Korea's AI Cost Screen Reaches 8% to 12% of Jobs. Current Practicality Cuts That to About 2%.

Korea Development Institute estimates that 8.1% to 11.7% of Korean employment would clear its cost test if every task with an identifiable AI modality became technically automatable. Weighting that already cost-eligible set by current practical automation scores cuts the share to 1.4% to 2.2%. The industry sales measure falls from 10.5% to 1.8% (KDI).

Neither range is raw technical exposure. Both apply labor and system costs. They are modeled task shares, not observed adoption, layoffs or forecasts of payroll change. The gap measures how much the current practical automation score reduces a future-capability economic case.

KDI applies three distinct filters

KDI decomposes 4,206 job descriptions across 537 occupations into 6,824 object-and-action tasks. It scores each task on whether outcomes are measurable, suitable data exist, context matters, human judgment is required and an agent can execute the work (KDI monograph).

Raw exposure is the first concept. It asks whether AI could affect a task and captures technical reach without proving that deployment would pay. KDI then constructs a future-capability economic case by setting the practical automation score to one for modality-defined tasks and applying its labor-versus-system cost threshold. Finally, it weights that cost-eligible set by the current practical automation score, which incorporates today's measurement, data, context, judgment and agent constraints.

The reported comparison is between the second and third concepts. In the occupational dataset, cost-eligible employment under assumed future capability is 11.7% of workers. Current practical weighting reduces it to 2.2%. In the industry dataset, the comparable move is from 8.1% to 1.4%, while the sales measure falls from 10.5% to 1.8% (KDI).

The future-capability measure still is not a count of jobs destined to disappear. It spreads occupation-level tasks across firms and treats a cost-eligible task as affected. Affected work can be automated partly, reorganized or complemented. It is not equivalent to a terminated position.

Raw exposure measures technical reach. The upper range measures cost eligibility under assumed future capability. The lower range applies today's practical automation weights to that set.

Integration costs set the minimum scale

The cost model explains most of the compression. KDI assumes a foundation-model system requires an initial team of architects, developers and a business analyst. Design, development and integration cost about KRW252 million. Fixed maintenance adds about KRW51 million a year. The model uses a five-year recovery period, a 5% discount rate and 22% annual hardware-efficiency improvement (KDI).

Data costs vary with labeling time, worker skill, modality, confidentiality and availability. Training uses an H100 rental assumption of KRW3,968 an hour. Inference scales with task time. Labor savings use 1.26 times reported wages to include employers' non-wage costs. For each task, KDI divides the system cost by the five-year labor cost per worker to derive a minimum viable headcount.

That structure makes wages and firm scale as important as model prices. Higher pay raises the saving from automation. More workers performing the same task spread fixed integration costs further. Lower training and inference prices help, but cheap tokens do not configure internal data, redesign approvals or maintain a production system.

KDI's sensitivity analysis makes that point explicit. Cutting AI costs by 50% improves the modeled total-factor-productivity result less than raising the productivity effect. The report says many firms still employ fewer people in a given occupation than the reduced break-even headcount requires (KDI). Reasonable model-cost relief therefore does not collapse the integration constraint.

This is also why a shared service can change the result more than another price cut. Reusing one system across firms or tasks spreads the fixed bill across more labor. The report models some sharing within occupational groups, but it does not assume frictionless access to a universal automation utility.

The firm result is a digital bundle

KDI then studies Korea's Survey of Business Activities from 2017 through 2023. The panel covers larger firms and uses firm and year effects. Its preferred instrumental-variable design combines industry-level economic viability with adoption rates that exclude the firm being measured (KDI).

In the preferred IV model, the estimated coefficient implies that firms using AI plus technologies including cloud, big data, internet-of-things systems and robotics had 20% higher sales per permanent employee than firms using none of the bundle. The same model estimates 13% higher total sales. Its 5% reduction in permanent employment is not statistically significant. A matched comparison estimates about a 5% increase in sales per worker in the adoption year, with no clear persistent effect later. KDI summarizes the cross-method productivity range as roughly 5% to 20%.

The AI-only test is weaker. After excluding firms using other fourth-industrial-revolution technologies, the coefficients for sales per worker were about plus 1% and minus 1% across two specifications, with neither distinguishable from zero. Cloud, data infrastructure, software, management quality and prior organizational capacity can all contribute to the bundle result. It cannot be assigned to AI alone.

The identification remains demanding. Firms choose when to adopt. The instrument is built partly from the same viability framework used elsewhere in the report, so its validity depends on industry cost conditions affecting performance through adoption rather than another route. Matching removes observed differences, not unobserved management quality, and its post-adoption window is short.

The useful forecast is smaller

KDI combines viable task shares, the firm productivity coefficient and industry sales weights in a general-equilibrium model. Its baseline projects a cumulative 2.3% TFP gain over a decade, or about 0.23 percentage points a year (KDI).

That is a scenario, not a measured growth contribution. Adoption paths and productivity parameters are imposed. The model focuses on task automation and omits structural transformation, intangible-capital accumulation, idea creation and discontinuous capability gains. It also feeds the 20% digital-bundle estimate into the baseline, not the statistically weak AI-only coefficient.

The practical implication is narrower than an economy-wide labor shock. Automation leads where wages are high, tasks repeat across enough workers and data systems already exist. Enterprise buyers should calculate break-even headcount by task before buying capacity. Workforce planners should treat exposure rankings as maps for job redesign, not redundancy schedules. Labor economists should keep the 2% current viability estimate separate from realized displacement.

Capability determines what enters the spreadsheet. Integration determines what gets funded.