Fiscal Consequences of AI Automation 

The data. The model is built on Korean data measuring the economy in 2025. Every simulation carries control totals that must stay the same for every run: 12.4 million wage workers in the modeled cells, a ₩667.7 trillion modeled wage bill, ₩2,658 trillion of value added, and ₩650.6 trillion of national revenue. If any of these fails to hold, the model stops. The model runs on 209 occupation-by wage cells (the populated cells of 9 KSCO occupation groups × 24 monthly-wage brackets), each carrying employment, wages, AI and robot exposure, contribution rules, and income-tax brackets. The cells cover 12.4 million of roughly 22 million wage workers: an establishment-survey frame that over-represents larger firms and excludes public administration, which is a headline limitation rather than a footnote. 

MOEL establishment wage survey (PAYM39, 2025) — employment and wages for 9 occupation groups × 24 wage brackets 

ILOSTAT occupation × industry matrix — how each occupation's wage bill spreads across industries 

Bank of Korea 이슈노트 2025-2 「AI와 한국경제」 — how cognitively automatable each occupation group is, as the displacement-prone share 

Webb (2020) robot exposure, mapped from US occupations — how physically automatable each group is; no Korean measure exists 

NABO fund projections — the National Pension (2025 reform), Health Insurance (Focus 162), and Employment Insurance (mid-term) reserve paths, taken as the baselines NABO 2025 receipts (Focus 137) and 총수입 — the corporate and VAT bases and total national revenue 

Statistics Korea 장래인구추계 2022–2072 — the working-age population path, low, medium, and high 

MOEL 구직급여 statistics — re-employment rates and benefit rules 

2026 statutory schedules — contribution rates (National Pension 9.5% with the ₩6.59 million monthly cap, Health Insurance 7.19%, long-term care 0.94%, Employment Insurance 1.8%, industrial accident 1.47%), income-tax brackets, the 10% local surtax, 10% VAT, and the 24.2% effective corporate rate 

지역별고용조사 2025 H1 — the provincial occupation mix behind the map 

The scenarios. Every scenario is an adoption path, the share of displacement-prone work that is automated in each year, applied to the same Korean data. Three headline scenarios form a diffusion family; the other seven translate published forecasts (Acemoglu, Brynjolfsson, Karger et al., Metaculus, AI 2027, Korinek–Suh) onto Korean occupations without re-calibration, so that Korea and the United States can be read side by side. 

Slow adoption — 10% of exposed work by 2035: a linear ramp from 0.5% in 2026. Korea's small and medium firms hold adoption back for the decade.

Central — 20% by 2035: from 1% in 2026. The early-adoption levels observed in the United States, discounted for Korea's SME lag; the end point is the share of exposed work that Acemoglu and Svanberg find profitably automatable within ten years, discounted for Korea's adoption lag. Rapid adoption — 40% by 2035: from 2% in 2026. Half of what is technically feasible is realized, with Korea's ICT readiness driving a catch-up. 

Exposure, meaning which work counts as displacement-prone, is the Bank of Korea's classification in all three: 27% of employment sits in high-exposure, low-complementarity occupations, concentrated in clerical work. The US Budget Lab measure, mapped onto the same occupation groups, agrees on clerical work (0.84 against BOK's 1.00) and disagrees on professionals (0.71 against 0.22), managers (0.69 against 0.00), and sales (0.52 against 0.36); it is kept as a disclosed cross-check, not the headline. The three share one labour-market convention. Displaced workers are re-employed at 25% a year at wages 13% lower, draw 구직급여 for 26 weeks, and are not displaced by robots, since the diffusion family is the AI-cognitive wave. The Korean anchors are the level and the lag: 9.2% of Korean firms used AI in 2024, up from 1.4% in 2017 (Statistics Korea 기업활동조사); 31% of SMEs against more than half in Germany (OECD 2025); and KDI's finding that 38.8% of jobs were more than 70% automatable at 2023 technology, so that realized automation is gated by adoption rather than by what is technically possible. 

Which inputs are Korean, and which are not. Wherever a Korean institution publishes the figure, the model uses it: NABO's fund projections and receipts, Statistics Korea's population projection, the Bank of Korea's exposure classification, MOEL's re-employment and benefit statistics, and the statutory contribution and tax schedules. No Korean institution publishes a path for how much exposed work will actually be automated, a robot-exposure measure by occupation, or the wage cut on re-employment after displacement. For those the model borrows, and says so: the starting level and end point of the adoption paths from the US adoption evidence and the Acemoglu– Svanberg bound; robot exposure from Webb's US patent measure mapped onto Korean occupation groups; the 13% wage cut from Farber's US displacement studies; and the behavioural conventions the US model shares, such as how firms dispose of saved wages. Each is a placeholder for a Korean figure the day one is published, and each is listed with its source in the scenario's own provenance. 

How a displaced worker is priced. For each of the 209 cells, the model answers one question: if this worker loses this wage, what happens to revenue at every level of government? There are five channels. 

Income tax: the national brackets are re-evaluated at the worker's wage, plus the 10% local income surtax. This does not use average rates or elasticities: it is how this worker's taxes would actually change. 

Social insurance: the five schemes at their statutory rates, taking into account the pension contribution cap, which sits in the wage range where automation of clerical work starts. Transfers: 구직급여 for the benefit window (60% of the daily wage between the statutory floor and cap, 120–270 days by age and insured years, tax-exempt), then the loss of 근로장려금 once earned income is gone; 기초연금 for those who leave the labour force at 65 and over. Basic livelihood support is not modeled. 

Corporate recapture: the wage that a firm stops paying becomes taxable profit, taxed at the 24.2% effective corporate rate.

Consumption: lost VAT from the drop in spending. 

The model integrates over the wage distribution inside each cell rather than using average workers, because the contribution cap and the benefit floor and ceiling are thresholds that an average wage would miss. 

The dynamic model. The simulation then runs all of the cells forward one year at a time: 

1. An adoption ceiling sets how much feasible work is automated that year, cumulatively. Physical automation arrives later than cognitive automation, on the scenario's own ramp. 2. Firms use the saved wage bill in one of four ways: compute spending, retained profit, price cuts for consumers, and raises for surviving workers. Every lost won goes to exactly one of these four. 3. Survivor raises are funded from the saved wage bill, never out of thin air, and are taxed through the same brackets and contribution schedules. 

4. The macro block applies the price cuts and productivity gains to real output. Deflation does not change nominal tax arithmetic. 

5. The general account books the net change and borrows to cover any deficit. The earmarked funds book their own losses against NABO's published reserve paths, which is where the depletion dates come from. 

6. Local government books its own gap through the local income surtax. There is no state tier and no austerity closure: the gap is reported, never closed. 

7. Lost take-home pay leads to demand destruction and induced layoffs the following year. The loop is built to converge rather than spiral, and this is checked on every run. 

Trustworthiness. Eight families of accounting identities are re-verified on every period of every run. Every worker is in exactly one of seven states that sum to baseline employment; every saved won goes to exactly one destination; the national ledger reconciles line by line; and the Korean variant asserts that no austerity ever closes the gap. If a fiscal flow were not conserved, the run would stop. With every behavioural lever switched off, the full system reproduces the static calculation exactly. There are 540 tests, 149 of them specific to Korea, covering the identities, the channel arithmetic, the published fund paths transcribed independently, and behavioural pins from the displacement literature. Every result is a change against the NABO and KOSIS baselines: this is not a forecast of the whole economy, only what automation changes. The sensitivity chart shows which assumptions move each headline. 

Model scope and limitations 

Saved compensation by firms is fully converted to taxable surplus, which in reality would lead to a larger fiscal gap. 

The Bank of Korea does not respond to deflating prices, so it would not be following its mandate; the model overstates the fiscal gap in crisis scenarios. 

Within-job augmentation is not represented, which one can expect to lead to more productive and profitable firms, which could result in higher income and corporate taxes, overstating the gap. 

Compute and robot taxes do not affect firm behaviour and prices, which would slow down adoption. Robot exposure is anchored to the current patent stock, which understates potential robotics adoption rates, and is based on US exposure numbers mapped onto Korean occupation groups.

Cognitive exposure is the Bank of Korea's displacement-prone classification, which puts nearly all of the exposure on clerical work; the US measure spreads it across professionals, managers, and sales, and the two are compared in the methodology. 

This model only uses wage employees, which is about 76% of Korean employment, and its survey frame covers 12.4 million of roughly 22 million wage workers, over-represents larger firms, and excludes public administration. 

Nine occupation groups by 24 wage brackets is deliberately coarse and transparently so. 

Erosion counts revenue plus EI benefit outlays. NHI and NPS outlay responses and forgone investment income are not counted, and both omissions understate the damage. 

Exposure is a reading of the Bank of Korea's published chart, accepted because it reconciles with the printed aggregates. The ±0.5pp reading error is carried into the band. 

AI-cognitive exposure is the Bank of Korea's displacement-prone read; robot exposure is Webb (2020) mapped from US occupations, since no Korean measure exists. Demography follows the published KOSIS medium scenario. The fiscal gap is reported and never closed.