
The sentence "AI replaces jobs" is imprecise.
The accurate unit is not the job but the task. The origin of this shift traces back to the task model that David Autor, Frank Levy, and Richard Murnane proposed in their 2003 paper. Their core distinction ran along two axes: routine versus non-routine; cognitive versus manual. Crossing those two axes yields four types.
Routine manual is the repetitive work of the assembly line. Industrial robots have been replacing it since the 1980s. The factory landscape we saw in Book 1.
Routine cognitive is accounting data processing, standardized legal document review, straightforward translation. Computers have been replacing it, and LLMs have accelerated the pace.
Non-routine manual is the domain of adaptive physical manipulation in the real world — plumbing, cleaning. Because the body must adapt to the environment, robots have not yet fully replaced it.
Non-routine cognitive is creative problem-solving, complex negotiation, strategic judgment. Until 2022, it was regarded as "the domain immune to automation."
The 2003 Autor-Levy-Murnane prediction was clear. Routine tasks would continue to be automated; non-routine tasks would remain the human domain. This prediction held broadly valid until GPT. When spreadsheets replaced accounting data processing and OCR replaced document classification, the pattern was "automation of routine cognitive tasks." Autor's analysis was correct.
But the model's implicit premise — "non-routine cognitive tasks will not be automated" — began to collapse after 2022. Writing, legal analysis, coding, medical diagnostic assistance — non-routine cognitive tasks entered the domain of LLMs. Autor himself acknowledged this at a 2022 MIT lecture: "My 2003 model did not anticipate LLMs."
The scale of the change was quantified in 2023. The study by Eloundou et al. analyzed LLM exposure using the U.S. Bureau of Labor Statistics O*NET occupational database. Three core findings emerged.
Eighty percent of U.S. occupations are exposed — at least 10 percent of their core tasks face LLM exposure. Occupations with "no exposure" amount to only 20 percent of the total.
Nineteen percent have 50 percent or more of their core tasks exposed. In those occupations, more than half the work could potentially be replaced or augmented by LLMs.
The third finding was the most counterintuitive. The higher the income and educational requirements of an occupation, the greater the exposure. Past automation replaced low-skill routine labor. LLMs operate in the opposite direction. It is not telemarketers or warehouse workers but lawyers, accountants, and analysts who show high exposure rates.
This finding stands in contrast to the 2013 Frey and Osborne "47 percent" prediction. Frey and Osborne estimated that 47 percent of the 702 U.S. occupations they analyzed were "at high risk of automation." But they used the whole job as their unit of analysis. When the OECD applied the same methodology at the task level, the figure dropped to 9 percent. The gap between 47 percent and 9 percent is a gap in what is being measured. Using the job as the unit overstates; using the task as the unit gives precision.
McKinsey analyzed 800 occupations and 2,100 detailed work activities, estimating that generative AI carries the technical potential to automate 60–70 percent of current work activities. This is because generative AI has extended into non-routine cognitive territory: natural-language understanding, document generation, coding.
Goldman Sachs analyzed that AI automation could expose the equivalent of 300 million full-time jobs worldwide. This does not mean 300 million people will lose their jobs. It means that when the hours of automatable tasks are summed, they add up to the equivalent of 300 million full-time positions. Goldman Sachs issued a self-correcting report in 2024, acknowledging that productivity gains relative to AI investment have not yet been proportional — that task redistribution is occurring but that institutional and human adaptation costs are required before that shows up in productivity statistics.
The IMF estimated that approximately 60 percent of occupations in advanced economies are exposed to AI. In low-income developing countries, the figure is 26 percent — because advanced economies have a higher concentration of cognitively intensive occupations. The paradox: workers in more developed countries face greater exposure.
The variance among these figures is wide. But the direction is consistent. AI is absorbing tasks, not jobs, and that absorption is proceeding faster among high-education cognitive workers.
Exposure is not replacement. But exposure is where it begins.
3. AI Exposure by Occupation in Korea — A Portrait of the KSCO's Eight Groups
How does this structural force operate in Korea?
When the analyses of Eloundou et al., the IMF, McKinsey, and the Korea Employment Information Service are cross-applied to the eight major groups of the Korean Standard Classification of Occupations (KSCO), the outline of AI task-replacement rates by Korean occupation comes into view. The Korea Employment Information Service's 2022 analysis estimated that occupations at high risk of automation account for approximately 52 percent of the total — a figure higher than the United States, reflecting the proportion of clerical workers in Korea's labor market and the degree of routine-ization across occupations.
The AI task-replacement rate for clerical workers is estimated at 60–75 percent. Bookkeeping clerks, secretaries, administrative clerks, bank tellers. Data entry, document classification, handling standard inquiries, ledger posting — these are their core tasks. They are direct targets of LLMs. This group shows the highest exposure among all eight.
The professional group — physicians, attorneys, accountants, teachers, journalists — falls at 45–65 percent. Document drafting, case-law searches, standard diagnostic procedures, repetitive calculations are being eroded. What remains is complex judgment, patient and client relationships, ethical responsibility. This is a group in which routine cognitive tasks are concentrated.
Managers fall at 30–45 percent. Report writing, data analysis, and schedule coordination transfer to AI, but organizational judgment, stakeholder negotiation, and accountability remain, as these belong to the domain of relationships and authority.
Sales workers fall at 40–55 percent. Insurance planners, real estate agents — product-information provision, standard consultation, and quotation calculation have entered the zone of automation. High-value relationship sales, personalized persuasion, and trust-building, however, remain in the human domain depending on transaction scale.
Service workers, by contrast — care workers, cooks, hairdressers — remain at 15–30 percent. Physical care and relational trust, physical manipulation — these are the core of such occupations. Work done by the hands, work done by the body, is what remains.
Skilled workers — electricians, welders, auto mechanics — fall at 20–35 percent. Because non-routine manual tasks are central, the current technology still provides a buffer.
Elementary occupations fall at 15–25 percent. Physical on-site response and work in non-standardized environments provide a buffer at the limits of physical robotics.
The most striking point in these figures is the paradoxically high exposure of the professional group. KSCO Group 2, professionals (45–65 percent), is markedly higher than service workers (15–30 percent), skilled workers (20–35 percent), and elementary occupations (15–25 percent). The core finding of Eloundou et al. — "high-education, high-wage occupations are more exposed to LLMs" — applies without modification to Korea's occupational structure.
Korea's university enrollment rate is 76 percent, first in the OECD. A substantial portion of the ₩29.2 trillion spent annually on private tutoring is invested in preparation for law school, medical school, and the certified public accountant examination. The typical investment is 9–11 years of education and ₩100 million to ₩300 million. Every one of these occupations carries medium-to-high AI exposure.
The structure these numbers reveal is paradoxical. The places into which Korean society has concentrated its greatest educational investment are the places most exposed to AI. AI is dismantling the core of the equation "high education = safe occupation" — the routine cognitive tasks that justify the premium of high-credential work.
Task-replacement rates are not probabilities of job extinction. They do not mean that 60–75 percent of clerical workers will be dismissed — they mean that that proportion of the tasks they perform can potentially be transferred to AI. Actual employment changes depend on institutional choices, wage structures, and the speed of corporate adoption.
But when tasks are transferred, what moves into the space they leave behind? When nothing moves in, the job becomes a shell.
4. Three Maps — How Replacement Unfolds Differently in the United States, China, and Korea
Even for the same occupation, the pace of AI replacement differs by country. The question is not one of technology. It is a compound interaction of institutions, wage structures, and technology-adoption rates.
In the United States, replacement is market-driven. The base salary of a junior associate at a large law firm is $225,000 — roughly ₩300 million. At that wage level, the return on AI investment is immediate.
The United States' employment flexibility (at-will employment) allows companies to adjust their workforces quickly after AI adoption. The cycle of "technical feasibility → economic adoption → employment adjustment" turns within two to three years.
Direct effects are already visible among paralegals and junior associates. Some large law firms have begun freezing new paralegal hiring.
The first occupational groups to be hollowed out are legal support, financial analysis, and content production. They share a common feature: all had high-education barriers to entry. As AI lowers those barriers, the "premium of high-credential qualifications" is coming under pressure.
In Book 2 we analyzed the AI competition between the United States and China. The ripple effects of that competition are appearing as different wave patterns in the labor market.
In China, replacement is state-designed. AI judges, algorithmic management, and automated public services are progressing under government direction. AI prosecutors are being piloted in Shanghai. The scope of replacement extends to the judiciary and public administration — broader than in the United States. At the same time, the buffers for maintaining employment in the name of social stability are also stronger.
China's distinguishing feature is that AI adoption is coupled with labor surveillance. Algorithms measure worker productivity in real time, and that data is directly linked to wages and employment. Beyond task replacement, the algorithmization of human labor itself is advancing. The OECD analyzes this as the spread of "algorithmic management."
In Korea, replacement is market-driven but the institutional response is slow. Kakao Bank and Toss Bank's non-face-to-face loan assessment, Lunit and Vuno's imaging AI, Samjeomssam's automated tax filing — these are spreading rapidly under market logic. But the redesign of the education system, vocational training framework, and social safety net is lagging behind. By OECD standards, Korea's social protection spending as a share of GDP is 14.8 percent, well below the OECD average of 21.1 percent.
The gap between the pace of the market and the response of institutions is widening.
The distinctive factor that determines the pace of replacement in Korea is the intensity of the licensing system. Attorneys, physicians, certified public accountants, and tax accountants have their scope of practice defined by statute. Even where AI replacement is technically possible, work that must legally be performed by a human license-holder remains. This is why "AI replacement does not appear to be happening" in Korea. Tasks are already being transferred to AI, but because of licensing requirements, the jobs themselves are maintained.
What is being maintained — the job, or the job title? That difference is the definition of hollowing-out.
Korea's structural particularity is "late start, fast diffusion." The typical pattern: Kakao Bank launched in 2017 and within three years had rapidly captured market share in bank lending from the major commercial banks. The automation of bank teller functions that took five to seven years in the United States happened in Korea in three to four years. "Cliff-edge replacement" — where institutional buffers hold and then suddenly collapse — is Korea's distinctive risk.
The same pattern may repeat in law and accounting.
Before that possibility becomes reality, individuals must redesign their paths. But the demand to redesign one's path is also society's way of transferring responsibility — avoiding the redesign of structures.
5. What Remains When the Tasks Are Gone — Accountants, Radiologists, Legal Clerks
Three occupations, dissected. They are among the most common destinations for high-cost educational investment in Korea, and they stand at the front line of AI task replacement.
An accountant's work is composed of seven core tasks. Bookkeeping and journal-entry processing, tax-return preparation, financial-statement anomaly detection, standard audit procedures — the AI-replacement potential of these top four tasks ranges from moderate to very high. Samjeomssam and the HomeTax autocomplete function are absorbing personal and small-business tax filing. AI audit tools — Deloitte's Argus, PwC's GL.ai — have been adopted across all four of the Big 4 accounting firms.
What remains is interpretation of gray zones in tax law, M&A financial advisory, and the formation of audit opinions — the domain in which legal liability accrues to the individual certified public accountant; the domain where trust and judgment intersect.
When the small self-employed businesses that form the core client base for Korean tax accountants migrate to AI platforms, the bottom tier of demand in the tax-accountant industry shrinks. What remains is corporate tax strategy, handling tax investigations, and complex international transaction taxation. Only the domain that simultaneously requires expertise and trust is left. Hiring at Korean accounting firms — including the Big 4 — has been on a declining trend since 2023.
Of the radiologist's seven tasks, the top four — chest X-ray interpretation, CT and MRI quantitative analysis, mammography reading, and prioritization of anomalous cases — carry high AI-replacement potential. Lunit's chest X-ray AI is in use at more than 50 hospitals in Korea. Some studies have reported performance on par with specialists. The number of FDA-approved AI-assisted radiology diagnostic systems exceeded 100 as of 2023.
Given Korea's National Health Insurance fee structure, the per-reading reimbursement for imaging is low, and some hospitals see one physician reading 200–300 images per day. In this high-burden environment, AI assistance is not a "choice" but a "means of survival."
Yet the paradox accelerates: the more physicians depend on AI, the more their independent judgment erodes. When AI filters first, the opportunity for residents to "train the eye across thousands of films to recognize the boundary between normal and abnormal" diminishes. Radiology residency application rates have been declining in the United States since 2020.
Industry estimates suggest that the top four tasks among a legal clerk's seven — standard contract review, case-law and statutory research, legal document drafting, and litigation record organization — account for 60–70 percent of working hours. Domestic legal-tech firms such as Lawform and Casenote are already providing services in this space. GPT-4-based legal document drafting tools are spreading rapidly. Technically, two-thirds are in a state ready to be transferred to AI.
There is a particular pressure in Korea's legal market. Since the introduction of the law school system in 2009, 1,700 attorneys are produced each year. In a market that has not grown sufficiently to absorb that supply, a structure is forming in which AI replaces junior-level work. Supply increase and demand decrease are operating simultaneously.
The three occupations share a common pattern. The tasks AI is replacing are precisely the tasks junior professionals repeat during the first three to five years of their careers — the process through which they build the foundation of their expertise.
Bookkeeping is the process by which accountants develop their numerical intuition.
Repeated chest X-ray reading is the process by which radiologists inscribe the boundary between normal and abnormal in their eyes.
Reviewing hundreds of contracts is the process by which legal clerks develop the ability to intuitively recognize risky clauses.
When this "experience of repetition" is transferred to AI, the senior professionals of five to ten years hence will have to render high-level judgments without that foundation. Knowledge can be transmitted, but perception must be embodied. Routine repetition is what builds perception. When AI absorbs the routine, the developmental path of perception is absorbed along with it.
The lower rungs of the experience ladder are being removed. The long-term cost of short-term efficiency gains lies here.
6. Two Chairs — The Weaver and the Loan Officer
Nottingham, England, 1815.
Thomas Higden stands before the factory gate. Until ten years ago, he threw the shuttle with his own hands, decided the pattern at the loom, confirmed the texture of the finished cloth with his fingertips.
In Book 1 we followed the handloom weaver's wages as they fell eighty-two percent — from twenty-five shillings a week to four and a half shillings. But something had fallen before the wages. Control over himself.
What Higden does now is inspect the cut ends of cloth woven by the power loom. The machine determines the texture of the cloth. He checks whether any threads have broken, whether the pattern has shifted. The eyes do this work. The hands are no longer needed.
The name of the occupation is still "weaver." But the essence of the weaver — the artisan's skill of handling thread by hand, the capacity to feel the tension of the yarn with the fingertips, the ability to read the balance of warp and weft by eye — is gone.
The Luddite movement of 1812 is commonly described as a revolt of machine-breaking fanatics. Its substance was different. It was resistance to the disappearance of the occupation's meaning while the job title remained. Not anger at dismissal — resistance to hollowing-out.
Two centuries have passed. The tools are different. The structure is the same.
Seoul, 2025. Park Jae-yeong, age forty-three, a loan assessment officer at a major commercial bank, begins his morning work. On the screen, 34 loan applications processed by AI are arrayed. For each entry, an AI credit score, risk grade, and recommended decision are recorded.
Park Jae-yeong affixes his electronic signature to the approved items. The rejected items are processed automatically.
Twenty years ago, Park Jae-yeong visited the applicant's office in person, reviewed the books, read the look in the owner's eyes. His expertise was reading what did not appear in the financial statements — the owner's resolve, the instinct for the market, the experience of navigating a crisis. What he reads now is numbers on a screen. Those numbers were produced by AI.
The job title is still "assessment officer." The substance of the role has changed.
The total number of employees at the four major commercial banks declined from approximately 72,000 in 2015 to approximately 59,000 in 2023 — an 18 percent reduction. Over the same period, the number of customers per employee at internet-only banks exceeded that of major commercial banks by a factor of ten. As the model of "processing more loans without human assessment officers" spreads, the role of the assessment officer grows thinner still.
Higden lost the expertise of reading the texture of cloth through the sensitivity of his hands. Park Jae-yeong has had his capacity to read a business owner's future taken from him.
Both men still come to work. Neither has been dismissed.
But their expertise is no longer required.
Higden ultimately lost the job itself. The machines did everything more cheaply and more quickly. Park Jae-yeong will not lose his job as long as he retains "the authority to sign." The institutional shield holds.
How long will that shield last?
7. The Hollowing-Out of Occupations — The Title Remains, the Content Disappears
After tasks are transferred to AI, what remains for humans converges on three roles.
First, monitoring. Watching AI output and detecting malfunctions. Not active judgment but passive surveillance. Like the quality-control inspector on the factory floor, like the safety driver in an autonomous vehicle, the human confirms that AI is operating normally.
But there is a paradox. For humans to monitor AI properly, they must understand what AI is doing. As the volume AI processes increases, the time humans invest in deeply understanding each result decreases. The "monitor" becomes in substance a "confirmer" — skimming the surface for anomalous signals. What the aviation industry calls "autopilot over-reliance" is structurally inherent in every monitoring role.
Second, approval. Affixing a human signature to decisions AI has already made. This role arises in processes where a human's final sign-off is required for legal or institutional reasons. The assessment officer who signs the loan documents. The physician who reviews the clinical records drafted by AI. The accountant who issues an opinion on financial statements analyzed by AI.
The content of expertise is emptied out; what remains is the "qualification to sign" — the license.
This is a structure in which judgment without responsibility and responsibility without judgment are separated.
AI judges; the human signs. When AI is wrong, the defendant in the lawsuit is the human. A structure of bearing responsibility without having exercised judgment.
Third, exception-handling. Managing the anomalous cases AI cannot process. The core of expertise is preserved, but the volume of work is sharply reduced. From a positive vantage, this is the deepening of expertise; from a negative one, it is the thinning of the occupation. High expertise is required, but there are not enough cases to sustain it.
There is a structural risk in the exception-handling role. In ordinary times AI handles everything; then, suddenly, a difficult exception appears and the human must render an immediately top-level judgment. The higher the degree of automation, the lower the human's capacity for manual judgment. The higher the reliance on autopilot, the more pilots' manual flying skills deteriorate — and the more judgment delays increase in emergencies. This was one of the contributing factors in the Boeing 737 MAX accidents.
This structural risk is inherent across the professions.
Monitoring, approval, exception-handling. What these three roles share is that they are "what is left." Not roles that actively create, but roles that clean up after AI is done.
Acemoglu has argued that the current trajectory of AI development carries an excessive "automation bias" — investment concentrated on replacement rather than augmentation of human capabilities, entrenching a structure in which productivity gains accrue to capital while the value of labor declines. This is not a technological inevitability but a choice. The cost of that choice is being paid by the middle tier.
Hollowing-out is different from dismissal. Dismissal is visible and captured in statistics. Hollowing-out is invisible. Employment statistics report that a job "exists" — but they do not report what that job is filled with. The same applies to GDP statistics. When tasks drain away but the job remains, employment statistics appear normal.
Individuals, organizations, and policymakers alike mistake the situation: "There's still a job, so things are fine."
The real question is how long an occupation that is only a shell can remain a shell.
The total number of employees in Korea's financial sector declined from approximately 270,000 in 2015 to approximately 230,000 in 2023. In legal services, 1,700 attorneys are produced each year while legal-tech reduces demand for paralegals. Supply increasing as demand decreases, crossing. Hiring at accounting firms has been on a declining trend since 2023.
These figures do not speak of the extinction of occupations. These fields still exist. But as tasks drain away, the density of the occupation thins, the value of expertise declines, barriers to entry lower, and wages come under pressure. This is the economic expression of hollowing-out.
The death certificate for an occupation does not read "closed" — it reads "hollowed out."
The paralegal who read every contract with diligence. The radiology resident who interpreted images one by one. The assessment officer who visited the loan applicant's office in person. Those who were most faithful to the system are being pushed aside by that system.
8. At the Threshold
A job is not a simple list of tasks. A job is a set of specific tasks institutionalized as a single bundle. Why does a physician perform diagnosis, prescription, and patient consultation as a single occupation? Why does an attorney handle legal research, contract negotiation, and courtroom argument together? These bundles are not technical necessities but historical and institutional choices.
AI is unbundling them. Absorbing tasks unit by unit, leaving humans with monitoring, approval, and exceptions.
A job that has lost its tasks can follow one of three paths. Reconstituting the bundle with new tasks is transition. Persisting, shrunken, with only the remaining tasks is hollowing-out. Disappearing because nothing remains is replacement.
What is currently happening most broadly is not replacement (the third path) but hollowing-out (the second). Hollowing-out is less visible than replacement, harder to address, and more widely distributed.
"AI will replace X" is not a prediction. It is an analysis of the structural forces at work. Where the outcome of those forces converges remains open. What is certain is that there is no certain answer.
Lee Jin-hee pressed the approval button again today. Forty-nine items in 12 minutes. Her three years of law school taught her things it did not teach her. Contract law, civil procedure, Constitutional Court precedents — those it taught her. That her occupation could take this shape, it did not teach her.
From the paralegal's office in Gangnam to the chicken franchise in Mapo-gu, the landscape of Seoul is changing. The dismantling of cognitive labor does not happen only inside offices. Its reverberations reach the signs on the street. Where signs come down, a different kind of displacement is beginning.
Threshold Question: What tasks in your occupation is AI already performing? Remove those — what remains? Is what remains a job, or a job title?