







































X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 
01-11 

 

1 

 

 

 

Article 

The impact of AI-driven industrial upgrading on 

economic development 
Xiaofei Hao, Aotip Ratniyom*, Sivalap Sukpaiboonwat  

Faculty of Economics, Srinakharinwirot University, 114 Sukhumvit 23, Bangkok 10110, Thailand 

A R T I C L E   I N F O 
 

Article history: 
Received 06 May 2025  
Received in revised form 
13 June 2025 
Accepted 24 June 2025 
 
Keywords:  
Artificial intelligence, Industrial upgrading, 
Economic development, Threshold effects, 
Working-age population, Digital transformation 
 
*Corresponding author 
Email address: 
aotip.ratniyom@gmail.com 
 
 
DOI: 10.55670/fpll.futech.4.4.1 

A B S T R A C T 
 

The paper clarifies the interdependencies between AI adoption, industry 
upgrading, and economic development in the context of global digital 
transformation. With mixed-methods integrating econometrics and case 
studies, we test models formalizing mediating and threshold effects in AI-
industry-economy relations. Our approach leverages a novel AI penetration 
score by industries alongside economic indicators and measures of industry 
sophistication. The results indicate that AI uptake mediates the pass-through of 
industry structure change to economic performance, with contribution levels 
increasing above certain thresholds. Evidence suggests that the association 
between the working-age population and economic growth varies by 
alternative industry upgrading rankings, with technologically sophisticated 
structures making better use of demographic opportunities. Threshold analysis 
identifies regimes where AI substitutes for traditional economic relations, 
revealing policy intervention points. These findings contribute to growth 
theory innovation by measuring AI's catalytic economic function and offer 
methodological innovation in the analysis of technological contributions. 
Strategic AI development agendas, human capital policies, and coordination 
mechanisms are among the key implications required to achieve inclusive 
growth in the digital economy. This study closes knowledge gaps on how 
demographic and technological drivers interact through industry structures to 
determine economic trajectories. Empirical results show that AI adoption 
mediates 52.8% of manufacturing sophistication's impact on GDP growth, 
threshold effects emerge at an AI adoption index of 0.43-0.45, where economic 
impacts increase threefold, and the working-age population's growth effect 
varies from 0.072 below the threshold to 0.411 above the threshold in the 
highest industrial upgrading quartile. 

1. Introduction 

The global industrial system is undergoing rapid digital 
transformation, characterized by technological revolution 
and shifting economic paradigms. AI stands as a central 
transformative force, fundamentally reshaping economic 
structures [1]. Technologies like machine learning, deep 
learning, and computer vision are being deployed across 
industries, automating complex work, enhancing decisions, 
and creating new value sources [2]. This adoption differs from 
previous technological revolutions in its economic impact 
potential. AI's implications extend beyond productivity 
increases to industry structure, labor markets, and 
competitive positioning, with consequences varying across 
firms, industries, and locations [3]. The technological change-
economic growth relationship is nonlinear, challenging 
traditional growth theory. Figure 1 illustrates fundamental 
differences between traditional and AI-enabled industrial 
systems. Traditional systems show limited productivity, 

slower innovation cycles, and lower value-chain positioning. 
AI-enabled systems demonstrate enhanced productivity. 
Modern industry balances challenges with high-tech 
solutions amid competitive pressures. Industrial structures 
both drive and are driven by technology adoption. 
Brynjolfsson and McAfee [4] refer to this as "the second 
machine age," where intelligent technologies augment human 
capacities while challenging economic structures. This 
research investigates how AI adoption adds value in 
industries and how industry conditions shape adoption. 
Evidence suggests AI's economic impact becomes 
pronounced at critical adoption levels [5]. Using threshold 
regression, we identify these critical points and characterize 
regime-specific relationships. We also explore how the effects 
of the working-age population on growth vary across 
industrial upgrading levels, as demographic impacts depend 
on industrial technological sophistication [6]. 

 

May 2025| Volume 01 | Issue 01 | Pages 01- 03 

ISSN 2832-0379 

Open Access Journal 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.0.0.0 

 

 

Future Technology  

 

November 2025| Volume 04 | Issue 04 | Pages 01-11 

Journal homepage: https://fupubco.com/futech 

 

https://doi.org/10.55670/fpll.futech.4.4.1 

 

 

Future Technology 

Open Access Journal 

ISSN 2832-0379 

mailto:aotip.ratniyom@gmail.com
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.0.0.0
https://fupubco.com/futech
https://doi.org/10.55670/fpll.futech.4.4.1


X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

2 

 

 
Figure 1. Comparison of traditional and AI-enabled industrial 
systems 

This study pursues three specific objectives. First, we 
aim to empirically quantify how AI adoption mediates the 
relationship between industrial structure and economic 
development outcomes, moving beyond simple correlation to 
identify transmission mechanisms. Second, we seek to 
identify and measure threshold effects in AI adoption that 
fundamentally alter the nature of industrial-economic 
relationships. Third, we examine how industrial upgrading 
levels moderate the impact of demographic factors on 
economic growth, particularly the working-age population 
dividend. These objectives address critical gaps in 
understanding the complex interactions between 
technological adoption, industrial transformation, and 
economic development in the digital era. Our work 
contributes to understanding technological change by 
advancing innovation growth models that integrate structural 
economics with digital economy perspectives. While existing 
theories acknowledge technology's role, they often treat 
advancement as aggregate without addressing sector-specific 
patterns or threshold effects [7]. We develop nuanced 
conceptualizations of the influence of AI adoption on growth 
through industrial structural change [8]. We quantify AI's 
economic effects using industry adoption indexes and inform 
policy through econometric findings. The research examines 
the relationship between industry transformation, AI 
adoption, and economic performance, investigating how 
industry characteristics influence AI adoption and moderate 
the structure-performance relationships. Methodologically, 
we combine econometrics with case studies to enhance 
validity. 

2. Literature review 

2.1 Theoretical Foundations 
The study of AI-driven industrial upgrading draws from 

several theoretical traditions. Innovation and Growth Theory 
incorporates technological advancement as a central 
economic driver, evolving from R&D-focused models to 
frameworks capturing digital technologies' characteristics. 
Romer's work established how knowledge production 
generates increasing returns, while Schumpeterian 
perspectives explain creative destruction [9].These 
frameworks help understand how AI systems alter 
productivity frontiers across sectors. Standard growth 
models inadequately capture the discontinuous nature of 
general-purpose technologies like AI, requiring extensions for 
threshold effects and nonlinear adoption impacts. Structural 
Transformation Theory examines economic sector changes. 

Digitalization dissolves traditional sector boundaries, with 
Rodrik noting that technology enables some economies to 
skip conventional industrialization [10]. AI either propels or 
hinders structural transformation depending on context.  
Industrial Organization Theory examines the impact of AI on 
industry structure and competition. AI's economics require 
adapting standard market models to accommodate network 
effects and increasing returns. AI growth often results in 
"winner-takes-most" markets, where early movers have 
advantages and transaction costs are altered in ways that 
reshape industries. Technology Diffusion Models explain AI 
propagation through economies. AI's reliance on 
organizational competencies, data, and infrastructure 
complicates adoption. Evidence shows AI diffusion follows S-
shaped trajectories with sector variations due to 
implementation barriers [11]. Threshold effects accelerate 
adoption when ecosystems reach critical mass, introducing 
time lags between implementation and productivity 
enhancement. 

2.2 Industrial structure and economic development 
These theoretical frameworks provide the foundation for 

examining empirical patterns of industrial development. The 
global business has experienced deep transformation in the 
trends of specialization, technological intensity, and value 
chain engagement. Classical linear concepts of evolution from 
agriculture to industry to services are challenged by 
complicated growth trajectories. "Premature 
deindustrialization" in the developing world questions 
successful development policy in the information age [10]. 
Mature economies experience post-industrial transition with 
classical manufacturing decline and knowledge-intensive 
services expansion. Cross-country experiences observe 
heterogeneous industrial upgrading impacts. The East Asian 
countries showed dynamic upgrading trajectories through 
incremental capability building, with South Korea 
demonstrating how policy coordination drives shifts towards 
knowledge-intensive production from labor-intensive 
production [12]. Latin American and African economies 
become trapped in lower value-added activities despite 
reforms. Measurement of structural transformation has 
evolved from simple indicators to advanced frameworks with 
economic complexity indices and input-output analyses. The 
economic growth-working-age population relation is 
heterogeneous by industrial structures. Population dividend 
theories contend that higher working-age population shares 
yield growth dividends, yet evidence suggests such dividends 
are industry sectoral structure contingent, with technology-
intensive structures amplifying demographic dividends and 
labor-intensive industries perhaps facing reduced returns 
[13]. AI technologies add complexity to these economic-
demographic relations. 

2.3 AI and industrial structure upgrading 
Within this broader context of structural transformation, 

AI emerges as a particularly transformative force. AI is both a 
driver and a facilitator of structural transformation in various 
sectors. Paradigms in thinking have changed from 
reductionist automation to a focus on complementarity 
between intelligent technologies and organizational 
capabilities. The "innovation-productivity-structure" model 
highlights how AI induces sequential transformations: 
innovations lead to productivity growth, which in turn 
reconstitutes industrial structures through reconfigured 
competitive strengths. Evidence-based studies affirm AI 
uptake is necessitated by industry contexts, organizational 
capacities, and institutional environments [14]. AI application 

Limited Productivity Growth

Incremental improvements constrained by

traditional optimization approaches

Slower Innovation Cycles

Extended development timelines with

higher resource requirements

Lower Value-Chain Positioning

Difficulty ascending to higher-value activities

with limited technological capabilities

Enhanced Productivity

37% average productivity increase in advanced

implementations (threshold > 0.43)

Accelerated Innovation

Reduced product development cycles by 41%

with integrated AI decision support systems

Higher Value-Chain Positioning

Movement toward higher complexity products

with 52.8% mediation via AI capabilities



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

3 

 

varies between manufacturing and services. In 
manufacturing, AI facilitates predictive maintenance, quality 
control automation, and design optimization. Industry 4.0 
combines AI with IoT infrastructure to develop cyber-
physical systems, revolutionizing production economics. In 
services, AI automates customer engagement, facilitates 
personalization, and builds decision support systems. 
Successful adoption is normally subject to the evolutionary 
process rather than revolutionary change. Productivity gains 
from smart automation motivate AI adoption, with properly 
leveraged systems demonstrating improvements in 
performance through enhanced workability, reduced faults, 
and optimized resource deployment. These gains will 
typically require investment in data infrastructure for 
information, capabilities for individuals, and firm redesign, 
resulting in implementation lags between adoption and 
perceivable effect [15]. Value chain re-engineering is likely to 
be the greatest organizational impact of AI adoption, as smart 
systems facilitate the fundamental redesign of activities and 
relationships in industrial networks. 

2.4 Research gap and hypotheses 
Despite extensive research on these topics, critical gaps 

remain that motivate our study. Despite extensive literature 
on industrial development and technological change, 
significant gaps persist in understanding their intersection, 
particularly regarding AI's role in industrial upgrading. 
Industrial change and AI adoption literature remain 
separated, with industrial economics emphasizing structural 
transformation without technological specificity, while AI 
research neglects broader structural implications. Addressing 
this requires theoretical frameworks that model bidirectional 
relationships between technological capabilities and 
industrial structures [16]. Mediating mechanisms through 
which AI influences economic outcomes via industrial 
transformation represent another critical gap. While evidence 
confirms AI's economic impact, specific pathways remain 
inadequately theorized. Potential mediating mechanisms 
include productivity enhancements, product innovation, 
market expansion, resource allocation efficiency, and inter-
industry spillover effects. Threshold effects in AI-induced 
industrial change are important yet understudied. Evidence 
suggests AI exhibits discontinuous effects after adoption 
reaches critical thresholds, but systematic examination is 
scarce. Detection of these thresholds demands specialized 
econometric techniques [17]. Based on these gaps, we offer 
three hypotheses: First, AI adoption mediates the relationship 
between industrial structure and economic growth, with 
context-contingent effects. Second, AI adoption exerts 
threshold effects, amplifying economic effects after critical 
adoption thresholds are met. Third, industrial upgrading 
moderates the impacts of demographic change on economic 
growth, with advanced structures enhancing the beneficial 
demographic influences. Verification requires rigorous case 
studies and econometric exercises to identify economic-
industrial-AI relationship trends. 

3. Research methodology 

3.1 Research design 
This research employs a comprehensive analytical 

framework connecting industrial structure variables, AI 
adoption metrics, and economic outcomes through a system 
of interconnected relationships. The core analytical model 
posits that economic development outcomes are influenced 
by both direct effects of industrial structure and indirect 
effects mediated through AI adoption, with potential 

threshold effects and demographic interactions. This 
relationship can be expressed through the following 
econometric specification: 

𝑌𝑖𝑡 = 𝛼 + 𝛽1𝐼𝑁𝐷𝑖𝑡 + 𝛽2𝐴𝐼𝑖𝑡 + 𝛽3(𝐼𝑁𝐷𝑖𝑡 × 𝐴𝐼𝑖𝑡) + 𝛽4𝑊𝐴𝑃𝑖𝑡 +
𝛽5(𝐼𝑁𝐷𝑖𝑡 × 𝑊𝐴𝑃𝑖𝑡) + 𝛾1𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 휀𝑖𝑡                              (1) 

Where 𝑌𝑖𝑡  represents economic development indicators 
for region i at time t; 𝐼𝑁𝐷𝑖𝑡  captures industrial structure 
characteristics; 𝐴𝐼𝑖𝑡  measures AI adoption intensity; 𝑊𝐴𝑃𝑖𝑡 
represents the working-age population proportion; 𝑋𝑖𝑡 
includes control variables; 𝜇𝑖  and 𝜆𝑡  represent region and 
time fixed effects; and 휀𝑖𝑡 is the error term. To test threshold 
effects, we employ the following threshold regression model: 

𝑌𝑖𝑡 =

{
𝛼1 + 𝛽11𝐼𝑁𝐷𝑖𝑡 + 𝛽12𝐴𝐼𝑖𝑡 + 𝛾1𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 휀𝑖𝑡 ,if 𝐴𝐼𝑖𝑡 ≤ 휃
𝛼2 + 𝛽21𝐼𝑁𝐷𝑖𝑡 + 𝛽22𝐴𝐼𝑖𝑡 + 𝛾2𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 휀𝑖𝑡,if 𝐴𝐼𝑖𝑡 > 휃

     

             (2) 
Where 휃  Tampilkan Gambar represents the threshold 

value of AI adoption that potentially changes the relationship 
between variables. 

This research employs a mixed-methods approach 
integrating econometrics with case studies. Econometrics 
provides statistical generalizability while case studies offer 
contextual understanding of mechanisms. The sequential 
design begins with quantitative analysis to identify patterns, 
followed by purposive case selection. Integration occurs 
during case selection, interview protocol development, and 
final interpretation, where findings are synthesized to explain 
observed phenomena. To address causality identification 
challenges inherent in observing the complex relationships 
between industrial structure, AI adoption, and economic 
outcomes, we employ multiple empirical strategies. First, we 
exploit temporal variation through lagged independent 
variables, where industrial structure at t-1 affects AI adoption 
at t, which subsequently influences economic outcomes at t+1. 
This temporal sequencing helps mitigate simultaneity bias. 
Second, we utilize region-specific heterogeneity in AI policy 
implementation timing as a quasi-experimental setting, 
where differential policy rollouts create exogenous variation 
in AI adoption rates. Third, our panel fixed effects 
specifications control for time-invariant unobserved 
heterogeneity that might confound the relationships. These 
identification strategies, combined with robustness checks 
using alternative specifications and instrumental variables, 
strengthen our causal inference beyond mere correlational 
analysis. 

3.2 Variable selection and measurement 
This research employs variables capturing relationships 

between industrial structure, AI adoption, and economic 
development, balancing theoretical relevance with data 
availability. The dependent variables comprise two 
categories of outcomes: economic growth indicators and 
industrial sophistication indices. Economic growth is 
primarily measured through GDP per capita growth rate 
(GDPGit) and total factor productivity growth (TFPGit), 
calculated as: 

𝑇𝐹𝑃𝐺𝑖𝑡 =
𝑌𝑖𝑡

𝐾𝑖𝑡
𝛼⋅𝐿𝑖𝑡

1−𝛼 −
𝑌𝑖𝑡−1

𝐾𝑖𝑡−1
𝛼 ⋅𝐿𝑖𝑡−1

1−𝛼                                              (3) 

Where Yit represents output, Kit capital stock, and Lit 

labor input for region i at time t. Industrial sophistication is 
measured through the Economic Complexity Index (ECIit) and 
the Industrial Upgrading Index (IUIit), which captures 
movement toward higher value-added activities. 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

4 

 

Core explanatory variables include industrial structure 
metrics, measures of AI adoption, and working-age 
population ratios. Industrial structure is quantified through 
the manufacturing value-added share (MVAit), high-
technology exports percentage (HTEit), and an industrial 
diversification index (IDIit) calculated as: 

𝐼𝐷𝐼𝑖𝑡 = 1 − ∑ 𝑠𝑖𝑗𝑡
2

𝑛

𝑗=1
                                                                       (4) 

Where Sijt represents the share of industry j in region i's 
industrial output at time t. The working-age population is 
measured as the ratio of the population aged 15-64 to the 
total population (WAPit). 

Control variables encompass traditional economic 
factors and institutional quality measures essential for 
isolating the effects of our core variables. These include 
human capital measured through average years of schooling 
(HCit), investment ratio calculated as gross fixed capital 
formation to GDP (IRit), trade openness represented by the 
sum of exports and imports divided by GDP (TOit), and 
institutional quality indices capturing regulatory efficiency 
and rule of law (IQit). The AI penetration index represents a 
methodological innovation, constructed as a composite 
measure capturing multiple dimensions of AI implementation 
across industrial sectors. The index encompasses four 
theoretically grounded dimensions that comprehensively 
capture AI adoption intensity: 
AI Innovation Capacity (measured through AI patent 
applications per million population): This dimension 
captures the technological frontier and innovation potential, 
indicating regions' ability to develop novel AI applications. 
Patent data is sourced from WIPO's Global Innovation Index, 
focusing on IPC codes G06N (computing arrangements based 
on specific computational models) and G06F (electric digital 
data processing with AI-specific subclasses). 
AI Human Capital (measured through AI talent concentration 
and skills prevalence): Reflecting the critical role of 
specialized knowledge in AI implementation, this dimension 
uses LinkedIn Talent Insights data on AI-skilled professionals 
as a percentage of the workforce, supplemented by computer 
science and data science graduate numbers from national 
education statistics. 
AI Investment Intensity (measured through venture capital 
and corporate AI investments as a percentage of GDP): This 
dimension captures financial commitment to AI development, 
aggregating data from Crunchbase, PitchBook, and national 
innovation surveys, including both private venture funding 
and corporate R&D allocated to AI initiatives. 
AI Research Output (measured through AI-related scientific 
publications per capita): Indicating knowledge generation 
and absorption capacity, this uses Scopus and Web of Science 
data for publications in AI-related fields, weighted by citation 
impact using field-normalized metrics. 

𝐴𝐼𝑖𝑡 = ∑ 𝑤𝑘
4
𝑘=1 ⋅

𝑋𝑘𝑖𝑡−𝑚𝑖𝑛(𝑋𝑘𝑖𝑡)

𝑚𝑎𝑥(𝑋𝑘𝑖𝑡)−𝑚𝑖𝑛(𝑋𝑘𝑖𝑡)
                                               (5) 

Where (Xkit) represents the value of the AI component k 
for region i at time t, and wk represents the component weight 
determined through principal component analysis. 

We employ PCA for weighting rather than arbitrary equal 
weights for several methodological reasons. First, PCA 
objectively determines weights based on the covariance 
structure of the data, capturing the common underlying factor 
of 'AI adoption intensity' while allowing components to 
contribute proportionally to their information content. 
Second, the high correlation among our four dimensions 

(ranging from 0.52 to 0.71) suggests a strong common factor 
that PCA efficiently extracts. Third, PCA addresses 
multicollinearity concerns that would arise from including all 
dimensions separately in regression models. The first 
principal component explains 67.4% of total variance, well 
above the 50% threshold conventionally required for index 
construction, with loadings of 0.412 (innovation), 0.387 
(human capital), 0.298 (investment), and 0.234 (research 
output). These loadings align with theoretical expectations, 
giving the highest weights to innovation and human capital, 
the fundamental drivers of AI capability. Robustness checks 
using alternative aggregation methods (geometric mean, 
equal weights, factor analysis) yield indices with correlations 
exceeding 0.92 with our PCA-based measure, confirming its 
validity. 

3.3 Econometric models 
Based on these variables, we specify the following 

econometric models to test our hypotheses. To investigate the 
mediating role of AI adoption in the relationship between 
industrial structure and economic development, we employ a 
three-step approach following Baron and Kenny [18]. First, 
we estimate the direct effect of industrial structure on 
economic outcomes: 

𝑌𝑖𝑡 = 𝛼0 + 𝛼1𝐼𝑆𝑖𝑡 + 𝛼2𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 휀𝑖𝑡                                   (6) 

Second, we examine the relationship between industrial 
structure and AI adoption: 

𝐴𝐼𝑖𝑡 = 𝛽0 + 𝛽1𝐼𝑆𝑖𝑡 + 𝛽2𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 𝜈𝑖𝑡                                  (7) 

Third, we estimate the full model including both industrial 
structure and AI adoption. To strengthen causal identification 
in our mediation analysis, we implement several robustness 
checks. We employ lagged values of industrial structure 
variables as instruments for current-period values, exploiting 
the persistence of industrial characteristics while breaking 
potential contemporaneous feedback loops. Additionally, we 
conduct Granger causality tests to verify the temporal 
precedence of industrial structure changes in relation to AI 
adoption, as well as AI adoption in relation to economic 
outcomes. The instrumental variable approach addresses 
potential endogeneity where: 

𝑌𝑖𝑡 = 𝛾0 + 𝛾1𝐼𝑆𝑖𝑡 + 𝛾2𝐴𝐼𝑖𝑡 + 𝛾3𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 휂𝑖𝑡               (8) 

Where Yit represents economic development indicators (GDP 
per capita growth or TFP growth) for region i at time t; ISit 

captures industrial structure characteristics (manufacturing 
value-added share, high-technology exports percentage, or 
industrial diversification index); AIit measures AI adoption 
intensity using our composite index; Xit includes control 
variables; 𝜇𝑖  and 𝜆𝑡  represent region and time fixed effects; 
and 휀𝑖𝑡, 𝜈𝑖𝑡, and 휂𝑖𝑡  are the respective error terms. 

Industrial upgrading and AI adoption may be 
endogenously determined through reverse causality or 
omitted variables. We address this using three strategies: (1) 
lagged industrial structure (t-2) as instruments, with first-
stage F-statistics exceeding 24.7 confirming relevance; (2) 
national AI strategy introduction timing (2016-2019) as 
exogenous variation; (3) Arellano-Bond GMM for dynamic 
endogeneity. Wu-Hausman tests reject exogeneity (p < 0.05), 
while IV estimates exceed OLS by 18-23%, suggesting 
downward bias if endogeneity is ignored. Core results remain 
robust: mediation effects range 48.2-57.4%, and thresholds 
stay within 0.426-0.449 across specifications. 

To identify threshold effects between AI adoption and 
economic outcomes, we use Hansen's threshold regression to 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

5 

 

determine critical AI adoption levels that alter economic 
relationships [19]. Our model is: 
𝑌𝑖𝑡 = 𝛿0 + 𝛿1𝐼𝑆𝑖𝑡 ⋅ 𝐼(𝐴𝐼𝑖𝑡 ≤ 휃) + 𝛿2𝐼𝑆𝑖𝑡 ⋅ 𝐼(𝐴𝐼𝑖𝑡 > 휃) +
𝛿3𝐴𝐼𝑖𝑡 + 𝛿4𝑋𝑖𝑡 + 𝜇𝑖 + 𝜆𝑡 + 𝜉𝑖𝑡                                                        (9) 

where I(.) is an indicator function that takes the value 1 
when the condition inside the parentheses is satisfied and 0, 
otherwise; 휃  represents the threshold value of AI adoption 
that potentially changes the relationship between industrial 
structure and economic outcomes. The threshold parameter 
휃 is estimated by minimizing the sum of squared residuals: 

휃̂ = argmin
𝜃

∑ ∑ 𝜉𝑖𝑡
2𝑇

𝑡=1

𝑁

𝑖=1
(휃)                                                   (10) 

Figure 2 illustrates the conceptual AI adoption-economic 
growth relationship. The relationship shows distinct regimes 
separated by a threshold θ. In the first regime (AI₍ᵢₜ₎ ≤ θ), the 
industrial structure's impact on growth is represented by 
coefficient δ₁, while in the second regime (AI₍ᵢₜ₎ > θ), this 
relationship strengthens to δ₂ (where δ₁ < δ₂). This 
demonstrates how AI's economic impact accelerates once 
implementation reaches critical threshold levels, creating 
nonlinear growth patterns. 

 
 
Figure 2. Conceptual illustration of threshold effects in AI adoption 

Threshold effect significance is evaluated using Hansen's 
likelihood ratio test with bootstrapped p-values. For multiple 
thresholds, we extend the model to accommodate up to three 
regimes following Bai and Perron [20], identifying potential 
multiple transition points as adoption increases. To examine 
how upgrading modifies demographic-growth relationships, 
we estimate: 

𝑌𝑖𝑡 = 휁0 + 휁1𝑊𝐴𝑃𝑖𝑡 + 휁2𝐼𝑈𝑖𝑡 + 휁3(𝑊𝐴𝑃𝑖𝑡 × 𝐼𝑈𝑖𝑡) + 휁4𝑋𝑖𝑡 +
𝜇𝑖 + 𝜆𝑡 + 𝜔𝑖𝑡                                                                                     (11) 

where 𝑊𝐴𝑃𝑖𝑡  represents the working-age population 
ratio (population aged 15-64 as a percentage of the total 
population) and 𝐼𝑈𝑖𝑡  is the industrial upgrading index. The 
coefficient 휁3  captures the interaction effect, indicating how 
the impact of the working-age population on economic 
growth varies across different levels of industrial upgrading. 

3.4 Case study methodology 
To complement our econometrics, we employ a multiple-

case study design following Yin's replication logic [21]. The 
case studies serve three critical functions in our research 
design: (1) validating econometric findings through 
mechanism identification, (2) revealing boundary conditions 
and contextual factors not captured in quantitative models, 
and (3) providing contradictory evidence that refines our 
theoretical understanding. To ensure methodological rigor 
and avoid confirmation bias, case selection was conducted 

after initial econometric analysis but before final model 
specification. This sequencing allowed us to identify puzzling 
patterns in the quantitative data—such as regions with high AI 

adoption but limited economic impact — that warranted 
deeper investigation. The preliminary econometric results 
(completed in March 2023) identified threshold effects and 
heterogeneous impacts across industrial contexts, which then 
guided our purposive sampling strategy to select cases that 
could illuminate these patterns. Importantly, insights from 
case studies conducted between April and September 2023 
led us to refine our econometric models, particularly in 
developing more nuanced measures of industrial upgrading 
and identifying omitted interaction effects. Case selection 
uses stratified sampling based on AI adoption intensity and 
industrial upgrading status, creating a matrix of high/low 
adoption and advanced/emerging industrial status. This 
facilitates both literal replication (similar results in similar 
contexts) and theoretical replication (contrasting results for 
anticipated reasons). Selection probability incorporates 
regional AI adoption level, industrial upgrading status, and 
relevant theoretical characteristics. 

𝑃(𝑠𝑒𝑙𝑒𝑐𝑡𝑖𝑜𝑛)𝑖 = 𝑓(𝐴𝐼𝑖 , 𝐼𝑈𝑖 , 𝑅𝑖)                                                     (12) 

 where 𝐴𝐼𝑖  represents the region's AI adoption level, 𝐼𝑈𝑖  
denotes industrial upgrading status, and 𝑅𝑖  encompasses 
regional characteristics relevant to theoretical heterogeneity. 
Data collection follows a triangulation strategy, incorporating 
multiple evidence sources [22]. We conducted 127 semi-
structured interviews across 16 cases, with 6-10 interviews 
per case spanning multiple organizational levels: senior 
executives (AI strategy), middle managers (implementation 
processes), technical staff (operational challenges), and 
external stakeholders (policy makers, industry associations). 
The interview protocol, developed through pilot testing with 
three organizations, contained 24 core questions organized 
around five themes: (1) AI adoption drivers and barriers, (2) 
implementation processes and timeline, (3) organizational 
changes and capability development, (4) performance 
impacts and measurement, and (5) external factors and 
ecosystem effects. Questions followed a funnel approach, 
beginning with open-ended prompts ('Describe your 
organization's AI journey') before probing specific 
mechanisms identified in our quantitative analysis. All 
interviews were recorded, transcribed verbatim, and 
returned to participants for validation. 

Our coding framework employed a hybrid deductive-
inductive approach. The initial codebook contained 31 
theory-derived codes mapped to our three hypotheses (e.g., 
'threshold_awareness,' 'capability_complementarity,' 
'demographic_interaction'). Through iterative coding of the 
first four cases, we inductively developed 19 additional codes 
capturing emergent themes. Two researchers independently 
coded 20% of transcripts, achieving inter-rater reliability of 
0.84 (Cohen's kappa), with discrepancies resolved through 
discussion. We used NVivo 12 for data management, 
employing matrix queries to identify patterns across cases 
and constant comparison techniques to refine theoretical 
categories. 

Triangulation occurred at multiple levels: data 
triangulation compared interview accounts with 
documentary evidence (annual reports, internal 
presentations, government statistics) and observational 
notes from 38 site visits; investigator triangulation involved 
three researchers analyzing each case independently before 
reaching consensus; and methodological triangulation 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

6 

 

integrated qualitative findings with case-specific quantitative 
indicators. For each key finding, we required corroboration 
from at least two data sources and two different stakeholder 
groups, documented in evidence tables linking claims to 
supporting data. 

To establish causal mechanisms in our qualitative 
analysis, we employ process tracing methodology to identify 
the temporal sequence of events and decision-making 
processes. We specifically document: (1) the timing of 
industrial policy changes and AI investment decisions, (2) the 
sequence of capability development and organizational 
adaptations, and (3) the lag structure between AI 
implementation and observed economic outcomes. This 
temporal evidence from case studies complements our 
econometric identification strategy by revealing the 'black 
box' of causal mechanisms that statistical analysis alone 
cannot fully capture. 

𝐸𝑖𝑗𝑘 = 𝜔1𝐼𝑖𝑗𝑘 + 𝜔2𝐷𝑖𝑗𝑘 + 𝜔3𝑂𝑖𝑗𝑘                                      (13) 

where 𝐸𝑖𝑗𝑘  represents the evidential strength for 

phenomenon k in organization j within region i; 𝐼𝑖𝑗𝑘 , 𝐷𝑖𝑗𝑘, and 

𝑂𝑖𝑗𝑘  represent interview, documentary, and observational 

evidence respectively; and 𝜔1 , 𝜔2 , and 𝜔3  denote source-
specific weights determined through reliability assessment. 

For cross-case analysis, we employ pattern-matching 
structured around our three hypotheses, systematically 
comparing empirical patterns with theoretical predictions. 
The analysis combines within-case analysis with cross-case 
comparison using both variable and process-oriented 
techniques [23]. We apply modified Qualitative Comparative 
Analysis to identify necessary and sufficient conditions for 
successful AI-driven industrial upgrading. 

𝑌𝑖 = 𝑓(𝐶1𝑖 , 𝐶2𝑖 , . . . , 𝐶𝑛𝑖)                                                                     (14) 

where 𝑌𝑖  represents the outcome of interest (successful 
industrial upgrading) for case i, and 𝐶1𝑖  through 𝐶𝑛𝑖 represent 
configurational conditions including institutional quality, 
complementary capabilities, and implementation 
approaches. Figure 3 illustrates our analytical framework for 
cross-case comparison, demonstrating how individual case 
findings are systematically integrated into pattern 
identification. 

 
Figure 3. Cross-case analytical framework for AI-driven industrial 
upgrading 

The integration of quantitative and qualitative findings 
follows a sequential explanatory design [24] where case 
studies elaborate and expand upon econometric results. 

 

3.5 Data Sources and Sample 
Our econometric analysis draws on multiple 

complementary data sources covering 2010-2023 for 87 
countries across five continents, selected based on data 
availability and economic significance. The panel dataset 
combines: 
GDP per capita growth and total factor productivity data from 
the World Bank's World Development Indicators (WDI) and 
Penn World Table 10.0, supplemented by OECD National 
Accounts for high-income countries. Industrial structure 
variables, including manufacturing value-added share and 
high-technology exports, are sourced from the UNIDO 
Industrial Statistics Database and the World Bank's World 
Integrated Trade Solution (WITS). 
We construct our composite AI adoption index using: (1) AI 
patent applications from WIPO Global Innovation Index and 
PATSTAT database, (2) AI talent concentration from LinkedIn 
Talent Insights and national labor force surveys, (3) AI 
investment data from Crunchbase, PitchBook, and national 
venture capital associations, and (4) AI research publications 
from Scopus and Web of Science. The index covers 28 
manufacturing sectors (ISIC Rev.4 two-digit codes) and 15 
service sectors. 
Working-age population ratios from UN Population Division, 
human capital indices from Barro-Lee Educational 
Attainment Dataset, institutional quality measures from 
Worldwide Governance Indicators, and trade openness from 
WTO Statistics Database. 

To address data quality issues, we implement a 
systematic approach. For missing observations constituting 
8.3% of the initial dataset, we employ multiple imputation 
using chained equations (MICE) when missingness is random, 
validated through Little's MCAR test (χ² = 1847.3, p = 0.092). 
For systematic gaps in AI metrics for developing countries, we 
utilize a two-stage approach: first, predicting missing values 
using observable correlates (ICT infrastructure, R&D 
expenditure, tertiary education enrolment), then adjusting 
predictions based on regional benchmarks. Cross-validation 
with alternative data sources ensures consistency — for 
instance, correlating our AI adoption index with Stanford's AI 
Index (r = 0.89) for overlapping country-years. Outliers 
beyond 3.5 standard deviations are investigated through 
news searches and government reports, retaining those 
reflecting genuine economic shocks while winsorizing 
measurement errors at the 1st and 99th percentiles. The final 
balanced panel comprises 1,131 country-year observations 
with complete data across all key variables. 

The geographical distribution of our sample includes 28 
high-income countries (32.2% of observations), 35 middle-
income countries (40.2%), and 24 low-income countries 
(27.6%), ensuring adequate representation across 
development levels. Sectoral coverage spans 28 
manufacturing industries following ISIC Rev.4 classification, 
from traditional sectors (food processing, textiles) to high-
technology industries (electronics, pharmaceuticals), plus 15 
service sectors. This comprehensive coverage enables 
examination of AI adoption patterns across diverse industrial 
contexts. The 2010-2023 timeframe captures both the 
emergence phase of industrial AI applications (2010-2015) 
and the acceleration period following breakthrough 
developments in deep learning (2016-2023), providing 
sufficient variation to identify threshold effects and structural 
changes in the AI-economy relationship. 

 

Case 1 Analysis Case 2 Analysis Case n Analysis

Cross-Case Pattern ldentification

Thematic Synthesis and Theory DevelopmentEconometric Results

··· ···



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

7 

 

4. Results 

4.1 Descriptive analysis 
This section highlights key dataset trends: High-income 

economies show higher manufacturing sophistication; East 
Asian economies lead in diversification. AI adoption varies 
regionally—North American and East Asian economies show 
the steepest curves, with nonlinear growth benefits 
intensifying beyond threshold levels. Figure 4 illustrates AI-
growth nonlinearities, with association intensifying once the 
AI index exceeds 0.45, particularly at higher adoption levels. 

 
(a) Regional AI adoption trends 

 

 
(b) Sectoral AI adoption by income level 

 

 
(c) AI adoption vs. GDP per capita growth 

 
Figure 4. Patterns of AI adoption and economic relationships 

AI penetration correlates strongly with manufacturing 
sophistication (r=0.68) and diversification (r=0.64), but 
weakly with manufacturing value-added share (r=0.21), 
suggesting AI links more with qualitative aspects than 
manufacturing scale. Working-age population's growth 
impact varies by industrial upgrading level (r=0.43 in top 
quartile vs r=0.18 in bottom quartile), supporting our 
upgrading-moderates-demographics hypothesis. These 
patterns motivate formal testing through mediation analysis. 

4.2 Mediating effect analysis 
Before presenting our mediation results, we first 

establish the temporal ordering and causal direction of our 
key relationships. Granger causality tests confirm that 
industrial structure changes temporally precede AI adoption 
(F-statistic = 18.73, p < 0.001), while AI adoption precedes 
economic outcome changes (F-statistic = 14.52, p < 0.001). 
Instrumental variable estimates using lagged values and 
policy shocks yield consistent but slightly larger effects, 
confirming robustness to endogeneity. Reverse causality tests 
show no significant effects in the opposite direction, 
supporting our hypothesized causal chain. Furthermore, our 
instrumental variable estimates using lagged industrial 
structure values yield consistent results with slightly larger 
coefficients, suggesting that endogeneity bias, if present, 
attenuates rather than inflates our estimates. AI adoption 
mediates industrial structure-economic outcomes 
relationships. Industrial diversification affects growth more 
strongly (β≈0.24) than manufacturing value-added (β≈
0.18). All structure indicators predict AI adoption, with 
manufacturing sophistication showing the strongest 
relationship (β≈0.46). When controlling for AI adoption, 
structure coefficients decrease while AI shows significant 
positive effects, confirming mediation—approximately 53% of 

manufacturing sophistication's growth effect occurs through 
AI. Mediation is stronger for technological sophistication than 
manufacturing scale, stronger for TFP than GDP growth, and 
higher in advanced economies and recent periods. Table 1 
reports standardized coefficients with standard errors in 
parentheses and 95% confidence intervals in brackets. All 
models include control variables (human capital, investment 
ratio, trade openness, institutional quality) and country and 
year fixed effects. Significance levels: * p < 0.10, ** p < 0.05, 
*** p < 0.01. 

Table 1. Direct effects of industrial structure on economic growth 
(step 1) 

 

 

 

Industrial 
Structure 
Variable 

GDP 
per 

Capita 
Growth 

 TFP 
Growth 

 Economic 
Complexity 

Change 

 

Manufacturing 
value-added 
share 

0.183*** 
(0.042) 

[0.101, 
0.265] 

0.146** 
(0.053) 

[0.043, 
0.249] 

0.124** 
(0.046) 

[0.034, 
0.214] 

High-
technology 
exports 
percentage 

0.216*** 
(0.047) 

[0.124, 
0.308] 

0.183*** 
(0.051) 

[0.083, 
0.283] 

0.318*** 
(0.045) 

[0.230, 
0.406] 

Industrial 
diversification 
index 

0.241*** 
(0.039) 

[0.165, 
0.317] 

0.235*** 
(0.044) 

[0.149, 
0.321] 

0.253*** 
(0.041) 

[0.173, 
0.333] 

Manufacturing 
sophistication 
index 

0.197*** 
(0.043) 

[0.113, 
0.281] 

0.172*** 
(0.048) 

[0.078, 
0.266] 

0.287*** 
(0.042) 

[0.205, 
0.369] 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

8 

 

Our second analysis step examines how industrial 
structure influences AI adoption. Results show all industrial 
structure indicators positively predict AI adoption, with 
manufacturing sophistication showing the strongest 
relationship ( β =0.463, p<0.001), followed by high-

technology exports ( β =0.385, p<0.001). Manufacturing 

value-added share shows a more modest association (β
=0.217, p<0.01), suggesting technological sophistication 
facilitates AI implementation more than industrial scale, 
likely by providing necessary absorptive capacity and 
complementary capabilities. These differential effects add 
important nuance to understanding technological diffusion 
patterns. Table 2 reports standardized coefficients with 
standard errors in parentheses and 95% confidence intervals 
in brackets. All models include control variables and fixed 
effects as in previous models. The composite AI adoption 
index is the primary outcome variable, with AI patent 
intensity and AI skills prevalence serving as alternative 
measures for robustness testing. Significance levels: * p < 
0.10, ** p < 0.05, *** p < 0.01. Based on these three analytical 
steps, we calculate the indirect effects of industrial structure 
on economic outcomes through AI adoption and decompose 
the total effects into direct and indirect components. Next, we 
examine whether these relationships exhibit threshold 
effects. 

4.3 Threshold Effect Analysis 
This section presents empirical evidence on threshold 

effects in the relationship between AI adoption, industrial 
structure, and economic outcomes. Applying Hansen's (1999) 
threshold regression methodology, we identify critical levels 
of AI adoption that fundamentally alter the nature of 
industrial-economic relationships. The results provide strong 
support for our second hypothesis regarding the existence of 
significant threshold effects in AI's economic impact. The 
threshold estimation employs a grid search over the 15th to 
85th percentiles of the AI adoption distribution, with 
increments of 0.001, testing 542 potential threshold values. 
For each candidate threshold, we calculate the sum of squared 
residuals and select the value minimizing this criterion. The 
estimated threshold of 0.438 (95% CI: 0.425-0.451) is 
statistically significant based on Hansen's likelihood ratio test 
(LR = 47.23, p < 0.001), where p-values are obtained through 
1,000 bootstrap replications following Hansen (2000). To 
assess the robustness of this critical threshold, we conduct 
extensive sensitivity analyses.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

First, bootstrap confidence intervals constructed using 
the percentile method across 5,000 replications consistently 
place the threshold between 0.422 and 0.454, confirming the 
stability of our point estimate. Second, subsample analysis 
reveals remarkable consistency: excluding any single country 
changes the threshold by at most 0.009, while rolling window 
estimation (using 10-year windows) produces thresholds 
ranging from 0.431 to 0.446. Third, alternative threshold 
detection methods yield similar results —Andrews' (1993) 

supremum Wald test identifies a break at 0.441, while Bai-
Perron sequential testing confirms a single threshold at 0.435. 
Fourth, we test sensitivity to functional form assumptions by 
estimating thresholds in models with quadratic terms and 
interaction effects, finding threshold values within 0.012 of 
our baseline estimates. The economic significance of the 
threshold is validated through placebo tests. When we 
artificially impose thresholds at the 25th percentile (0.287) or 
75th percentile (0.614) of AI adoption, the regime-specific 
coefficients show no statistically significant differences (p > 
0.10), confirming that the identified threshold represents a 
genuine structural break rather than a statistical artifact. 
Moreover, the threshold's stability across different industrial 
structure measures—varying by only 0.008-0.021 when using 

alternative indicators — suggests it captures a fundamental 
characteristic of AI's economic impact rather than 
measurement-specific variations. Figure 5 illustrates this 
interaction effect by plotting the estimated marginal effect of 
working-age population on GDP per capita growth across 
different levels of AI adoption, clearly showing the 
strengthening relationship as AI adoption increases. 

 
Figure 5. AI threshold effects on demographic-growth relationships 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Table 2. Effects of industrial structure on AI adoption (Step 2) 

Industrial Structure 
Variable 

AI Adoption 
(Composite Index) 

 AI Patent 
Intensity 

 AI Skills 
Prevalence 

 

Manufacturing value-
added share 

0.217*** (0.051) [0.117, 
0.317] 

0.186*** 
(0.052) 

[0.084, 
0.288] 

0.228*** (0.053) [0.124, 
0.332] 

High-technology 
exports percentage 

0.385*** (0.047) [0.293, 
0.477] 

0.427*** 
(0.046) 

[0.337, 
0.517] 

0.356*** (0.048) [0.262, 
0.450] 

Industrial 
diversification index 

0.321*** (0.045) [0.233, 
0.409] 

0.276*** 
(0.047) 

[0.184, 
0.368] 

0.307*** (0.046) [0.217, 
0.397] 

Manufacturing 
sophistication index 

0.463*** (0.042) [0.381, 
0.545] 

0.439*** 
(0.044) 

[0.353, 
0.525] 

0.412*** (0.045) [0.324, 
0.500] 

 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

9 

 

Figure 5 shows working-age population's growth impact 
changes dramatically at the AI threshold (0.438). Below this 
level, effects are modest and marginally significant; above it, 
they strengthen substantially with high significance. This 
suggests demographic advantages require AI capabilities to 
effectively translate to growth. Analysis across upgrading 
quartiles reveals that AI adoption thresholds decrease with 
industrial sophistication, while demographic impact 
differentials increase (0.287 in the highest quartile versus 
0.138 in the lowest), indicating that advanced regions 
experience stronger threshold effects. Table 3 presents 
threshold estimates and regime-specific coefficients for the 
working-age population across industrial upgrading 
quartiles. All models include the full set of control variables 
and fixed effects. Coefficient differential represents the 
absolute difference between above-threshold and below-
threshold coefficients. Standard errors in parentheses. 
Significance levels: * p < 0.10, ** p < 0.05, *** p < 0.01. 

Our analysis confirms nonlinear relationships between 
AI adoption, industrial structure, and economic outcomes. 
Statistically significant thresholds across specifications 
support our second hypothesis on critical AI adoption levels 
that alter industrial-economic relationships. Regime-specific 
analysis supports our third hypothesis that industrial 
upgrading moderates demographic-economic growth 
relationships. These findings suggest AI adoption strategies 
should target surpassing critical thresholds to maximize 
economic benefits. Case studies provide deeper insights into 
these quantitative patterns. 

 

Table 3. Working-age population effects across industrial upgrading 
quartiles 

Industrial 
Upgrading 

Level 

AI 
Adoption 

Threshold 

Working-
Age 

Population 
Coefficient 

 Coefficient 
Differential 

  Below 
Threshold 

Above 
Threshold 

 

First 
Quartile 
(Lowest) 

0.246 0.072 
(0.062) 

0.210** 
(0.095) 

0.138 

Second 
Quartile 

0.318 0.097* 
(0.059) 

0.283*** 
(0.082) 

0.186 

Third 
Quartile 

0.386 0.118** 
(0.057) 

0.362*** 
(0.076) 

0.244 

Fourth 
Quartile 
(Highest) 

0.467 0.124** 
(0.054) 

0.411*** 
(0.071) 

0.287 

 
 

4.4 Case Study Findings 
This section presents findings from our 16-case study 

analysis across diverse contexts, complementing our 
econometric results with a deeper understanding of AI-driven 
industrial upgrading processes. Regional analysis reveals 
substantial implementation disparities. Figure 6 shows that 
high-income regions demonstrate balanced development 
across all dimensions, with strengths in data infrastructure 
and technical capabilities. Middle-income regions show 
strong strategic prioritization but weaknesses in 
infrastructure, while low-income regions have promising 
workforce engagement despite technical capability 
challenges. These patterns highlight the importance of 
contextually adapted implementation approaches addressing 
region-specific constraints. 

 
Figure 6. Regional comparison of AI implementation characteristics 

As illustrated in Figure 6, the most pronounced regional 
disparities appear in data infrastructure and technical 
capabilities dimensions, with high-income regions scoring 
approximately 2.7 times higher than low-income regions on 
these factors. These fundamental enablers create significant 
implementation barriers in resource-constrained contexts. 
However, the relatively smaller gap in workforce engagement 
(high-income regions scoring only 1.04 times higher than 
low-income regions) suggests an opportunity for low-income 
regions to leverage human capital development as an entry 
point for AI implementation. This pattern aligns with our 
econometric findings on the interaction between AI adoption 
and working-age population, suggesting that regions with 
demographic advantages can potentially offset some 
technical limitations through effective human capital 
development. 

5. Discussion 

Empirical evidence confirms that AI has a significant 
impact on economic growth through industrial development. 
Our analysis shows that AI channels have structural impacts 
on economic performance, extending theories that previously 
treated technology and industry separately. Our research 
identifies crucial threshold effects at AI adoption levels of 
0.43-0.45. Beyond this threshold, industrial capabilities' 
impact increases dramatically — manufacturing 
sophistication's effect on GDP growth triples, quantifying 
nonlinearities in technological adoption. AI fundamentally 
transforms demographic-economic relationships, with the 
working-age population showing minimal growth impact 
below thresholds but strong effects above. This modification 
of demographic dividends is crucial for countries facing 
demographic transitions alongside digital transformation. 
The integration of case study evidence significantly refines 
our understanding of these threshold effects. While 
econometric analysis suggested a sharp discontinuity at the 
threshold, case studies reveal a more gradual transition zone 
spanning approximately ± 0.05 around the estimated 
threshold value. The German automotive sector case 
exemplifies this: firms began experiencing productivity gains 
at AI adoption of 0.40, but the transformative reorganization 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

10 

 

of production networks occurred only after reaching 0.48. 
This finding prompted us to test alternative threshold 
specifications with smoother transition functions, though the 
discrete threshold model ultimately provided a superior fit. 
Additionally, case studies uncovered two patterns absent 
from our initial quantitative analysis: (1) the critical role of 
inter-firm knowledge spillovers in achieving threshold 
effects, observed in all successful high-income cases but only 
40% of middle-income cases, and (2) the existence of 
'adoption traps' where regions achieve moderate AI adoption 
(0.35-0.42) but lack the complementary investments to push 
beyond the threshold. These insights directly informed our 
policy recommendations regarding the importance of 
coordinated AI ecosystem development rather than isolated 
firm-level adoption. 

The causal nature of our findings is supported by 
multiple empirical strategies that go beyond correlational 
evidence. The temporal sequencing in our panel data, where 
we observe industrial structure changes preceding AI 
adoption and subsequently affecting economic outcomes, 
provides strong evidence for the hypothesized causal chain. 
Our instrumental variable approach addresses potential 
endogeneity concerns, while the threshold effects identified 
through quasi-experimental variation in AI policy 
implementation further strengthen causal interpretation. The 
consistency of results across different identification 
strategies—including fixed effects, instrumental variables, and 

threshold regression discontinuities — provides robust 

evidence that the relationships we document reflect causal 
mechanisms rather than spurious correlations. Moreover, our 
case study evidence reveals specific mechanisms through 
which causality operates, such as the development of 
complementary capabilities and organizational restructuring 
that follow AI adoption decisions. 

AI contributes through multiple pathways: immediate 
process optimization gains (23.6% efficiency) and later-stage 
value chain innovations, following a J-curve pattern as 
capabilities develop. Implementation success varies by 
region. High-income economies lead through strong 
infrastructure and capabilities, middle-income regions show 
intent but implementation gaps, while low-income regions 
display workforce engagement despite infrastructure 
constraints. This necessitates context-specific policies. 
Sustainable AI-driven growth depends on translating 
productivity into inclusive benefits, with balanced human-AI 
collaboration achieving more sustainable improvements than 
wholesale automation. 

6. Conclusion 

This research examined the impact of AI adoption on 
industrial upgrading and economic development. We found 
AI mediates between industrial structure and economic 
outcomes, with stronger effects for qualitative aspects (52.8% 
for manufacturing sophistication) than quantitative measures 
(32.1% for manufacturing value-added). Clear threshold 
effects exist (AI index ≈0.43-0.45) where capabilities' growth 
impact increases dramatically, and demographic factors' 
influence strengthens. Theoretically, we extend growth 
models by connecting AI to structural change, identifying 
critical threshold effects challenging linear models, and 
demonstrating technology's moderation of demographic-
economic relationships. Methodologically, our AI penetration 
index addresses measurement challenges for general-
purpose technologies. Policy implications include prioritizing 
implementation critical mass, calibrating approaches to 
development contexts, developing "translational capacity," 

and investing in data infrastructure for low-income regions. 
Limitations include data availability constraints, particularly 
for AI metrics in low-income countries before 2015, despite 
our systematic imputation approach. The reliance on proxy 
measures for AI adoption in some contexts may 
underestimate actual implementation in informal sectors. 
Future research should explore specific AI applications' 
effects, sectoral variations, and long-term sustainability 
regarding distributional outcomes. 

Ethical issue 

The authors are aware of and comply with best practices in 

publication ethics, specifically with regard to authorship 

(avoidance of guest authorship), dual submission, 

manipulation of figures, competing interests, and compliance 

with policies on research ethics. The authors adhere to 

publication requirements that the submitted work is original 

and has not been published elsewhere. 

Data availability statement 

The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1]  K. Schwab, The fourth industrial revolution, World 
Economic Forum, 2016. 
https://www.weforum.org/about/the-fourth-
industrial-revolution-by-klaus-schwab/ 

[2]  D.R. Heath, Prediction machines: the simple 
economics of artificial intelligence: by Ajay Agrawal, 
Joshua Gans and Avi Goldfarb, Published in 2018 by 
Harvard Business Review Press, 272 pp., 
30.00(hardcover),KindleEdition: 16.19, ISBN: 978-1-
633695672, Taylor & Francis, 2019. 
https://www.predictionmachines.ai/ 

[3]  E. Brynjolfsson, D. Rock, C. Syverson, The productivity 
J-curve: How intangibles complement general 
purpose technologies, American Economic Journal: 
Macroeconomics 13(1) (2021) 333-372. DOI：
10.1257/mac.20180386 

[4]  E. Brynjolfsson, A. McAfee, The second machine age: 
Work, progress, and prosperity in a time of brilliant 
technologies, WW Norton & company2014. 
https://wwnorton.com/books/the-second-machine-
age/ 

[5]  N. Zolas, Z. Kroff, E. Brynjolfsson, K. McElheran, D.N. 
Beede, C. Buffington, N. Goldschlag, L. Foster, E. 
Dinlersoz, Advanced technologies adoption and use 
by us firms: Evidence from the annual business 
survey, National Bureau of Economic Research, 2021. 
DOI：10.3386/w28290 

[6]  D.E. Bloom, D. Canning, G. Fink, Implications of 
population ageing for economic growth, Oxford 
review of economic policy 26(4) (2010) 583-612. DOI
：10.1093/oxrep/grq038 

[7]  D. Acemoglu, P. Restrepo, The race between man and 
machine: Implications of technology for growth, 
factor shares, and employment, American economic 
review 108(6) (2018) 1488-1542. DOI：

10.1257/aer.20160696 
[8]  J. Furman, R. Seamans, AI and the Economy, 

Innovation policy and the economy 19(1) (2019) 
161-191. DOI：10.1086/699936 



X. Hao et al. /Future Technology                                                                                           November 2025| Volume 04 | Issue 04 | Pages 01-11 

 

11 

 

[9]  P. Aghion, U. Akcigit, P. Howitt, What do we learn 
from Schumpeterian growth theory?, Handbook of 
economic growth, Elsevier2014, pp. 515-563. DOI：
10.1016/B978-0-444-53540-5.00001-X 

[10]  D. Rodrik, Premature deindustrialization, Journal of 
economic growth 21 (2016) 1-33. DOI: 
10.1007/s10887-015-9122-3 

[11]  D. Andrews, C. Criscuolo, P.N. Gal, The best versus the 
rest: the global productivity slowdown, divergence 
across firms and the role of public policy,  (2016). 
DOI: http://dx.doi.org/10.1787/63629cc9-en 

[12]  K. Lee, Schumpeterian analysis of economic catch-up: 
Knowledge, path-creation, and the middle-income 
trap, Cambridge University Press2013. DOI: 
https://doi.org/10.1017/CBO9781107337244 

[13]  D. Canning, J. Sevilla, The Demographic Dividend: A 
New Perspective on the Economic Consequences of 
Population Change, Rand Corporation2000. 
https://www.rand.org/pubs/monograph_reports/M
R1274.html 

[14]  E. Brynjolfsson, K. McElheran, The rapid adoption of 
data-driven decision-making, American Economic 
Review 106(5) (2016) 133-139. DOI: 
10.1257/aer.p20161016 

[15]  E. Brynjolfsson, D. Rock, C. Syverson, Artificial 
intelligence and the modern productivity paradox, 
The economics of artificial intelligence: An agenda 
23(2019) (2019) 23-57. DOI: 
https://doi.org/10.7208/9780226613475-003 

[16]  D.J. Teece, Profiting from innovation in the digital 
economy: Enabling technologies, standards, and 
licensing models in the wireless world, Research 
policy 47(8) (2018) 1367-1387. DOI: 
10.1016/j.respol.2017.01.015 

[17]  B.E. Hansen, Sample splitting and threshold 
estimation, Econometrica 68(3) (2000) 575-603. DOI: 
10.1111/1468-0262.00124 

[18]  R.M. Baron, D.A. Kenny, The moderator–mediator 
variable distinction in social psychological research: 
Conceptual, strategic, and statistical considerations, 
Journal of personality and social psychology 51(6) 
(1986) 1173. DOI: 10.1037/0022-3514.51.6.1173 

[19]  B.E. Hansen, Threshold effects in non-dynamic 
panels: Estimation, testing, and inference, Journal of 
econometrics 93(2) (1999) 345-368. DOI: 
10.1016/S0304-4076(99)00025-1 

[20]  J. Bai, P. Perron, Estimating and testing linear models 
with multiple structural changes, Econometrica  
(1998) 47-78. DOI: 10.2307/2998540 

[21]  R.K. Yin, Case study research and applications: Design 
and methods, Sage publications2017. 
https://us.sagepub.com/en-us/nam/case-study-
research-and-applications/book250150 

[22]  K.M. Eisenhardt, M.E. Graebner, Theory building from 
cases: Opportunities and challenges, Academy of 
management journal 50(1) (2007) 25-32. 
10.5465/amj.2007.24160888 

[23]  M.B. Miles, A.M. Huberman, J. Saldaña, Qualitative 
data analysis: A methods sourcebook. 3rd, Thousand 
Oaks, CA: Sage, 2014. https://us.sagepub.com/en-
us/nam/qualitative-data-analysis/book246128 

[24]  J.W. Creswell, V.L.P. Clark, Designing and conducting 
mixed methods research, Sage publications2017. 
https://us.sagepub.com/en-us/nam/designing-and-
conducting-mixed-methods-research/book241842 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/

