Korean firms saw rising overheads after the staged 52‑hour workweek, a pattern the authors interpret as the early-stage ‘overhead‑pressure’ of AI-driven time compression; firms that switch to output-based talent accounting are forecast to achieve roughly 1.5–2.0 percentage points higher TFP growth by 2032.

Main Finding

Shin & Kang (2026) develop a theoretical and empirical forecasting framework for a regime shift from time-based talent accounting to output-based talent ROI under AI augmentation. The paper’s core result (Theorem 3, "ROI Inversion at τ") predicts a single, derivable utilization threshold τ at which output-based accounting strictly outperforms time-based accounting for firm-level TFP. Empirically, a Korea DART panel (365 listed firms, 2018–2024) shows a rising SG&A-to-revenue ratio during staged implementation of a 52‑hour workweek and concurrent diffusion of generative AI — interpreted as the first documented signature of the pre-τ “overhead-pressure” regime. The authors forecast that firms adopting output-based talent ROI will exceed time-based peers in firm-level TFP growth by roughly 1.5–2.0 percentage points by 2032.

Key Points

• Conceptual shift: Talent ROI accounting is treated as a sociotechnical regime; AI makes an hour of work a heterogeneous mix of human and computational contribution, breaking the time-based unit-of-account assumption.
• Five theorems (hierarchy):
• Overhead Decomposition — the seven standard overhead components (wage, insurance, space, management, training, communication, motivation) become non-additive under augmentation; cross-partials matter.
• Slack‑Augmentation Synergy — augmentation-saved time splits into four mutually exclusive pathways: work intensification, hidden leisure, overemployment, and creative‑slack reinvestment; only creative slack yields innovative ROI, and its share grows with output-orientation and autonomy.
• ROI Inversion at τ — as AI utilization rises, time-based ROI falls and output-based ROI rises; they cross at τ (a function of output-orientation, autonomy, convergence capacity C). Below τ time-based accounting can remain optimal; above τ it drags TFP.
• Innovative ROI Premium — slack policies (e.g., Google 80/20) are amplified by a factor k > 1 when convergence capacity is high; amplification rises with the augmentable cognitive share.
• Information Asymmetry in human‑AI dyads — attribution uncertainty adds a third measurement difficulty to multitask principal-agent models, increasing agency costs under time-based evaluation and reducing them under output-based evaluation.
• Theory backbone: augmented human capital Ĥ = H · [1 + φ(A, C)], where A = AI utilization intensity, C = convergence capacity; φ captures interaction premium.
• Forecasts & falsifiability: central, testable forecast that output-based adopters will outperform by 1.5–2.0 ppt TFP growth by 2032; each theorem has independent falsifiability conditions.

Data & Methods

• Primary dataset: Korea Financial Supervisory Service DART disclosures. Panel covers KOSPI + KOSDAQ Top 500 sample filtered to 365 firms, 2,281 firm‑year observations (2018–2024). A 2015–2017 backward extension (224 firms, 601 obs) supplies pre-treatment checks.
• Empirical anchor: staged implementation of Korea’s 52‑hour workweek is used as an externally timed institutional shock to induce labor‑time compression; mapping to firms uses a revenue‑percentile cohort proxy (authors treat this mapping as an identifying assumption and interpret results directionally).
• Outcome: SG&A (selling, general, and administrative) expense as a share of revenue — used as an early-warning indicator of rising internal talent overhead under time-based accounting.
• Estimation approaches:
  • Two‑way fixed effects: +1.56 percentage points SG&A/revenue (clustered s.e. 0.79, p = 0.049).
  • Pooled event‑study: ramps to +4.21 pp at t = +3 (p = 0.001).
  • Callaway–Sant’Anna doubly‑robust staggered DiD: +4.51 pp at t = +4.
  • Pre-trend checks using 2015–2017 extension argue against simple pre-existing upward trends.
• Identification caveats: authors emphasize these estimates are not a point causal magnitude of the 52‑hour law or a direct estimate of τ*. The revenue‑percentile cohort is a proxy, and AI adoption is modeled as a concurrent diffusion process; results are read as evidence of a pre‑τ overhead‑pressure regime rather than definitive causal inference.

Implications for AI Economics

• Measurement & accounting:
  • The standard time-based overhead accounting is increasingly misspecified under AI augmentation. Economists and managers should treat firm talent ROI as regime-dependent and monitor indicators (e.g., SG&A/revenue) as early-warning signals.
  • Researchers should develop microdata measures that separate human vs. AI contribution (attribution metrics) and quantify φ(A, C) and convergence capacity C.
• Managerial practice:
  • Firms should track three levers before switching evaluation regimes: AI utilization intensity (A), convergence capacity (C), and institutional features (output‑orientation, employee autonomy). The τ* threshold is endogenous to these variables.
  • Increasing convergence capacity and autonomy raises the probability that creative‑slack reinvestment will generate innovative ROI; thus, investments in integration processes, knowledge codification, and decision rights can raise k and lower the effective τ*.
  • Firms can experiment with output-based pay and evaluation in controlled pilots; monitor TFP growth and SG&A dynamics to assess when a regime inversion is welfare‑enhancing.
• Policy:
  • Labor regulation (e.g., compressed workweeks) interacts with AI diffusion to produce overhead pressure in the pre‑τ phase. Policymakers should expect non-linear transitional outcomes and consider supporting firm investments in convergence capacity and attribution infrastructure.
  • Anticipate distributional consequences: as accounting regimes shift, compensation structures, bargaining positions, and employment contracts may need redesign.
• Research agenda:
  • Estimate τ* empirically across sectors and countries (requires richer microdata on AI tool adoption, task composition, output measures, and internal accounting categories).
  • Test the four augmentation‑saved‑time pathways directly (time‑use and task‑level data) and measure the amplification factor k under varying C.
  • Extend cross-country firm‑level analyses (planned Denmark/Japan comparisons suggested by authors) to understand how institutional context (flexicurity, governance, management norms) shifts the τ* mapping.
• Macroeconomic outlook:
  • If the framework is borne out, a structural shift to output-based talent ROI could raise aggregate TFP growth as output-orientated, high‑C firms capture more innovative surplus; however, transition dynamics may produce temporary overhead pressures, reallocations, and adjustment costs.

Limitations noted by authors: Korean case is a critical/early‑warning prototype, not a universal estimate of τ; empirical identification uses a cohort proxy and concurrent AI diffusion, so causal magnitudes are directional. Further microdata and international replication are needed to quantify τ and to operationalize managerial switching rules.