AI deployments depress operating margins initially for KOSDAQ firms — consistent with short‑term transition costs — while investors value AI as a growth option in ICT firms, which show significant increases in market valuation.
The Dynamic Causal Effects of Corporate AI Adoption on Profitability and Market Value : An Empirical Analysis of KOSDAQ Panel Data
Using DART-based adoption timing and TWFE/PSM on KOSDAQ firms (2018–2025), the study finds AI adoption lowers operating profit margins in the short run (consistent with a J‑curve), leaves ROA unchanged on average, and raises Tobin's Q selectively in ICT firms.
This study analyzes the causal effects of AI adoption on the corporate performance of KOSDAQ-listed companies from 2018 to 2025. The timing of adoption was identified through a multi-step, contextually validated text analysis of DART business reports, and endogeneity was controlled using a two-way fixed effects (TWFE) model and Propensity Score Matching (PSM). The analysis results reveal that AI adoption had a significantly negative (−) impact on the operating profit margin (OPM), consistent with short-term “J-curve” transition costs including process redesign and capability buildup. No statistically significant change was observed in ROA, reflecting the offsetting dynamics of income compression and asset expansion during the early adoption phase. While the average effect on market value (Tobin's Q) was not significant across all firms, a heterogeneous effect was observed in the ICT industry, where Tobin's Q significantly increased following AI adoption. This result indicates that the capital market evaluates AI investment as a future “growth option” selectively within industrial contexts characterized by strong data infrastructure, digital workforce readiness, and technological absorptive capacity. By demonstrating that the value creation process of AI adoption varies according to industry-specific complementary asset structures and ecosystem conditions, this study provides significant implications for digital transformation strategies, investment decision-making, and AI diffusion policy.
Summary
Main Finding
AI adoption by KOSDAQ-listed firms (2018–2025) produced a significant short-term decline in operating profit margin (OPM), consistent with a “J‑curve” of upfront transition costs. Return on assets (ROA) showed no statistically significant change, while market valuation (Tobin’s Q) did not move on average — except in the ICT industry, where Tobin’s Q rose significantly after adoption, indicating that markets selectively price AI as a future growth option when complementary assets and ecosystem readiness exist.
Key Points
- Short-run operating performance: AI adoption → significantly negative effect on OPM, interpreted as transition costs (process redesign, training, capability buildup).
- Accounting profitability: ROA shows no statistically significant change, implying income compression from costs was offset by asset expansion in the early adoption phase.
- Market valuation: No average effect on Tobin’s Q across all firms, but a positive, significant effect within the ICT sector.
- Heterogeneity matters: Industry context (data infrastructure, digital workforce readiness, absorptive capacity) conditions whether capital markets reward AI investment.
- Methodological rigor: Adoption timing identified via validated text analysis of corporate filings; endogeneity addressed via TWFE and PSM.
Data & Methods
- Sample: KOSDAQ-listed firms, 2018–2025.
- Adoption identification: Multi-step, contextually validated text analysis of DART business reports to determine timing of AI adoption announcements/implementation.
- Econometric strategy:
- Two-way fixed effects (TWFE) models to estimate causal effects while controlling for firm and year fixed effects and time-varying shocks common to firms.
- Propensity Score Matching (PSM) used to create comparable treated and control groups and reduce selection bias from non-random adoption.
- Heterogeneity analysis: Industry-stratified estimation, focusing on ICT vs non-ICT firms to reveal differential valuation and performance responses.
Implications for AI Economics
- Dynamics and measurement:
- Evidence supports a dynamic “J‑curve” model of AI adoption: initial negative operating returns followed by expected longer-run gains. Empirical studies should model time-varying adoption effects rather than only average treatment effects.
- Accurate, validated measures of adoption timing (text analysis of filings) are crucial for causal inference in diffusion research.
- Complementarities and heterogeneity:
- Returns to AI depend on complementary assets (data infrastructure, skilled labor, absorptive capacity). Models of AI-induced productivity must incorporate industry- and firm-level complementarities.
- Capital markets rationally discriminate across industries: AI investment is priced as an option in contexts with strong complementary assets.
- Policy and firm strategy:
- Policymakers should target support (training, data infrastructure, financing) to sectors and firms lacking complementary assets to avoid widening gaps in diffusion and value capture.
- Firms should anticipate short-term profit compression; plan staged investment, capability building, and metrics that capture longer-run option value.
- Future research directions:
- Study longer-run post-adoption horizons to confirm recovery and net gains beyond the short-term J‑curve.
- Quantify the role of specific complementary assets (data, talent, workflow redesign) and the threshold levels required to convert AI investments into value.
- Extend the identification approach to other markets/countries to assess generalizability of heterogeneous valuation patterns.
Assessment
Paper Typequasi_experimental
Evidence Strengthmedium — Strengths: validated textual identification of adoption timing, panel fixed effects that absorb firm fixed traits and common time trends, and use of PSM to improve covariate balance. Limitations: potential remaining confounding from time-varying unobservables (managerial intent, concurrent investments, demand shocks), possible measurement error in text-based adoption coding, and use of conventional TWFE in a staggered-adoption setting (2018–2025) which can produce biased average treatment estimates when treatment effects are heterogeneous; PSM does not resolve unobserved confounding.
Methods Rigormedium — Appropriate and standard tools (TWFE, firm and year fixed effects, PSM) were applied and adoption timing was validated contextually, which shows care in design; however, the analysis would be more rigorous if it addressed recent methodological concerns for staggered DiD (e.g., using event-study estimators robust to heterogeneous effects such as Callaway & Sant'Anna or Sun & Abraham), reported pre-trends/event-study evidence, corrected for measurement error in adoption coding, and provided sensitivity analyses for unobserved confounding.
SamplePanel of KOSDAQ-listed firms in South Korea observed 2018–2025, with AI adoption dates inferred from firm DART business reports via text analysis; firm-level outcomes include operating profit margin (OPM), return on assets (ROA), and market value (Tobin's Q); sample spans multiple industries with a highlighted subsample (ICT industry) where heterogeneous effects were explored. (Exact sample size not provided.)
Themesproductivity adoption innovation
IdentificationTiming of AI adoption is identified via a multi-step, contextually validated text analysis of DART (Korean corporate filing) business reports; causal effects are estimated using panel two-way fixed effects (firm and year) to control for time-invariant firm heterogeneity and common time shocks, and robustness is sought via propensity score matching (PSM) to balance observable covariates between adopters and non-adopters.
GeneralizabilitySingle-country (South Korea) and single-exchange (KOSDAQ) sample — findings may not generalize to other countries or larger-cap exchanges., KOSDAQ firms skew toward smaller, growth-oriented companies; results may differ for large incumbents or SMEs in other markets., Early-adoption window (2018–2025) captures short- to medium-run effects; longer-run effects of AI adoption remain unobserved., Adoption measure relies on business-report text disclosures and may miss informal/partial adoption or misclassify timing., Industry heterogeneity (ICT vs non-ICT) limits average effect interpretation; complementary assets and ecosystem conditions vary across contexts., Macroeconomic events during 2018–2025 (e.g., COVID-19, supply-chain shocks) could interact with AI adoption effects.
Claims (9)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| AI adoption had a significantly negative impact on the operating profit margin (OPM). Firm Productivity | negative | high | operating profit margin (OPM) |
0.48
|
| No statistically significant change was observed in return on assets (ROA) following AI adoption. Firm Productivity | null_result | high | return on assets (ROA) |
0.48
|
| The average effect of AI adoption on market value (Tobin's Q) was not statistically significant across all firms. Firm Revenue | null_result | high | Tobin's Q (market value) |
0.48
|
| In the ICT industry, Tobin's Q significantly increased following AI adoption (heterogeneous positive effect). Firm Revenue | positive | high | Tobin's Q (market value) in ICT-industry firms |
0.48
|
| The timing of AI adoption was identified through a multi-step, contextually validated text analysis of DART business reports. Adoption Rate | null_result | high | AI adoption timing identification |
0.8
|
| Endogeneity in estimating AI's effects was controlled using a two-way fixed effects (TWFE) model and Propensity Score Matching (PSM). Other | null_result | high | causal identification / endogeneity control |
0.8
|
| The observed negative OPM effect is consistent with short-term 'J-curve' transition costs (process redesign and capability buildup) during early AI adoption. Firm Productivity | negative | high | operating profit margin dynamics / transition costs interpretation |
0.08
|
| The capital market evaluates AI investment as a future 'growth option' selectively in industries with strong data infrastructure, digital workforce readiness, and absorptive capacity. Market Structure | positive | medium | market valuation response to AI investment (interpreted as growth-option pricing) |
0.05
|
| The value-creation process of AI adoption varies according to industry-specific complementary asset structures and ecosystem conditions, with implications for digital transformation strategies, investment decisions, and AI diffusion policy. Innovation Output | mixed | medium | variation in value-creation from AI adoption across industries |
0.05
|