Integration of Artificial Intelligence into Human Resource Management in Manufacturing Enterprises: A Systematic Literature Review of Challenges, Approaches, and Evolution (2000–2025)

Abstract

With the advancement of digital technology and Industry 4.0, artificial intelligence (AI) is gradually embedded in human resource management and has become an important digital foundation to support the sustainable transformation of enterprises. However, the research in the manufacturing context, particularly through the challenge perspective at different levels, remains fragmented. This work represents a systematic review of 347 articles from Scopus and Web of Science from 2000 to 2025 and employs a dual-method analysis strategy embracing metrics and in-depth coding on 100 core publications. Excel, Bibliometrix, CiteSpace, Latent Dirichlet Allocation (LDA), and VOSviewer were utilized for quantitative analysis, while open–axial–selective coding of the Grounded theory approach was applied to generate qualitative results. The findings revealed six key challenges in integrating AI-HRM within manufacturing and six approaches to solve the identified issues. The Challenge–Approach Matching Matrix was constructed, illustrating the suitability of different pathways for addressing specific challenges. Analysis of thematic evolution in AI-HRM research resulted in the identification of three distinctive phases and demonstrated a consistent shift from technology-centric approaches towards human–machine collaboration. The primary contribution of this research lies in proposing a Multi-Level Embedded Framework providing a complex view of AI-HRM in a manufacturing sector at micro, meso, and macro levels. The absence of sustainable HR transformation through AI integration was identified as the critical challenge at the macro level. This research provides theoretical and practical implications for designing the sustainable HRM system based on ESG principles and favors the United Nations Sustainable Development Goals 9 and 12.

1. Introduction

As a pillar industry of global economic development, manufacturing is undergoing a process of deep integration with Industry 4.0, smart manufacturing, and the Industrial Internet. Artificial intelligence (AI) is expanding from application scenarios, such as production scheduling, quality inspection, and equipment operation, into the domain of human resource management (HRM) [1,2]. From CV screening and interview assessment to personalized training and safety checks, AI-HRM systems are becoming vital tools for enhancing organizational efficiency and improving employee experience [3,4].

This trajectory is corroborated by authoritative international and governmental reports. The World Economic Forum indicates that over the next decade, the job structure within manufacturing will undergo a significant transformation [5]. Whilst AI and automation will replace highly standardized and repetitive tasks, new roles requiring data literacy and interdisciplinary skills for human–machine collaboration will emerge. Policies issued by China’s Ministry of Industry and Information Technology on manufacturing digital transformation and smart manufacturing further emphasize that while establishing smart factories and digital workshops, AI-enhanced HRM reforms must be advanced concurrently [6]. This entails applying AI capabilities to talent recruitment, skills training, and performance appraisal processes. Consequently, the integration of AI-HRM within manufacturing demands not only technological advancement but also systematic enhancement of organizational capabilities and human resource allocation [7].

However, AI technology has not been smoothly applied in management. When implementing AI-HRM systems in manufacturing, some projects demonstrate promising results during proof-of-concept phases. Yet, during actual deployment, systems often become idle due to insufficient data literacy among personnel [8]. Furthermore, ethical controversies surrounding algorithmic bias, fairness, and privacy protection have eroded employee trust in AI [9,10]. Furthermore, inadequate interface design between certain HRM processes and production operations systems has hindered the integration of AI-supported decision-making into daily workflows [11]. This sufficiently demonstrates that integrating AI-HRM within manufacturing contexts constitutes a complex endeavor requiring the coordinated alignment of multiple factors, including technology, organizational structures, and human elements [12].

1.1. Research Gaps

In recent years, systematic reviews concerning artificial intelligence and human resource management have proliferated rapidly (e.g., [13,14,15]). While these reviews elucidate the application and outcomes of AI within HR functions, gaps persist when examining research specifically focused on manufacturing contexts.

Firstly, most existing reviews adopt cross-sectoral perspectives and therefore provide limited discussion of how manufacturing-specific conditions—such as safety-sensitive workflows, heterogeneous data infrastructures, and shop-floor workforce structures—shape AI-HRM challenges and solutions. As a result, insights derived from service- or knowledge-intensive settings may not fully capture the contextual pressures faced by manufacturing enterprises. At the same time, manufacturing-oriented studies remain relatively fragmented and predominantly application-driven, lacking integrative synthesis [16].

Secondly, prior reviews rarely incorporate structured ‘challenge–pathway’ mappings, offering limited guidance on which integration approaches are most suitable for addressing specific implementation barriers. Third, the theorization of multi-level perspectives (macro–meso–micro) and their feedback mechanisms remains underdeveloped, particularly regarding how institutional constraints and organizational governance interact with employee trust and adoption.

Table 1. Comparison between this study and representative AI-HRM reviews.

Study	Manufacturing Focus	Analytical Structure	Multi-Level Perspective
[13]	No	AI-HRM interactions	No
[14]	No	Application-oriented review	No
[15]	Partial	Theoretical tensions	Limited
This study	Yes	Challenges–Approaches–Evolution	Macro–Meso–Micro (MLEF)

shows the distinctive features of the present review in comparison to the earlier works.

To address these gaps, this study reviews 347 manufacturing-related AI-HRM articles (2000–2025), combining bibliometric analysis with in-depth coding of 100 core papers to identify key challenges, synthesize integration pathways, map challenge–approach relationships, and propose a multi-level embedded framework.

1.2. Research Questions

To address the aforementioned gaps in both academic and practical contexts, a systematic search and screening of the relevant English-language literature indexed on Scopus and Web of Science from 2000 to 2025 was conducted, adhering to the Preferred Reporting Items for Systematic reviews and Meta-Analyses, PRISMA guidelines 2020 [17]. Specifically, this study aims to address the following three questions:

RQ1 (Challenge identification): What key challenges do manufacturing enterprises encounter during the AI-HRM integration process? How are these challenges distributed and co-occur across different HR modules (e.g., recruitment, training, and performance management)?

RQ2 (Pathway integration): What AI-HRM integration pathways and theoretical frameworks have existing studies proposed? What matching relationships exist between these pathways and challenges? Which pathways demonstrate greater practicality and feasibility within manufacturing contexts?

RQ3 (Evolutionary insights): How did AI-HRM research themes in manufacturing evolve between 2000 and 2025? What shifts and changes in focus emerged across macro (national and policy), meso (organizational and factory), and micro (employee and job) levels?

To address these research questions, bibliometric tools such as Bibliometrix and VOSviewer were employed to visualize annual trends in the literature, national and journal distributions, and co-occurrence networks of author keywords. Concurrently, manual coding of challenge types and resolution pathways underpinned the construction of a Challenge–Approach Matching Matrix and a Multi-Level Embedded Framework. This approach emphasizes the systematic coupling of technology, organization, and personnel, rather than merely listing fragmented challenges or singular pathways. It resonates with recent interdisciplinary AI-HRM review trends towards “multi-level integration” and “human-centered orientation” [2,18,19].

To systematically address these research questions, this study employs a structured review design that distinguishes between methodological procedures, empirical findings, and theoretical interpretations. Specifically, Section 2 details the research design and PRISMA-based screening process. Section 3 derives descriptive and thematic findings through bibliometric analysis and qualitative coding. Section 4 synthesizes these findings via a multi-level interpretative framework, whilst Section 5 summarizes relevant implications, limitations, and directions for future research.

2. Methodology

This section details the process of the literature screening, appraisal, and analysis to ensure transparency and reproducibility. It addresses the research design, data sources, selection criteria, and analytical procedures solely, without interpreting the empirical findings. The methodological choices made here lay the groundwork for the descriptive and thematic results presented in subsequent chapters.

This study adhered to the PRISMA 2020 guidelines and followed a predefined review protocol established by the authors prior to the literature retrieval. The review systematically searched and screened the English-language literature on artificial intelligence and human resource management in manufacturing from 2000 to 2025. The protocol specified research questions (RQ1–RQ3), databases, search strategies, inclusion and exclusion criteria, screening procedures, and coding frameworks.

Given that this study constitutes a bibliometrics-driven, qualitative systematic review without effect size synthesis or clinical outcomes, the protocol was not registered in PROSPERO or OSF. All protocol components are transparently reported within the main text, appendices, and Supplementary Materials to ensure methodological reproducibility.

A dual-track strategy combining bibliometric analysis and content coding was employed for comprehensive research [17,20,21,22]. Methodologically, this study incorporates three distinctive designs: first, adjusting the threshold for identifying challenges to uncover emerging, under-researched topics within manufacturing contexts; second, it categorizes AI-HRM integration into six pathways from an interdisciplinary perspective, enhancing classification granularity [23]; and third, it employs emergent term detection and Latent Dirichlet Allocation (LDA), latent topic modeling to chart research evolution from 2000 to 2025 [24,25].

2.1. Review Framework and Scope

This study constructs a Technology–Organization–Institutional Review framework to address three core research questions (RQ1–RQ3). The literature examined spans 2000–2025, covering the complete evolutionary cycle from the partial application of expert systems and traditional automation in HR tasks to the widespread adoption of Industry 4.0, smart manufacturing, and AI-driven human resources within manufacturing [15,26].

In terms of scope, this paper focuses on English-language journal articles and reviews the literature concerning the integration of AI and HRM within manufacturing contexts. Specific inclusion and exclusion criteria are detailed in Section 2.3. This defined scope ensures the review’s comprehensiveness while emphasizing the manufacturing context, providing clear boundaries for subsequent challenge identification and pathway mapping.

2.2. Search Strategy and Screening Process

The literature retrieval was conducted within the Web of Science (WoS) and Scopus databases, two of the most commonly used and comprehensive databases for systematic reviews [20,27]. The database search was conducted and completed on 4 November 2025, covering all records indexed in Scopus and Web of Science up to that date.

The search strategy design centered on three core concepts: AI, HRM, and manufacturing. After multiple rounds of preliminary searches and keyword optimization, with field restrictions applied to Title, Abstract, and Author Keywords, the following Boolean logic search terms were finalized (see Appendix A).

The initial search yielded 1490 articles in WoS and 57,062 in Scopus, totaling 58,552 records. Restricting the search fields to Title, Abstract, and Author Keywords reduced the literature count to 1198 articles. Further limitations on publication years (2000–2025), language (English), and document type (Article or Review) resulted in 552 relevant articles.

Risk of bias was not assessed using standard appraisal tools (e.g., RoB and CASP), as this review does not synthesize effect sizes or primary empirical outcomes. Instead, methodological rigor was ensured through predefined inclusion criteria, metadata completeness checks, systematic relevance screening aligned with the research questions, and transparent thematic coding procedures.

2.3. PRISMA Screening Process

This systematic literature review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [17,22]. A PRISMA flow diagram illustrating the literature identification, screening, eligibility, and inclusion process is presented in Figure 1.

Following Figure 1, the specific operational steps included: identification of sources and records, deduplication, screening of titles and abstracts, and controlling for eligibility. At the identification phase, searches in both the Web of Science and Scopus databases yielded 58,552 records. Following filtering by field, period, language, and document type, the specified number was reduced to 552 articles. Employing Zotero (Version: 7.0.32) literature management software and Python (Anaconda) scripts for dual deduplication based on DOI and title yielded 379 remaining papers. Initial title and abstract screening resulted in excluding papers that lacked abstracts or keywords. At this step, 356 papers were retained.

By the eligibility assessment, screening of titles and abstracts of 356 papers excluded sources clearly irrelevant to the research question, not addressing HR scenarios, or not dedicated to the manufacturing sector. This yielded 347 papers retained as the total study sample.

2.4. Inclusion and Exclusion Criteria

During the literature screening process, this study established inclusion and exclusion criteria by comprehensively referencing the PICO framework and the characteristics of the AI-HRM field [17].

First, the analysis subjects were primarily enterprises, employees, managers, or production units within the manufacturing sector. Second, the technological applications focused on AI, machine learning, deep learning-based intelligent decision-making, and algorithmic management applied to one or multiple HR modules. Third, the research findings demonstrated the impact of AI applications on HR aspects such as employee experience, organizational performance, and sustainability, or highlighted challenges and risks encountered during practical implementation. Finally, research designs encompassed the quantitative, qualitative, and mixed-methods literature.

The exclusion criteria embraced several aspects. Initially, the gray literature, including conference papers, monograph chapters, industry reports, and theses that are not peer-reviewed [28], was excluded. Further, studies with mismatched contexts, such as those examining samples from finance, education, or other sectors, where manufacturing distinctions cannot be properly analyzed, were excluded. Additionally, the literature deviating from the core theme, covering works solely discussing AI technology, HRM theoretical evolution, or digital transformation without substantive exploration of AI-HRM integration, was expelled. Finally, studies with missing information, such as abstracts, author keywords, or full texts unavailable for reliable coding, were excluded.

2.5. Data Extraction and Analysis Methods

This study employs a mixed-methods approach combining quantitative bibliometric analysis with qualitative content analysis to comprehensively address Research Questions RQ1–RQ3.

The final selection of 347 literature entries formed the basis for constructing the comprehensive research map and establishing the foundational dataset for bibliometric and evolutionary analysis. From this complete sample, a further 100 highly relevant and information-rich articles were selected to form a “core in-depth reading sample”. This sample provided support for manual coding related to challenges, convergence pathways, and framework development for Research Questions 1 and 2.

All records meeting the search criteria were exported from Web of Science and Scopus. Key bibliographic information—including article titles, publication years, journals, author keywords, and DOIs—was consolidated by standardizing fields and eliminating duplicates from the database exports, thereby forming a structured dataset for subsequent analysis.

This study clearly distinguishes between the complete sample (n = 347) and the in-depth coding sample (n = 100), each serving distinct analytical purposes. The full sample underpinned bibliometric analysis and thematic evolution mapping (Question 3), whilst the core sample was employed for qualitative coding. This was because Questions 1 (challenge identification) and 2 (integration pathways) required deeply narrative evidence, rather than reliance solely on bibliometric patterns.

The selection of these 100 core articles followed a progressive, criteria-based screening logic rather than methods based on citation counts or quantitative scoring. Four sequential criteria were applied:

(1)

Relevance to a manufacturing context;

(2)

Substantive exploration of AI-related HRM challenges and integration approaches;

(3)

Sufficient theoretical or informational density to support open-ended, axial, and selective coding;

(4)

Achieving a fundamental representational balance across research methodologies, geographical regions, and AI-HR application scenarios.

Within the quantitative research dimension, this study primarily utilized Bibliometrix and VOSviewer for bibliometric analysis [20,29]. Descriptive statistics were conducted on 347 articles concerning annual publication trends, geographical distribution, journal distribution, and disciplinary categories, thereby generating the data for descriptive analysis. High-frequency thematic keywords were identified by setting a minimum occurrence threshold (≥5 times) for author keywords in VOSviewer, and co-occurrence networks were plotted to reveal the structural relationships between research hotspots and themes within AI-HRM studies. Emergent keywords were detected using CiteSpace to identify rapidly rising terms during specific periods, thereby characterizing the temporal dynamics of research themes [25]. An LDA thematic model was also constructed based on abstract texts from the sample literature, identifying representative core themes and analyzing their evolution trends from 2000 to 2025. This generated a thematic latent change map, providing the foundation for the evolutionary trend analysis to answer RQ3 [24].

For qualitative analysis, this study adopted Grounded theory’s open–axial–selective coding approach [30], conducting systematic induction around RQ1 and RQ2. In the coding process for identifying challenges (RQ1), this study selected passages related to “challenges, obstacles, limitations, and risks” from 100 core literature sources. These were subjected to open coding, then synthesized into primary concepts. During the axial coding stage, these were further consolidated to distill six major categories of challenges in AI-HRM integration, as detailed in

Table 2. Major categories of challenge within AI-HRM.

Coding	Core Concept	No of Documents = 100	High-Frequency Association with HR Modules
C1 Technical & data infrastructure	Insufficient foundational conditions—such as algorithmic performance, data quality and integration, IT infrastructure, and security—are required for deploying AI systems in HR scenarios, leading to unstable model implementation or biased outcomes.	48	Training and development, employee safety and health, employee relations and engagement
C2 Organizational & strategic alignment	AI-HRM initiatives lack alignment with corporate strategy, HR strategy, and business processes. Unclear organizational structures and responsibility boundaries, coupled with inadequate cross-departmental collaboration and resource allocation, often result in AI projects operating in isolation.	37	Training and development, performance management, recruitment and staffing
C3 Skills & capability gaps	Managers and frontline staff commonly lack data literacy and AI proficiency, while HR teams demonstrate insufficient capabilities in algorithmic understanding, project management, and change facilitation. This results in “technology being available but personnel unable to utilize it” and “systems lying idle or being used in simplified ways”.	67	Training and development, employee safety and health, performance management
C4 Trust & ethical concerns	Employees and managers harbor concerns regarding fairness, transparency, and privacy protection in sensitive areas such as recruitment, performance evaluation, and remuneration. Fears of being “replaced by machines” or subjected to unfair treatment undermine trust in and willingness to adopt the systems.	52	Training and development, employee safety and health, performance management
C5 Policy & institutional environment	External regulations and standards lag or remain uncertain, while internal governance systems are incomplete. AI-HRM practices lack clear boundaries regarding data compliance, labor legal liabilities, and algorithmic accountability, leaving enterprises exposed to compliance risks and regulatory vacuums.	8	Training and development, employee safety and health, employee relations and engagement
C6 Sustainability & green transition	Insufficient mechanisms exist to align AI-HRM with green manufacturing, environmental performance, and social responsibility objectives. Green capability development and employee engagement remain inadequate, with AI predominantly serving efficiency and cost reduction rather than genuinely supporting sustainable and green transition.	12	Training and development, performance management, employee relations and engagement

.

Regarding integration pathways (RQ2), this study synthesized practical approaches, management methods, and theoretical models from the literature to define AI-HRM integration pathways. These were categorized through interdisciplinary temporal analysis. Ultimately, six integration pathways were established to construct a Challenge–Approach Matching Matrix, with specific classifications presented in Table 2 and

Table 3. Six categories of strategies and theoretical frameworks for AI-HRM integration.

Approach Category	Core Methods/Solutions	Typical Theoretical Foundations	Number of Supporting Documents
A1: Algorithmic & XAI Solutions	Explainable Artificial Intelligence (XAI) Differential privacy & federated learning Algorithmic matching for recruitment and performance Bias detection and mitigation	Algorithm transparency frameworks; Organizational Information Processing Theory (OIPT)	51
A2: Human–AI Collaboration	AI-HR co-decision interfaces Chatbot-assisted HR processes Human–AI interaction in feedback and development	Human–AI Teaming (HAT); Socio-Technical Systems Theory (STST)	28
A3: Organizational Theories	Technology adoption mechanisms Digital HRM governance Strategic integration of AI within HR systems	TAM; Resource-Based View (RBV); Self-Determination Theory (SDT); Social Exchange Theory (SET); OIPT	37
A4: Change & Training Interventions	Reskilling and upskilling programs AI literacy and competence development Leadership-driven change management	Organizational Change Theory; Learning Theory	64
A5: Ethics and Governance Mechanisms	Algorithm audits Fairness and bias mitigation Transparency and risk management protocols	Responsible AI (RAI) models; AI Ethics Frameworks	42
A6: Policy & Standardization	ISO 30414 human capital reporting Regulatory compliance (e.g., EU AI Act) Industry/sector guidelines for HR analytics	Institutional Theory; Compliance Theory	22

.

To ensure coding consistency and traceability, a progressive self-check approach was adopted. Open and axial coding were performed against predefined categories, with the results of coding continuously refined through revisiting source texts, cross-referencing category definitions, and multiple rounds of verification. Both challenges and approaches classification frameworks maintained high internal consistency, safeguarding subsequent analysis reliability [30].

2.6. Descriptive Statistics of the Literature Sample

A total of 347 articles were ultimately obtained for the overall sample. Statistical analysis of these 347 articles yielded descriptive statistics for the overall sample, including publication distribution by year, document type, disciplinary category, and research field (

Table 4. Descriptive statistics of the reviewed literature sample (n = 347).

Dimension	Category	Frequency N	Percentage %
Year	2000–2010	7	2
	2011–2015	10	2.9
	2016–2020	49	14.1
	2021–2025	281	81
document type	article	308	88.8
	review	39	11.2
categories (top 5)	management	46	13.3
	business	21	6.1
	environmental sciences	19	5.5
	engineering, electrical & electronic	19	5.5
	computer science, information systems	18	5.2
research areas (top 5)	business & economics	66	19
	engineering	65	18.7
	computer science	41	11.8
	environmental sciences & ecology	21	6.1
	science & technology—other topics	19	5.5

).

Statistical findings indicate that publications from 2021 to 2025 constituted over 80% of the total, reflecting an explosive growth in research on AI-HRM integration in recent years. Studies predominantly spanned management, business research, engineering technology, and computer science, highlighting the topic’s pronounced interdisciplinary nature [2,23]. This literature analysis not only underpinned the overall structure of Chapter Three but also provided crucial foundations for subsequent thematic evolution and the construction of a Three-Tier Analytical Framework (Macro–meso–micro).

3. Results

This section reports empirical findings derived from bibliometric analysis and qualitative coding, without theoretical interpretation. It presents factual patterns concerning publication trends, geographical distribution, thematic structures, and challenges in coding—such as path frequencies. The interpretation of these findings and their theoretical implications will be elaborated in the subsequent “Discussion” section.

3.1. Descriptive Analysis

As the first part of the examination, a multidimensional overview of AI-HRM research in manufacturing from 2000 to 2025 is provided, based on a sample of 347 documents. Annual publication trends, geographical distribution, journal representation, and keyword co-occurrence networks were investigated in detail. All quantitative results were generated using Excel, Bibliometrix, and VOSviewer, with robustness validated through cross-referencing against the core literature data [31]. Annual publication trends for the first quarter of the XXI century were traced in Figure 2.

Following Figure 2, analysis of annual publication volumes revealed a phased evolution in AI-HRM research within manufacturing, synchronized with advancements in artificial intelligence technology, Industry 4.0, and the digital transformation of manufacturing [32,33]. This evolution can be broadly divided into four periods, as described below.

The period 2000–2010 constituted an embryonic phase, with only seven relevant publications. Content centered on exploring the feasibility of applying expert systems and algorithms within human resource modules. A systematic theoretical framework had yet to emerge during this stage. The subsequent period from 2011 to 2015 represented a phase of gradual accumulation. The diffusion of the Industry 4.0 concept spurred growth in related research, though the overall scale remained modest. That period’s research primarily centered on AI-HR process alignment and conceptual analysis [34]. The phase from 2016 to 2020 constituted an accelerated development phase. Accompanied by advancements in big data and cloud computing, empirical research within manufacturing contexts on AI has been conducted. Particularly, technology adoption models based on the Technology Acceptance Model (TAM) and Technology, Organization, and Environments (TOE) frameworks provided theoretical foundations for applications such as intelligent recruitment and training within the HR domain [35]. The period from 2021 to 2025 represents a concentrated surge, accounting for over 80% of the total sample volume. The observed reduction in the number of publications in 2025 is primarily attributable to delays in database indexing and incomplete coverage of publication years during data collection. The research themes of this phase encompassed mostly trust, algorithmic bias, and human–machine collaboration [36].

Based on cross-validation of 100 core documents through manual coding, results indicate that emergent terms such as “Trust”, “Responsible AI”, and “Skills” also exhibited a significant rise after 2021, aligning closely with quantitative analysis. Examination of the national and regional distribution of publications dedicated to AI-HRM in manufacturing was the next step in the present study (Figure 3).

Analysis of 347 sample documents in Bibliometrix allowed us to identify a dual-dominant pattern in publication origins: “manufacturing powers + technological powers.” This finding is corroborated by existing reviews [32].

The geographical distribution uncovered that China, the United States, and Germany collectively contributed a substantial proportion of publications. Chinese research primarily focused on security detection and skill enhancement, U.S. studies emphasized algorithmic transparency, fairness, and trust [37], whereas German research accentuated human–machine collaboration within the context of Industry 4.0.

Countries such as the United Kingdom, Italy, South Korea, and India followed the aforementioned states with secondary publication volumes, where research predominantly addressed industrial applications and case studies within electronics, automotive, and equipment manufacturing sectors. The remaining studies concentrated on challenges, including inadequate data infrastructure, skills mismatches, and difficulties in technology dissemination [34]. This geographical concentration indicates that the observed frequency of challenges and the sequence of integration pathways are at least partly determined by the problem frameworks and solutions prevailing in the major manufacturing economies.

Analysis of article distribution identified that most of the works on the examined topic were allocated across over 120 journals, exhibiting distinct interdisciplinary characteristics. Bibliometrix statistics indicated that Sustainability (13 articles) ranked highest in publication volume, with research primarily centered on digital manufacturing, green HRM, and AI governance—reflecting the convergence of sustainability research and HR digitalization [36] (Figure 4).

Following Figure 4, the second-most prolific journal was the International Journal of Advanced Manufacturing Technology (11 articles), primarily addressing applications of intelligent manufacturing and machine learning within HR contexts, such as skill monitoring and safety management [32].

IEEE Access (9 articles) and the International Journal of Production Research (8 articles) predominantly publish research within the domains of information systems and production management. The former emphasizes the application of technologies such as algorithms and visual recognition within HR modules, while the latter focuses on research directions including human–machine collaboration, personnel scheduling, and performance forecasting within production systems and organizational processes [37].

Overall, research on AI-HRM exhibits a parallel publication landscape across three domains: engineering technology, management science, and information systems. Engineering journals focus on the implementation of algorithms and system applications, management journals emphasize organizational adoption and sustainable development, and information journals concentrate on technology governance and data security. This cross-disciplinary publication pattern reveals the multidimensional nature of AI-HRM research within manufacturing and provides a foundational disciplinary reference for subsequent thematic studies.

At the next step of the conducted descriptive analysis, VOSviewer was utilized to analyze author keywords that yielded a co-occurrence network (Figure 5), illustrating the overall structure of AI-HRM research themes within manufacturing. The analysis revealed a multi-cluster distribution pattern centered on “Industry 4.0-AI-HRM”. Further, the defined network structure was cross-validated through qualitative coding of 100 core publications [31].

Based on the co-occurrence network diagram, “Industry 4.0” occupies the geometric center, functioning as a pivotal hub connecting nodes representing the technological dimension (machine learning), manufacturing context (smart manufacturing), and management dimension (HRM and performance management) [33]. Three thematic clusters have formed around this core: technological foundations, manufacturing systems, and HR applications.

Keywords such as “Machine learning”, “Big data”, and “Internet of things” within the technological foundations layer reflect the foundational context of AI-HRM research. Within the manufacturing scenarios, “Smart manufacturing and industrial Internet of things” alongside the “Industry 5.0” cluster served as primary connecting points, illustrating the manufacturing sector’s transition from Industry 4.0 towards the more advanced human–machine collaborative paradigm of Industry 5.0. The detected keywords in the HR application layer—“HRM”, “Automation”, “Digital transformation”, and “Competencies”—indicated that AI implementation was concentrated in process automation, employee competency modeling, and organizational digital transformation.

Sustainability and governance keywords such as “Sustainability”, “Green innovation”, and “Blockchain” exhibited fewer connections in the co-occurrence network diagram, indicating weaker associations with core themes. This finding aligned with the rising prominence of topics like “Responsible AI” and “Ethics” in the core literature [36].

The co-occurrence map exhibited distinct thematic segmentation, with weak connections between technology, manufacturing, and governance clusters. It indicated that existing research had yet to establish an integrated theoretical framework bridging technology, organization, and institutions. The specified fragmentation provided research grounds for identifying challenges in the subsequent section. The conducted quantitative analysis of 347 articles supplemented by qualitative validation of 100 core publications allowed us to reveal certain characteristics of AI-HRM research in manufacturing: temporal concentration, uneven geographical distribution, journal coverage, and thematic clustering. It is worth noting that the fragmentation between the themes of technology, organization, and ethics engendered a research gap, lacking a theoretical framework capable of integrating these perspectives through a multi-level lens.

3.2. Thematic Findings

Based on the aforementioned descriptive findings, this subsection reports the outcomes of manual coding across 100 core studies. The analysis identified key challenge categories, integration pathways, and their co-occurrence patterns within the human resources module (Appendix B). The report maintains a descriptive style, focusing on the content uncovered rather than the reasons behind these patterns.

RQ1. Key Challenges in AI-HRM Integration within Manufacturing

To answer RQ1, open-ended and axial coding were applied to segments containing “Challenge”, “Barrier”, and “Limitation” across 100 core articles. A threshold of at least five core articles mentioning each term was applied, ultimately identifying six challenge categories (C1–C6). These were ranked by frequency to construct a classification table of primary AI-HRM challenges. Similarly, scholars in their systematic review of AI-HRM relationships [38,39] and bibliometric knowledge graph study [38,40] indicated that existing research generally viewed technological foundations, organizational capabilities, employee skills, and ethical risks as barriers to the integration process, critical to comply with the international legislation on AI usage [39], rather than solely technical issues.

Following Table 2, among 100 core publications, the six challenge categories were referenced in descending order of frequency. Skills and capability gaps were the most widespread challenge, whereas policy and institutional environment were the least common among 100 examined publications.

After further cross-coding of these six challenge categories with seven human resource modules (HR1–HR7), it was found that their distribution along different functional areas fluctuated. The definition of these seven human resource modules is as follows: HR1 = recruitment and staffing; HR2 = training and development; HR3 = performance management; HR4 = salary and reward; HR5 = safety and welfare of employees; HR6 = employee relationship and opinion expression; and HR7 = strategic human resource management and talent management.

The purpose of adopting this classification method is to establish a clear analysis mapping relationship between the challenges related to AI and various functions in the field of human resources. It was not aimed to comprehensively cover all human resource management functions, but to focus on those that are frequently discussed and observable in the reviewed literature. The labels HR1–HR7 were applied only for analysis reference, not as a normative or sequential human resource process model.

As illustrated in Figure 6, training and development (HR2) exhibits a high correlation with virtually all challenges, representing the module most heavily impacted. Employee safety and well-being (HR5) and performance management (HR3) followed in prominence, whilst compensation and rewards and strategic HRM and talent management demonstrated comparatively lower challenge associations. Therefore, it appears significantly more frequently in the training and development and performance management modules [18]. The sequential explanation of the revealed challenge categories was presented further in the text.

As shown in Figure 6, HR2 shows a high co-occurrence frequency in almost all challenge categories, indicating that it is the HR module with the widest impact on AI-related challenges. HR5 and HR3 also showed a strong correlation in multiple challenge dimensions. In contrast, the relationship between HR4, strategic HRM, and HR7 and specific challenge categories is more selective and situational. In general, the heat map shows the different distribution of challenge intensity between different HR modules, in which the training-oriented function shows an obvious concentration.

Challenge category 3 “Skills and capability gaps” stood out as the most prominent challenge, being directly personnel-related. The existing literature consistently highlighted significant gaps among frontline staff in understanding data, intelligent operating systems, and algorithms [41]. This pattern has been documented in multiple research reports, which highlight instances where system availability does not align with personnel capabilities [18,42]. This challenge was particularly pronounced in training and development, performance management, and safety and health modules. Similarly, mid-level managers often struggled to utilize models effectively, hindering their potential as decision-support tools. Similarly, European manufacturing research revealed that within Industry 4.0 and 5.0 contexts, enterprises faced talent shortages in digital skills, interdisciplinary capabilities, and human–machine collaboration literacy. Multiple studies indicate that a persistent gap exists between an organization’s overall capabilities and the level of technological advancement [43,44].

Over half of the literature highlighted insufficient trust among employees and managers (C4) regarding AI applications in sensitive areas such as recruitment, performance evaluation, and safety monitoring. Concerns primarily centered on algorithmic fairness, privacy protection, and anxieties about job displacement. Existing research indicated that when employees could not comprehend how AI made hiring and promotion decisions, they were more likely to attribute negative outcomes to “algorithmic bias”, thereby amplifying distrust [18]. Particularly in the absence of explanation and appeal mechanisms, employee distrust could escalate into organizational-level resistance, potentially drawing the attention of trade unions and regulatory bodies [42].

Nearly half of the literature identified technological integration and high-quality data (C1) as prerequisites for AI-HRM implementation. The manufacturing sector featured numerous heterogeneous systems, with inconsistent data standards across HR, production, and supply chain domains, alongside closed system interfaces. This challenge hindered AI models from accessing comprehensive, high-quality training data [19,45]. Some studies further indicated that inadequate recording of critical data, such as safety incidents and non-compliance behaviors, impeded predictive models’ ability to accurately identify risks, with significant discrepancies existing between models operating in simulation environments and real workshop conditions [45].

Approximately one-third of the analyzed literature indicated that AI-HRM projects are often misaligned with corporate strategy, digital roadmaps, and HR strategy (C2). Typical issues included: projects being led by IT or other business units, marginalizing the HR role, inconsistencies between departmental objectives and performance metrics, and unclear budget allocation and division of responsibilities [42]. Multiple studies describe artificial intelligence human resource management initiatives as primarily positioned as pilot projects or compliance initiatives, rather than long-term strategic capability development programs. While some organizations frequently invoked slogans like “intelligent HR” and “data-driven”, these principles remained largely absent in practical job design, performance management, and organizational structures.

Although only 8% of the current literature addressed the challenge of policy and institutional environment (C5), recent research indicated that lagging external regulations and internal governance systems became significant constraints on the integration of AI-HRM in manufacturing [2,45]. These challenges included EU regulations on high-risk AI systems, which placed AI applications in HR under heightened scrutiny. Corporate methodologies had yet to establish robust management systems for algorithmic documentation, impact assessments, and internal accountability, leading to unclear liability allocation when issues arise. For resource-constrained SMEs, compliance costs and policy uncertainties significantly dampened their willingness to adopt AI-HRM [18].

Though the literature remains sparse, it is gradually expanding. Examining the integration of AI-HRM through sustainability and Industry 5.0 lenses revealed a marked disconnect between green manufacturing and ESG objectives (Challenge 6). Current applications primarily targeted efficiency gains and cost savings, lacking mechanisms for green skills training and AI assessment integration. Few studies explicitly discussed incorporating AI-HRM integration metrics into ESG disclosures [2,19,46,47].

The coding results of this study indicated that the challenges of AI-HRM integration in manufacturing were not solely technical issues, but rather systemic problems involving the interplay of technology, organization, institutions, and personnel. In the cross-sectoral comprehensive assessment of artificial intelligence and human resource management, a similar multi-dimensional pattern of obstacles has also been identified [18], partially addressing the limitations of existing reviews that adopt singular perspectives, such as the “technical dimension” or “ethical dimension”. Technical, organizational, and institutional challenges are predominantly “coupled problems” that overlap and mutually reinforce each other across most studies, rather than easily separable individual issues. Earlier studies indicate that single-dimensional technical optimization or training interventions are insufficient to address multiple coexisting barriers [48].

RQ2. Integration Pathways, Strategies, and Theoretical Frameworks

To answer RQ2, the present study synthesized solutions from 100 core articles by identifying typical theoretical foundations and primary challenges addressed, thereby formulating six categories of strategies and theoretical frameworks for AI-HRM integration (Table 3). While answering these questions, the coauthors relied on AI-HRM systematic reviews categorizing approaches through technological pathways, organizational mechanisms, and outcome variables, further underscoring the importance of examining AI-HRM integration from technological, organizational, and personnel perspectives [14,15,49].

Based on the coding results from the close reading of 100 core articles, the AI-HRM integration pathways exhibited distinct stratification in terms of literature support. Pathway A4 (64 papers) and Pathway A1 (51 papers) were defined as the two most frequently occurring approaches, forming relatively mature and stable research threads. Path A5 (42 papers) and Path A3 (37 papers) exhibited moderate intensity, demonstrating persistent yet context-dependent trends. In contrast, Path A2 (28 papers) and Path A6 (22 papers) received limited support within the sample, primarily stemming from scattered case studies or conceptual research, and remain largely exploratory in nature. It is noteworthy that this stratification aligns with the thematic focus of the principal contributing countries reported in Section 3.1, where differing national contexts portend distinct problem-solving approaches (e.g., skills enhancement, trust-building, and human–machine collaboration).

Based on this frequency-driven stratification, the present study systematically examines the operational mechanisms, representative technologies, organizational characteristics, and corresponding core challenges for each of the six convergence pathways.

A1 Algorithmic and XAI

Solutions enhance the transparency, fairness, and security of AI-HRM systems through technologies such as explainable AI, differential privacy, and bias detection, thereby addressing challenges C1 and C4. Some studies demonstrated that integrating computer vision with explainable modules in industrial safety inspections enabled frontline managers to rapidly comprehend decision-making rationale, thereby reducing ‘black-box anxiety’ [19,45]. Such solutions focused on embedding ‘explainability’ and ‘privacy protection’ at the algorithmic level to foster employee trust in the system.

A2 Human–AI Collaboration

Addressing challenges, C3 and C4 involved utilizing AI-HRM decision support and chatbot human–machine collaboration functions, positioning AI as a decision-support partner rather than a decision-replacement entity [42]. The research indicated that when systems explicitly stated final decision-making, authorities resided with humans and provided interfaces for employees to interact with and refine AI outputs; staff were more likely to perceive AI as a tool rather than a threat, with significant reductions in resistance. Recent studies demonstrated that redefining human–machine task boundaries while preserving human dominance in ethical judgements and complex task analysis enhanced employee engagement and decision quality [50,51]. Within manufacturing contexts, research has explored integrating machine learning, the Internet of Things, and human resource planning. By optimizing scheduling and capability allocation, this approach achieved synergy between smart manufacturing and human resource management [52,53].

A3 Organizational Theories

Theories such as the Technology Acceptance Model (TAM), Resource-Based View (RBV), and Organizational Information Processing Theory (OIPT) were employed to develop adoption mechanisms for AI-HRM, primarily addressing C2 and C3 challenges [42]. This research emphasized that AI-HRM should become an integral part of organizational capability rather than merely a short-term tool. Sustainable competitive advantage was only achieved when AI was genuinely embedded within corporate strategy and human resource development processes [54].

A4 Change and Training Interventions

Addressing C3 challenges involved cultivating employees’ AI literacy, reshaping job skills, and enhancing leadership capabilities [18,42]. Research indicated that sustained training and job redesign enhanced AI-HRM integration more effectively than mere system upgrades. The on-the-job learning nature of manufacturing ensured the A4 pathway transformed abstract algorithms into practical knowledge, yielding the strongest alignment with C3 in the matching matrix. Existing research indicated that successful AI project implementation relied heavily on sustained training and skill development, alongside organizational change management mechanisms across all levels, rather than solely on technological upgrades [13,55].

A5 Ethics and Governance Mechanisms

Addressing C4 and C5 challenges involved institutionalizing trust and fairness through algorithmic audits, ethics committees, and employee right-to-explanation frameworks [2,42]. This research, grounded in Responsible AI frameworks, emphasized that AI-HRM projects should establish ethical principles, oversight, and accountability mechanisms from inception. Path A1 reduced information asymmetry and enhanced comprehensibility but struggled to resolve performance losses and misattribution of responsibility. Path A5 provided accountability and grievance channels but incurred higher implementation costs. Complementary to each other, they formed a dual-path approach of “technical transparency + institutional safeguards,” offering a more robust response to challenge C4.

A6 Policy and Standardization

The present approach addresses challenges C5 and C6 by adopting external institutional frameworks and industry standards—such as human capital disclosure frameworks and international AI governance principles—to bridge labor laws, data protection regulations, and other legal frameworks [2,45]. Although empirical evidence remains limited, existing research indicates that linking AI-HRM with ESG disclosure, ergonomics, and safety standards helped establish clear compliance boundaries and accountability mechanisms [19]. Recent research on small and medium-sized manufacturing enterprises further indicated that integrating AI adoption with green human resource management, amplified through organizational culture, constituted a key mechanism for driving sustainable performance [56].

To avoid a mere enumeration, this study further proposes the designed Challenge–Approach Matching Matrix that quantifies the co-occurrence frequency of C1–C6 with A1–A6, thereby indicating high matching (✮✮✮), medium matching (✮✮), and low matching (✮) to illustrate their interrelationships (Table 5).

Following Table 5, segmenting the challenge-path matrix into strong, medium, and weak gradients allowed us to reveal distinct structural coupling patterns.

First, C3 and A4 demonstrated the strongest overall alignment, reflecting the ‘learning by doing’ characteristic of manufacturing contexts. Abstract algorithms struggled to directly enhance self-efficacy; however, Path A4 contextualized technical knowledge through on-the-job training, role redesign, and peer learning. This approach propels employees from merely knowing how to use tools to genuinely wanting to use them, thereby becoming the core pathway for addressing skill gaps [18].

Second, C4 exhibited high concentration at both A1 and A5 levels, illustrating typical technical–institutional complementarity. Path A1 enhanced transparency by mitigating information asymmetry, yet struggled to resolve performance risks and misjudgment accountability; Path A5 provided audit, appeal, and accountability mechanisms, albeit with relatively high implementation costs. The combination of these two pathways formed a dual-track approach to addressing trust challenges, further supplemented by Pathway A2’s complementary support in system functional collaboration and human final decision-making authority [42,50,51].

Finally, limited evidence exists for challenges C5 and C6 on Path A6, indicating that AI-HRM integration remains nascent in institutional and green dimensions. Earlier perspectives in the green HRM and Industry 5.0 literature provided similar inferences [2,19].

Overall, singular challenges typically correspond to dominant pathways (e.g., C3 → A4), whereas compound challenges rely more heavily on combined pathways (e.g., C4 → A1 + A5, supplemented by A2 collaboration). This statement suggests manufacturing enterprises should prioritize investment in three highly compatible pathways—training, human–machine collaboration, and ethical governance—while avoiding indiscriminate technological upgrades [14,15,49].

Verification of RQ2 not only identified six reusable convergence pathways but also revealed challenge–pathway matching patterns, providing crucial foundations for subsequent model development and scenario-based applications.

RQ3. Research Evolution Trends and Macro-Meso-Micro Transfers

To address RQ3, the given study employed keyword co-occurrence analysis, emergent term detection, and LDA thematic latent variable analysis on 347 sample documents that yielded an evolutionary trajectory of AI-HRM research in manufacturing from 2000 to 2025 [57]. The thematic evolution map is depicted in Figure 7.

To ensure the interpretability and robustness of LDA-based thematic evolution analysis, this study does not rely on topic modeling as an independent classification tool. Instead, thematic identification and phase delineation were validated through triangulation. Specifically, LDA results were cross-checked against keyword co-occurrence networks (VOSviewer), thematic emergence detection (CiteSpace), and manual coding from in-depth reading of 100 representative articles. Thus, thematic labels and phase boundaries were determined based on the convergence of quantitative thematic distribution and qualitative thematic consistency, rather than solely on statistical consistency scores.

Regarding model stability, alternative theme numbering settings were explored during modeling, and the identified three-phase evolutionary structure maintained substantial consistency across various configurations. While the relative prominence of individual themes varied with parameter adjustments, the overall temporal shift from technology-driven automation to data-driven human resource optimization, and finally to trust, governance, and people-centered integration proved robust. These findings suggest that the identified evolutionary phases reflect stable thematic transitions within the literature rather than artificial constructs resulting from specific model parameter choices.

The thematic map has signified the abrupt growth of academic interest towards artificial intelligence, human resource management, and Industry 4.0 since 2020. The keywords burst timeline for the first quarter of the XXI century is given in Figure 8.

Synthesizing quantitative findings and content analysis, the research broadly unfolds across three distinct phases, each characterized by markedly different focal points and driving forces. This progression largely aligns with the temporal evolution of the cross-industry AI-HRM literature [2,18]. Latent topic modeling is visualized in Figure 9.

Following Figure 9, trust, ethics, and algorithmic governance constituted the key theme in the examined field. The more substantiated explanation of the identified trends in Figure 7, Figure 8 and Figure 9 is presented in the text below.

Phase One (2000–2015): Technology-driven and the dawn of automation

Publications of this period were relatively scarce, with keywords primarily centering on “automation” and “expert systems”. This phase focused chiefly on AI applications within isolated HR tasks such as payroll processing and rostering. Research at this time emphasized technological feasibility and cost savings, with little consideration for organizational or employee perspectives. While laying the groundwork for subsequent expansion, this phase also sowed the seeds of neglecting human-centered factors. A small number of studies began exploring the application of AI methods, such as neural networks, to critical HR forecasting areas like employee attrition prediction, establishing the methodological foundation for later “HR analytics” research [58].

Phase Two (2016–2020): Data Analysis and HR Module Optimization Period

During this period, the proliferation of Industry 4.0 and big data technologies shifted key terms towards “HR analytics, big data, Industry 4.0,” with research focus moving to core modules such as recruitment optimization, training evaluation, and performance prediction [2]. Concurrently, theories such as TAM, OIPT, and RBV gained widespread application, emphasizing the impact of data capabilities and organizational resources on AI-HRM integration outcomes. However, most research remained centered on individual HR modules, with limited attention to systematic studies and ethical considerations.

Phase Three (2021–2025): Trust Mechanisms, Ethical Governance, and System Integration

This phase accounts for most of the published literature. Emergent keywords and LDA analysis detected rapid thematic shifts towards “trust, fairness, governance, behavioral responses, and sustainability” [18,19]. Research expanded beyond merely discussing AI-enabled management to contemplating how to manage for people. On one hand, it highlighted the dual challenges of C4 and C5 in RQ1 through trustworthy AI and ethical governance; on the other hand, it explored the convergence of AI-HRM with production systems and green transition through coordinated action at macro-policy, meso-organizational, and micro-individual levels. This finding relates to the human-centered Industry 5.0 agenda [2,19,59].

Transition between these three phases stems not merely from technological maturation but from dual impetus: institutional disruption coupled with practical dilemmas. Firstly, as algorithmic management and data-driven approaches proliferate in personnel decision-making, academia has begun exposing algorithmic risks concerning fairness and transparency, advocating for ethical constraints within HR practices [60,61]. Concurrently, the widespread adoption of remote working during the COVID-19 pandemic accelerated the deployment of electronic monitoring systems, exacerbating concerns regarding employee privacy and psychological burdens [62,63]. Subsequently, the discourse on AI-HRM convergence has shifted from early focuses on technical feasibility and efficiency prioritization towards a socio-technical perspective centered on responsible AI, human-centered governance, and institutional design [2,19,59].

To further characterize this evolutionary “hierarchical migration”, this study constructed a Multi-Level Embedded Framework—based on close reading—and conducted comparative analyses of representative research across these dimensions [64,65].

Building upon the three-tier comparative framework presented in Table 6, the following sections elaborate on key research themes and evidence gaps across macro, meso, and micro levels.

Macro-level (policy and institutional) research focuses on national AI strategies, data governance regulations, and industry standards, highlighting how boundaries for AI-HRM in manufacturing are delineated at this tier—particularly concerning constraints on AI system design within occupational safety and health frameworks [45]. However, quantitative evidence remains scarce on topics such as the inhibitory effects of policy adaptability and compliance costs on SME adoption behavior in regions like Southeast Asia and Africa [18].

Research at the meso level (organizations and processes) is most concentrated, focusing on decision support, process re-engineering, and governance mechanisms within enterprises [42]. Nevertheless, the integration of the three dimensions—AI, HR, and production systems—remains insufficiently validated in practice, particularly lacking longitudinal studies across different developmental stages within the same enterprise.

Micro-level research (employees and teams) examines the influence of trust, risk perception, and AI anxiety on adoption willingness [18,45,66]. Nevertheless, research centered on frontline blue-collar workers is still scarce, with a significant proportion of studies still drawing from knowledge workers or service industry personnel. Longitudinal studies and cross-scenario comparative research are also relatively fragmentary, a situation particularly pronounced within manufacturing contexts.

Overall, this evolutionary analysis indicated that AI-HRM research in manufacturing had progressively shifted from early technological tool perspectives towards a human-centered, system-integrated paradigm centered on human–machine collaboration and responsible AI. It complies with the human-centered research trends observed in Green HRM, Industry 5.0, and occupational safety [2,19,45]. The thematic evolution from “technological tools → data-driven approaches → human-centered governance” was corroborated by high alignment with C3 × A4 and C4 × (A1 + A5), forming a cross-level linkage trajectory of “macro policy → meso governance → micro trust”. The results gained provided direct empirical support for subsequent discussion (Table 5).

Table 5. Challenge–approach matching matrix.

Challenge Approach	A1	A2	A3	A4	A5	A6
C1	33	19	17	27	28	12
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C2	24	9	20	25	14	10
	(✮✮✮)	(✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)	(✮✮)
C3	29	22	25	54	25	13
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C4	23	17	24	35	31	10
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C5	5	2	4	5	6	5
	(✮)	(✮)	(✮)	(✮)	(✮✮)	(✮)
C6	5	3	5	8	2	1
	(✮)	(✮)	(✮)	(✮✮)	(✮)	(✮)

For interpretability, the star ratings in Table 5 are defined by co-occurrence frequencies across the 100 core articles (✮✮✮: ≥30; ✮✮: 15–29; ✮: <15).

Table 5. Challenge–approach matching matrix.

Challenge Approach	A1	A2	A3	A4	A5	A6
C1	33	19	17	27	28	12
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C2	24	9	20	25	14	10
	(✮✮✮)	(✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)	(✮✮)
C3	29	22	25	54	25	13
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C4	23	17	24	35	31	10
	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮✮)	(✮✮)
C5	5	2	4	5	6	5
	(✮)	(✮)	(✮)	(✮)	(✮✮)	(✮)
C6	5	3	5	8	2	1
	(✮)	(✮)	(✮)	(✮✮)	(✮)	(✮)

For interpretability, the star ratings in Table 5 are defined by co-occurrence frequencies across the 100 core articles (✮✮✮: ≥30; ✮✮: 15–29; ✮: <15).

Figure 10 illustrates the distribution of six primary challenge categories across three developmental phases. Challenges from the early period (2000–2015) are relatively sparse, whilst those pertaining to skills and trust have become increasingly prominent since 2021. This indicates a marked shift in focus towards people-centered and governance-oriented approaches.

Table 6. Comparison of studies under the three-level framework.

Level	Core Focus	Representative Study	Key Insights
	Challenges & HR Module	(Derived from 100 in-Depth Studies)
Macro	C3, C4, C1 HR: Accent on training and safety	[3,15]	Policies significantly drive AI-HR pilot schemes yet lack long-term designs for employee trust and capacity building; empirical evidence is scarce across the policy–organization–employee chain.
Meso	C3, C1, C2 HR: Focus on training, performance, safety	[2,7,8]	The organizational layer represents the core bottleneck for AI implementation; cross-departmental collaboration, governance mechanisms, and digital capabilities determine AI effectiveness; research on manufacturing sector variations remains insufficient.
Micro	C3, C4 HR: Focus on safety, performance, and training	[9,11]	Employee trust and perceived fairness are fundamental to AI system success; research on frontline workers remains insufficient; longitudinal studies tracking shifts in attitudes and skills are deficit.

Table 6. Comparison of studies under the three-level framework.

Level	Core Focus	Representative Study	Key Insights
	Challenges & HR Module	(Derived from 100 in-Depth Studies)
Macro	C3, C4, C1 HR: Accent on training and safety	[3,15]	Policies significantly drive AI-HR pilot schemes yet lack long-term designs for employee trust and capacity building; empirical evidence is scarce across the policy–organization–employee chain.
Meso	C3, C1, C2 HR: Focus on training, performance, safety	[2,7,8]	The organizational layer represents the core bottleneck for AI implementation; cross-departmental collaboration, governance mechanisms, and digital capabilities determine AI effectiveness; research on manufacturing sector variations remains insufficient.
Micro	C3, C4 HR: Focus on safety, performance, and training	[9,11]	Employee trust and perceived fairness are fundamental to AI system success; research on frontline workers remains insufficient; longitudinal studies tracking shifts in attitudes and skills are deficit.

4. Discussion

This study reveals that AI-HR integration within manufacturing is not a linear or phased process, but rather a dynamic interaction mechanism spanning macro, meso, and micro levels. The challenges exposed by Research Question 1, the pathways summarized by Question 2, and the evolutionary patterns revealed by Question 3 collectively demonstrate that technological deployment alone is insufficient. Success hinges upon how institutional environments, organizational governance structures, and employee cognitive behaviors become mutually embedded, continuously reconfiguring over time. To capture this dynamic process, this study proposes the Multi-level Embedded Framework (MLEF) illustrated in Figure 11.

Importantly, MLEF does not posit a direct causal relationship from the macro to the micro level; rather, macro-level institutions are primarily embedded within organizational structures, with only limited and indirect signaling effects observable at the employee level.

The MLEF transforms AI-HR integration into a bidirectional interaction and feedback loop system, wherein each level functions both as an influence source and a feedback receiver. The framework’s embedded mechanism structurally couples otherwise dispersed challenges by institutionalizing, translating, and internalizing AI-HRM principles across levels.

In high-pressure manufacturing environments, AI-HRM typically delivers immediate gains in efficiency, cost control, and operational stability. However, efficiency demands may directly conflict with ethical governance requirements, particularly when AI is applied to safety monitoring, performance evaluation, and employee discipline. Crucially, MLEF does not assume efficiency and ethical compliance are inherently aligned. Instead, it conceptualizes this tension as a recurring cross-level trade-off that must be managed through embedding and feedback loops. At the macro level, regulations and norms establish non-negotiable boundaries for data usage and algorithmic accountability; at the meso level, organizations translate efficiency goals into controlled socio-technical arrangements; and at the micro level, employees’ perceptions of legitimacy and trust determine whether efficiency-driven systems are accepted, questioned, or resisted. These micro-level responses feed back into organizational redesign and, over time, inform the foundational basis for institutional learning and improvement.

The macro-level institutional environment also shapes the boundary conditions for AI-HR integration within manufacturing. Regulatory initiatives such as the EU’s Artificial Intelligence Act and data protection frameworks indicate a growing emphasis on accountability, transparency, and risk management in AI-enabled human resource practices (e.g., [39,67]). Within this study, these regulations are treated as part of the broader institutional context rather than subjects for legal analysis.

Geographical concentration and boundary conditions

The identified challenge–pathway patterns (Figure 3) should be interpreted in light of the uneven geographical distribution of the reviewed literature, wherein China, the United States, and Germany account for a substantial share of publications. This dominance may amplify prominent challenges and integration logics within advanced manufacturing ecosystems. Research originating from China tends to emphasize deployment readiness, workforce upskilling, and training-intensive solutions. This aligns with the high salience of skills and capability gaps (C3) and strong consistency with change and training interventions (A4). Research from the United States more frequently highlights issues of transparency, fairness, privacy, and employee trust, reinforcing challenges related to trust (C4) and the coupling between algorithmic, explainable AI and governance-oriented pathways (A1 and A5). German and broader European research, rooted in Industry 4.0 and Industry 5.0 discourses, places greater emphasis on human–machine collaboration and organizational embedding, aligning with pathways centered on human–AI collaboration and organizational integration (A2 and A3).

By contrast, challenges associated with policy institutionalization and sustainability transitions (C5–C6), alongside standardization-oriented approaches (A6), appear relatively underrepresented. This suggests the resulting challenge–pathway matrix more strongly reflects contexts characterized by resource abundance and regulatory maturity. Consequently, for SMEs in developing economies, AI-HRM adoption may be constrained by more fundamental barriers such as limited data infrastructure, capability gaps, and relatively high compliance and governance costs. This implies a phased, context-sensitive adoption logic where capacity building, participatory work design, and low-cost AI collaborative mechanisms (A4 and A2) precede resource-intensive governance and standardization solutions.

Within the MLEF, the nesting mechanism operates through three distinct modes:

First, at the macro–meso interface, institutional nesting is achieved through regulatory pressure, normative pressure, and imitative pressure. National AI strategies, labor regulations, data governance rules, and ESG standards collectively shape how organizations configure AI-HRM systems, delineate accountability boundaries, and determine governance priority pathways. These institutional constraints and incentives directly influence organizational decisions regarding system design, training investment, and the extent to which ethical and sustainability considerations are embedded within HR processes. Secondly, Macro-level institutions shape micro-level perceptions primarily through meso-level organizational arrangements, while also exerting limited direct influence via highly visible regulatory signals (e.g., compliance requirements, data protection rights, and ethical standards) that frame employee expectations.

Labor protection laws enhance employees’ perceptions of employment security; data protection and privacy regulations shape trust in algorithmic transparency, while national or sectoral AI ethics guidelines influence expectations regarding the fairness and legitimacy of AI-assisted HR decisions. These institutional signals constitute employees’ initial cognitive frameworks for AI systems, influencing trust formation even before encountering specific organizational practices.

Finally, organizational embedding mechanisms, situated between meso and micro levels, translate institutional requirements and strategic intentions into concrete socio-technical arrangements. Through system architecture, governance practices, training programs, and change management, organizations embed AI within daily workflows. Structural design influences perceptions of fairness and transparency, while practice-based embedding—such as retraining initiatives and participatory governance—shapes actual usage patterns, resistance behaviors, and acceptance attitudes. Interactions within the MLEF framework are inherently dynamic and recursive. Organizational design decisions shape employee trust and behavioral responses, while employee usage patterns, resistance, and experiential feedback simultaneously reshape organizational practices. When trust deteriorates, or resistance intensifies, organizations are compelled to recalibrate governance mechanisms, restructure training systems, or adjust human–machine task allocation. Conversely, efficient human–machine collaboration and sustained system usage bolster organizational confidence in AI-HRM investments and capability development.

Crucially, over time, these micro-level responses coalesce into feedback loops whose impact extends beyond organizational boundaries. At the meso level, accumulated usage data, performance outcomes, and resistance patterns drive organizational learning and systemic recalibration. At the macro level, the aggregate organizational experience manifested through compliance challenges, ethical controversies, skills shortages, or successful implementation cases provides the basis for policy learning and regulatory adjustments. Thus, employee-level behavior becomes an indirect yet critical input factor in institutional evolution, achieving a closed-loop connection between micro-level practice and macro-level governance. In summary, the MLEF framework redefines AI-HR integration as a process of continuous learning and adaptation, rather than a one-off technological intervention. At the macro level, institutional frameworks establish boundary conditions and normative expectations; at the meso level, organizations translate these into governance structures and capability allocations; while at the micro level, employees ultimately implement, challenge, or reinforce AI-HRM systems through trust, adoption, or resistance. This framework thus reveals that sustainable AI-HR integration in manufacturing stems not solely from technological refinement but from the coordinated evolution and alignment of institutional legitimacy, organizational capability, and employee trust over time.

Figure 11 provides a structured overview of the multi-tiered embedding of AI-HRM within the manufacturing sector. It demonstrates that when regulatory and compliance pressures are imposed upon enterprises, the adoption of AI-HRM at the organizational level becomes more cautious. This, in turn, reduces investment in areas such as training, thereby undermining employee trust in the system and hindering integration efforts. Conversely, frontline resistance further impedes integration at the meso level, compelling industry and policy refinement to foster governance models featuring explainability (A1), auditability (A5), and continuous learning (A4). This framework emphasizes the interplay of challenges across tiers rather than attributing obstacles solely to technological factors.

Building upon this, the discussion reinterprets the findings regarding challenges, pathways, and evolution through the MLEF analytical lens. The macro level focuses on policy and institutional constraints, addressing RQ3, the meso level centers on governance approaches involving organizational change and pathway optimization, reflecting RQ2, while the micro level examines employee cognition, trust, and behavior, accentuating RQ1. This approach aligns with the Multi-level socio-technical systems analytical framework in technological change studies, emphasizing the mutual embedding of macro-level institutions, industrial structures, and micro-level practices [68]. It also resonates with organizational research analyzing the relationship between “learning algorithms” and organizational routines, treating AI as an embedded element within systems rather than a neutral tool [69,70]. Consequently, the key constraint on AI implementation in manufacturing lies not in the technology itself but in the mismatch across multiple levels of individuals, organizations, and institutions.

The manufacturing environment introduces structural characteristics that fundamentally distinguish AI-HR integration from service-oriented or knowledge-intensive industries. Manufacturing enterprises typically employ large numbers of blue-collar workers whose tasks are intrinsically linked to physical processes and shop-floor operations. Compared to knowledge workers, these employees have limited exposure to data analytics and algorithmic reasoning, widening the skills gap (C3). This amplifies resistance when introducing AI systems into HR functions such as training, performance evaluation, or safety-related operations. Within safety-critical production environments, algorithmic decisions directly correlate with occupational risks and liabilities, rather than abstract performance metrics. Consequently, trust-related issues (C4) in manufacturing are intrinsically tied to perceptions of physical safety and liability. This explains why resistance and trust erosion may escalate more rapidly in manufacturing settings than in digitally dominated environments.

Secondly, AI-HR in manufacturing operates as a tightly integrated component within coupled human–machine physical systems, rather than functioning as an isolated digital workflow. AI applications related to human resources, such as skills matching or performance analytics, are frequently interconnected with production scheduling, industrial IoT infrastructure, and physical information systems at the shop-floor level. This systemic coupling increases implementation complexity, as misalignment between HR systems and production systems compromises operational stability, not merely information transmission efficiency. Consequently, coordination challenges (C2) in manufacturing extend beyond aligning HR with IT, encompassing production units, safety governance, and operational controls. This embeddedness underscores why reliance on technological or algorithmic solutions alone is insufficient, and why effective AI-HR integration hinges on the co-configuration of organizational structures, governance mechanisms, and workforce capabilities.

Macro Level: Institutional Lag, Policy-Practice Disconnect, and Absence of Sustainable Transformation

The macro-institutional environment, encompassing national AI strategies, Industry 4.0 policies, data protection regulations, and industry standards, provides directional guidance and a legal foundation for enterprise AI-HRM integration. Earlier research on human-centered Industry 5.0 and management has highlighted that macro policies should not only drive digitalization and enhance competitiveness but also incorporate worker behavioral responses and social sustainability as ultimate objectives [71,72]. However, this study has identified that institutional lag and the disconnect between policy and practice at this level constrain the integration of AI-HRM in manufacturing.

Firstly, policies across multiple nations have spurred AI adoption in manufacturing across training, safety, and performance domains. China’s “Made in China 2025” and Germany’s “Industrial 4.0” initiatives have both driven corporate AI implementation to varying degrees [10,16]. Da Silva et al. [7], in their study of automotive manufacturing across China, Germany, and Brazil, noted that policies often prioritized technological innovation and industrial competitiveness while lacking guidance on critical HR issues such as “employee reskilling, labor protection, and AI ethics.” Consequently, although enterprises receive incentives for technological investment, they lack complementary mechanisms to safeguard employee rights.

Secondly, ambiguities in data governance and privacy regulations further exacerbate institutional uncertainty. While the EU AI Act and GDPR establish principles for AI systems, gaps remain concerning employee data privacy and liability allocation for algorithmic errors [3]. Certain multinational automotive and electronics manufacturers have been compelled to abandon certain AI-HRM functionalities due to cross-border data flow restrictions [7].

Thirdly, the absence of C6 at the policy level warrants attention. Within this study’s sample, only 12% of the literature linked AI sustainability with HRM, predominantly remaining at the initiative stage. A mere handful of studies (e.g., [2]) empirically demonstrated that AI-HRM enhanced organizational performance when integrated with scheduling and skills training.

Finally, policy research predominantly adheres to textual analysis and cross-national comparisons, lacking empirical studies across the policy, organizational, and employee levels. Similarly, research on platform economies and algorithmic management largely focuses on textual and macro-level aspects, with insufficient multi-level empirical investigations into policy design, corporate practices, and individual work experiences [73,74]. Consequently, the emergent term timeline in Figure 8 indicated that “governance and ethics” became prominent themes after 2021, yet systematic research on how macro-level institutions influence employee trust was still deficit.

Meso Level: Organizational Capability Traps, Cross-Departmental Collaboration Deficits, and Sub-Industry Heterogeneity

The meso level represents a critical tier determining whether AI-HRM can transition from technological advancement to organizational implementation. This study identified the primary meso-level contradiction as the decoupling between advanced technological applications and lagging organizational capabilities. Research on algorithmic management and HRM similarly indicated that if organizational structures, employee roles, and collaborative mechanisms were not synchronously adjusted, AI projects would be in their initial stages and struggle to integrate into daily management practices [75].

Organizations frequently fall into capability traps. Table 5 indicates that although C2 represented a core challenge, the Path A3’s literature coverage is only approximately 20%, while A4’s alignment with C3 reached 54%. This suggests enterprises prefer short-term training to bridge skill gaps rather than addressing deep-seated issues through governance restructuring [2]. This results in AI systems becoming tools confined to the IT department, hindering unified resource allocation between HR and production units.

Additionally, inadequate cross-departmental collaboration hinders technological implementation. While a few scholars have proposed tripartite collaboration mechanisms involving HR, IT, and production [76], most enterprises became trapped in a scenario where IT dominates, HR cooperates, and production resists. Without a shared governance framework, practical application proved to be challenging.

Moreover, significant heterogeneity exists across manufacturing sub-sectors. However, no stratified review of sub-sectors was conducted; consequently, we treat sub-sector heterogeneity as a boundary condition and prioritize comparative sub-sector designs in future empirical work. Highly automated industries like automotive and electronics prioritize AI applications in safety alerts, quality inspection, and performance analysis. Conversely, labor-intensive sectors focus more on low-cost training and scheduling solutions. Existing research, however, disproportionately emphasizes high-tech industries while neglecting labor-intensive ones. The literature on manufacturing digital transformation also proclaimed that differences in automation levels, capital intensity, and skill structures across industries significantly impacted AI project implementation pathways [77].

Therefore, the integration challenge at the meso level is not whether to adopt AI, but whether it can be embedded within cross-functional processes, governance structures, and strategic logic to ultimately foster sustainable organizational capabilities.

Micro Level: Weakened Employee Agency, Fractured Human–Machine Trust, and Unique Challenges in Developing Countries

The micro level represents the ultimate determinant of whether AI is genuinely accepted, utilized, and trusted, constituting the final stage of AI-HRM integration. This study has uncovered three phenomena.

Initially, employee anxiety regarding algorithms has become one of the most prominent challenges in manufacturing. Challenge 3 refers not only to operational inadequacy but also to employees’ difficulty in comprehending the rationale behind AI decisions without algorithmic explanations, thereby amplifying their skepticism [9].

Second, the “trust rupture” exhibited a pronounced threshold effect. As illustrated in Figure 6, C3 and C4 showed significant overlap within the HR5 module. Particularly in safety-critical scenarios, if AI monitoring systems produce major misjudgments without human-led redress mechanisms, employee trust in AI can suffer irreparable damage [9]. Delpla et al. [11] similarly noted that without human-centered feedback mechanisms, employees were more likely to perceive AI as a constraint. Preceding research confirmed that the absence of explanatory or appeal mechanisms in personnel decisions might erode organizational trust [78,79], thereby triggering irreversible trust thresholds.

Furthermore, samples from developing nations revealed distinct trust-generation dynamics. Malik et al. observed that in manufacturing sectors of emerging economies, worker anxiety towards algorithms stemmed predominantly from inadequate skills training and employment instability, whereas Western employees’ concerns primarily revolved around privacy and surveillance [10]. This inference illustrates how micro-level trust is shaped by institutional environments and cultural contexts.

Overall, the micro-level issues lie not in whether employees are willing to adopt technology, but in whether explainability, redressability, and capacity-building can enhance their agency, transforming them from passive recipients into active participants.

Synthesizing macro, meso, and micro perspectives reveals that research on AI-HRM in manufacturing has evolved from a technology-driven focus on automation and efficiency (2000–2015) towards a human–machine collaboration and contextual governance paradigm (2021–2025). During this evolution, the traditional Technology Acceptance Model (TAM) overemphasized individual perceptions of usefulness and ease of use, while the Resource-Based View (RBV) prioritized the importance of resources like technology and data. Neither adequately explains the role of institutional pressures, organizational governance, and employee trust across different levels.

Consequently, the proposed Multi-Level Embedded Framework encompassing the links between technology, organization, and trust offers three theoretical contributions. Firstly, at the micro level, it addresses TAM’s limitations by recognizing that employee acceptance of AI-HRM depends not only on perceived usefulness and ease of use but also on organizational governance and external institutional factors. Second, at the meso level, it expands the scope of organizational capability within the Resource-Based View (RBV), incorporating not only technological and data resources but also algorithmic governance capabilities, cross-departmental coordination abilities, and employee trust-building capacities into the realm of key strategic resources [2,3]. Thirdly, at the integrated macro, meso, and micro levels, the elaborated model employs a Socio-technical systems theory [68] to explain the nonlinear evolutionary trajectory of AI-HRM integration through dynamic feedback loops involving macro-level institutions, meso-level governance, and micro-level trust.

Overall, the model reveals that AI-HRM can only transcend fragmented pilot schemes and progress towards multi-level embedding and systemic integration—thereby supporting manufacturing’s transition towards human-centered intelligence when technological explainability is ensured, organizational governance structures are both agile and robust, and employee trust is continuously nurtured.

5. Conclusions

By leveraging the MLEF framework to synthesize diverse challenges, pathways, and evolutionary dynamics, this discussion elucidates the intrinsic mechanisms underpinning the integration of AI-HR within manufacturing. The subsequent section will summarize key findings and theoretical and practical implications, while outlining limitations and future research directions.

This study adheres to the PRISMA 2020 guidelines, systematically reviewing 347 English-language articles on AI-HRM integration in manufacturing from 2000 to 2025, with in-depth coding based on 100 core publications. The investigation centered on three core research questions (RQ1–RQ3). Consequently, a knowledge architecture comprising six categories of challenges, six pathways, and a three-stage evolutionary framework was constructed. The Multi-Level Embedded Framework (MLEF) was proposed, offering a structured perspective for understanding the deep coupling of technology, organization, and personnel within the context of intelligent manufacturing.

First, the integration of AI and HRM in manufacturing has exhibited a multidimensional, nested challenge structure. The identified Challenge 3. Skills and capability gaps and C4. Trust and ethical concerns were frequently identified as two core challenge categories across diverse studies, aligning strongly with existing research on inadequate personnel AI literacy, employment anxieties, and trust crises [9,10]. Furthermore, Challenge 1, Technical and data infrastructure, and Challenge 2, Organizational and strategic alignment, represented pervasive structural bottlenecks [2]. Although challenges C5, Institutional and Regulatory Constraints, and C6, Sustainability and green transition, at the macro level appeared less frequently in the sample, institutional factors significantly influenced the diffusion pathways and boundaries of AI-HRM [3].

Secondly, the six integrated pathways (approaches) provided differentiated solutions for distinct challenge types. Employing a structured matching approach via the Challenge–Approach Matching Matrix enabled the six approaches to address the six challenge types. It is evident that the stipulated Approach 1, Algorithms and Explainable AI, exhibited the highest matching strength for resolving Challenge 3, Technical and data infrastructure [11]. Approach 4, Change and Training Interventions, proved most effective against Challenge 3, Skills and capability gaps [8], while the combination of Approach 2, Human–AI Collaboration, and Approach 5, Ethics and Governance Mechanisms, can effectively mitigate Challenge 4, Trust, Fairness and Ethical Concerns [80]. Although Approach 6, Policy and Standardization, was relatively infrequently referenced in the literature, it played an irreplaceable role in addressing institutional challenges such as data governance, legal liability, and industry standards [7].

Thirdly, the research theme underwent a three-stage evolution from technology-oriented to human-centered and governance-oriented approaches. Based on the thematic evolution diagram (Figure 8), three distinct phases were identified: the technology-driven period (2000–2015) primarily focused on automation, expert systems, and process optimization; the data analytics phase (2016–2020) centered on machine learning, predictive analytics, and recruitment optimization; and the governance and collaboration phase (2021–2025) revolved around explainability, fairness, human–machine collaboration, and system integration. This reflects a shifting research paradigm from an “efficiency and performance” perspective towards a comprehensive view of “human-centered values and governance mechanisms” [81].

Fourthly, the MLEF framework has revealed structural differentiation and integration gaps across research levels. Macro-level research primarily focuses on national AI strategies, data governance, and industrial policies. Meso-level studies concentrate on enterprise-level AI-HRM system development and process re-engineering. Micro-level research is focused on employee acceptance, safety behaviors, and recruitment decisions [2,7,82]. However, existing research generally lacks cross-level empirical testing that integrates macro policies, meso organizational governance, and micro employee responses within a unified analytical framework [83]. Through the Multi-Level Embedded Framework (MLEF), this study has synthesized these fragmented research strands into a longitudinal linkage structure of “Policy-organization-employee”, providing a clear basis for subsequent cross-level, systematic research.

5.1. Theoretical and Practical Implications

Theoretical contributions comprise three key aspects, the first of which is establishing a three-dimensional Challenge–pathway–evolution framework that integrates research scattered across technological, organizational, and individual levels, addressing the fragmentation in the existing literature. Additionally, this study has proposed six categories of integration pathways and employed a Challenge–Approach Matching Matrix to establish explicit correspondences between problems and solutions, which renders previously implicit theories testable and reusable, enhancing the explanatory power of the review. MLEF, incorporating policy, organization, and employee behavior into a unified analytical system, represents the third contribution, since it expands the application of the technological systems perspective within intelligent manufacturing scenarios, providing a foundational theoretical basis for future cross-level research.

Regarding practical contributions, this paper provides manufacturing enterprises with a toolchain encompassing “challenge diagnosis—path selection—system integration”. In the initial phase, enterprises may prioritize Pathways A4 and A2 to mitigate skill gaps and trust barriers. During the deepening stage, enterprises should establish a standardized AI application system through Pathways A3 and A5. Policy makers should refine systems concerning data governance, algorithmic accountability, and skills training to form a “policy–organization–individual” collaborative enhancement mechanism, thereby advancing the practical implementation of intelligent manufacturing.

5.2. Research Limitations

Although the methodology combining bibliometric mapping (n = 347) with deep coding (n = 100) is rigorous, certain limitations should be acknowledged.

Firstly, database coverage and indexing practices may introduce selection bias. This review relies on Scopus and Web of Science, which may inadequately represent regionally indexed outlets and practitioner-oriented venues. Moreover, the reduction in publications by 2025 may partly reflect indexing delays rather than a substantive decline in research output.

Secondly, search string design and field restrictions (title/abstract/author keywords) may have overlooked studies primarily discussing AI integration with human resource management within full texts, industry case sections, or Supplementary Materials. Despite iterative keyword optimization, the retrieval strategy inevitably reflected the boundaries of predefined query terms and metadata availability.

Thirdly, this review is restricted to English-language journal articles and reviews, excluding conference proceedings and the gray literature. Given the rapid proliferation of AI applications in manufacturing, this choice may reduce coverage of emerging industrial implementations and technical prototypes initially appearing in non-journal media.

Fourth, screening and qualitative coding entailed interpretative judgements. Whilst a structured open–axial–selective coding procedure and iterative self-checking were applied, category assignments (e.g., distinguishing trust-related barriers from governance constraints) may still contain subjectivity.

Fifth, the core sample (n = 100) for in-depth coding was selected through criteria-based logical choices emphasizing manufacturing relevance and information density, rather than citation-based ranking. This enhances the interpretability of RQ1–RQ2 but may compromise representativeness across regions, sub-sectors, and methodological traditions.

Finally, method-specific constraints apply. Co-occurrence networks depend on authors’ keyword selections, while abstract-based LDA thematic modeling may fail to fully capture methodological nuances embedded within full texts. Consequently, identified phase and thematic transitions should be interpreted as robust patterns at the corpus level rather than exhaustive representations of all sub-streams. Future research may broaden coverage by further validating multi-level dynamics through integrating additional databases, multilingual resources, and full-text thematic models, alongside conducting cross-sectoral comparisons.

5.3. Future Research Directions

Based on the findings of this systematic review and the proposed MLEF framework, future research may advance AI-HRM studies in manufacturing along several specific avenues.

Firstly, forthcoming investigations could develop multi-level empirical designs that explicitly link macro-level institutional conditions, meso-level organizational governance, and micro-level employee responses within a single analytical structure, while accounting for variations across different institutional and cultural contexts. Researchers may employ cross-level datasets or nested study designs, rather than examining these levels in isolation, to explore how policy signals (such as data governance or AI regulation), organizational embedding mechanisms (such as training systems or governance structures), and employee trust or capability formation collectively shape AI-HRM outcomes over time. Such designs would directly address the cross-level gaps identified in the existing literature and empirically operationalize the MLEF logic.

Secondly, future research may prioritize longitudinal and process-oriented approaches to capture the dynamic nature of AI-HRM integration within manufacturing. Given that skill development, trust formation, and employee capability building constitute an evolving socio-cognitive process, longitudinal case studies, cohort-based analyses, or sequential mixed-methods designs could be employed to trace how AI-HRM initiatives progress from pilot phases to organizational practices. This direction will enable researchers to move beyond static snapshots and better understand feedback loops between organizational learning, employee behavior, and institutional adjustments.

Finally, scholars may extend AI-HRM research further into sustainability and green transition contexts by integrating human resource analytics with ESG-related outcomes. Whilst existing research primarily focuses on efficiency and performance, future studies could explore how AI-supported HRM facilitates green skills development, workforce transitions in low-carbon manufacturing, and societal sustainability goals. Comparative studies across manufacturing sub-sectors or differing institutional environments may prove particularly valuable in revealing distinctions between sustainability-oriented AI-HRM practices and efficiency-driven applications.

Collectively, these directions encourage future research to shift from fragmented, single-level analyses towards integrated, longitudinal, sustainability-focused investigations. This approach would enhance both the theoretical depth and practical relevance of AI-HRM research within manufacturing contexts.