AI may widen Korea’s professional labor divide before it raises productivity

Picture

Member for

1 year 10 months

Real name

Gordon Institute Policy Forum Editor

Bio

Gordon Institute Policy Forum Editor

Input

Sep 4, 2026 09:50

Modified

Sep 7, 2026 06:55

AI can raise the productivity of some Korean professionals while weakening entry routes and routine roles elsewhere.
The distributional effect depends on task structure, supervisory capacity, firm organization, and access to learning—not adoption alone.
Korea needs transition indicators that reveal displacement, degraded career ladders, and concentrated gains before aggregate productivity responds.

The Sequence of the Labor-Market Shock

Korea’s AI Paradox: High Adoption, Low Productivity describes a high-adoption economy in which measured productivity gains remain uneven. That pattern should not be read as evidence that AI is economically unimportant. Labor-market restructuring can occur before national accounts show a clear productivity acceleration. Firms may reduce outsourced work, compress junior teams, or expand output without proportionate hiring while the aggregate effect remains difficult to isolate.

The ILO’s Generative AI and Jobs: A Refined Global Index of Occupational Exposure emphasizes that exposure is not identical to job loss. Most occupations contain task bundles, and transformation is often more plausible than complete automation. The immediate Korean risk is therefore polarization inside occupations: workers able to define and supervise AI-supported work gain scope, while workers whose contribution consisted mainly of producing routine intermediate output lose a training route.

This sequence matters in a labor market where credentials and entry positions have traditionally provided the bridge to professional expertise. If the bridge narrows, the long-run supply of supervisors may eventually weaken even as incumbent specialists become more productive.

A Production Function with Organizational Capital

Let firm output be

Y = A H^{\alpha} I^{\beta} P^{\gamma} O^{\delta},

where $H$ is human capital, $I$ is intellectual capital, $P$ is physical and technological capital, and $O$ is organizational capital.

The marginal product of AI-related physical capital is

\frac{\partial Y}{\partial P} = \gamma A H^\alpha I^\beta P^{\gamma-1} O^\delta.

The return to technology rises with $O$. The cross-partial derivative

\frac{\partial^2Y}{\partial P\,\partial O} > 0

captures the complementarity. A better organizational system increases the value of technical investment.

The Workflow as a Causal System

AI adoption changes behavior. Once a score affects a decision, it can change the future data.

Suppose a lender approves applicants when $\widehat p(x)\geq\tau$. Repayment is observed primarily for approved applicants. The policy shapes the next training sample:

\mathcal D_{t+1} \sim P \left( X,Y \mid a_t(X)=1 \right).

If the organization treats the model as a static forecasting object, it may overlook selective labels, feedback loops, and changes in applicant behavior.

Organizational capital is needed to define:

who monitors the data-generating process;
which outcomes are observed;
when thresholds change;
how overrides are recorded;
who investigates distribution shift; and
when the system should be suspended.

Model governance is part of production, not an administrative afterthought.

Automation Can Scale Weakness

Before automation, a flawed process may be slow and inconsistent. Automation can make it fast and consistent without making it correct.

Let each decision create expected value $v$ and let the system make $n$ decisions. Total expected value is

$$ V = n v. $$

If $v>0$, scale is beneficial. If $v<0$, scale increases loss.

This simple equation explains why efficiency metrics are insufficient. An AI system that processes twice as many cases may produce twice as much value or twice as much harm.

The organization must validate the sign and distribution of $v$, not only the speed of execution.

Human Capital Inside Organizational Capital

Specialists need more than technical depth. They must understand the institutional interfaces through which their work creates value.

This includes the ability to:

translate an operational concern into a modelable question;
identify missing stakeholders;
distinguish predictive accuracy from decision value;
communicate uncertainty to non-specialists;
document assumptions and revisions;
recognize incentives that distort data; and
design a feedback loop.

These capabilities are not “soft” alternatives to mathematics. They determine whether the mathematics belongs to the decision.

At the same time, organizations should not expect one technical employee to repair every interface. Management is responsible for constructing the system in which expertise can operate.

Task Bundles and Unequal Gains

Let occupation $j$ contain tasks $i=1,\ldots,n$ with employment shares $s_{ij}$. AI changes the cost or productivity of each task by $a_i$, while complementary human capability contributes $h_i$. A simple exposure balance is

\Delta V_j=\sum_i s_{ij}(h_i-a_i).

A negative value does not mechanically predict the disappearance of the occupation. Tasks can be reorganized, demand can expand, and regulation may preserve human responsibility. But the expression identifies why one title can contain winners and losers. Research, negotiation, diagnosis, and accountability may become more valuable while first drafts, routine analysis, and standard documentation become cheaper.

The firm captures the gain only if it redesigns the workflow. Otherwise workers produce more drafts without changing decisions, or senior staff spend the saved time correcting a larger volume of weak output. Productivity and labor displacement are therefore both mediated by organizational capital.

The Missing Entry-Level Ladder

The displacement of routine work creates a developmental problem. Entry-Level Jobs Vanish as the Superhuman Labor Premium Widens highlights the widening premium around workers who can already combine domain expertise with AI. Yet many of those workers learned through the very junior tasks now being automated: preparing comparisons, checking data, drafting standard documents, and observing how senior colleagues corrected them.

Employers cannot assume that universities will replace this apprenticeship automatically. Academic knowledge and workplace judgment are complements, not substitutes. Firms may need explicit supervised pathways in which junior employees review generated work, investigate exceptions, and defend changes. These roles produce less immediate output than full automation but preserve the future supply of capable reviewers.

Korea’s large-firm and small-firm divide makes this challenge distributional. Large organizations can redesign roles, provide proprietary data, and absorb training costs. Smaller firms may simply purchase tools or external services, leaving their employees with less opportunity to build supervisory capability.

Indicators for an Earlier Policy Response

Aggregate unemployment will arrive too late as the only warning indicator. Korea should track vacancy changes by task content, entry-level hiring, promotion time, subcontracting volumes, wage dispersion within occupations, and the share of AI output reviewed by workers with relevant authority. These measures reveal the mechanism of transition.

Training policy should follow the mechanism. Workers do not need a generic promise to become AI experts. They need routes from displaced production tasks into verification, domain integration, exception handling, and client responsibility. Programs should disclose which occupational transition they support and test whether participants can perform it.

A policy that concentrates only on model access may increase the productivity of already capable workers while doing little for those whose jobs are being decomposed. Distribution is not an afterthought to adoption. It is part of the implementation system.

Different Transitions Across Korean Firms

The transition will differ sharply by firm size and market exposure. Export-oriented manufacturers and large service firms face international cost benchmarks and can justify investments in internal data, security, and model governance. A small domestic supplier may receive the same tool through a cloud subscription but lack the scale to redesign roles or measure whether the output improves margins. National adoption statistics can therefore rise while the productive divide between firms widens.

Subcontracting complicates the picture further. A large company can use AI to internalize analytical or software work that previously supported employment at specialist vendors. The measured productivity gain appears at the client, while displacement and lost learning appear elsewhere in the supply chain. Labor statistics organized by employer may miss the connection unless task flows and contract spending are examined together.

Regional effects may also be uneven. Firms in Seoul and other dense professional markets can recruit supervisors and change vendors when an implementation fails. Organizations in thinner markets may have access to the same model but fewer people able to evaluate it. Transition support should therefore combine digital access with shared assessment resources, sector institutions, and routes to external expertise.

Worker transitions will also differ by contract status. Permanent employees may be reassigned into review or client-facing roles, while temporary workers and freelancers lose assignments without appearing in a formal redundancy count. Survey evidence should therefore follow hours, earnings, and task composition across employment forms. Otherwise, a smooth headline employment series can conceal a rapid decline in the volume of work available to people outside protected internal labor markets.

The Distribution of Bargaining Power

Productivity does not determine how gains are divided. A worker who owns scarce domain knowledge, client relationships, or authority over a regulated decision may capture part of the surplus. A worker performing a standardized intermediate task may face competition from both domestic automation and global providers using the same tools. Firm concentration and employment protection will shape how quickly those differences appear in wages.

Korean policy should distinguish protection of a particular task from protection of a transition pathway. Preserving obsolete production steps can reduce firm competitiveness without building capability. Wage insurance, targeted learning accounts, portable benefits, and assessed work-based training can support movement while allowing the production system to change.

The distributional objective should be stated explicitly. If public support lowers adoption costs for firms, policymakers should ask whether complementary support reaches workers and smaller suppliers. Otherwise, subsidies can accelerate capital deepening while leaving the adjustment cost concentrated among those least able to finance their own transition.

Table 1. Early signals of professional labor-market polarization

Signal	Possible interpretation	Necessary qualification
Fewer junior vacancies	Routine production is being automated	Demand or cohort size may also have changed
Higher senior productivity	Expertise complements AI	Output quality and hours must be measured
Falling subcontracting	Work is returning in-house with AI	Tasks may have moved to another vendor
Wider within-role wages	Supervisory capability earns a premium	Firm and industry composition matter

Source: Author’s analysis

Conclusion

AI may alter Korea’s professional labor market well before it produces a clean aggregate productivity result. The earliest effects are likely to appear in task allocation, junior recruitment, outsourcing, and the scope of high-capability workers.

The appropriate response is neither to promise universal replacement nor to dismiss exposure because occupations survive. Korea needs institutions that observe the transition and preserve credible routes into the judgment-intensive work that remains. Those routes should be evaluated through earnings, mobility, durable employment, and demonstrated responsibility rather than course participation alone over time and across different employers.