Generative Artificial Intelligence in Supply Chain: Review, Trends, and Future Directions

Abstract

Background: Generative artificial intelligence (GenAI) has attracted significant attention in supply chain management (SCM) due to its potential to improve data-driven decision-making and operational performance. However, existing studies mainly focus on individual GenAI models or specific supply chain applications, lacking a comprehensive understanding of how different GenAI architectures support decision-making across the supply chain. Methods: This study conducts a systematic literature review using the PRISMA framework to examine the applications of Generative Adversarial Networks (GANs), Transformers, Variational Autoencoders (VAEs), and flow-based models within a six-level supply chain decision-making framework. A total of 692 peer-reviewed publications were analyzed using bibliometric methods, including keyword co-occurrence, temporal and density analyses, and Supervised Embedding Visualization. Results: Current research is concentrated on Transformer and GAN applications, particularly in data analytics, optimization, forecasting, manufacturing, transportation, logistics, and quality management. The analyses also reveal major research themes, the evolution of GenAI in SCM, and limited attention to sustainability, cybersecurity, resilience, and reverse logistics. Conclusions: This study provides a comprehensive overview of GenAI applications in SCM, identifies key research gaps, and offers a foundation for future research while helping practitioners evaluate opportunities and limitations of GenAI for supply chain decision-making.

1. Introduction

In recent years, companies have operated in increasingly volatile environments with demand uncertainty, supply disruptions, and geopolitical risks. Global events, such as the COVID-19 pandemic, trade conflicts, and natural disasters, have shown the extent of the fragility of traditional supply chain systems [1]. This highlights the need for resilience, flexibility, and data-driven decision-making [2,3]. Meanwhile, the digital transformation of supply chains has also led to large amounts of real-time data being collected from sensors, logistics systems, and online platforms. This data offers new opportunities to analyze and improve supply chain operations but also creates challenges in analyzing and using it effectively [4]. To manage this complexity, supply chains have recently relied on artificial intelligence (AI). AI helps identify patterns, improve forecasts, and support faster and more accurate decision-making [5,6]. As modern supply chains become more connected and dynamic, AI tools can improve visibility, responsiveness, and resilience, which are now essential for competing in uncertain and fast-changing global markets [7].

AI supports critical supply chain functions, such as demand forecasting, inventory optimization, supplier selection, production scheduling, and logistics planning. Unlike traditional methods, AI-based systems can process vast amounts of real-time data to uncover hidden patterns, anticipate disruptions, and suggest proactive interventions. However, realizing these benefits in practice often requires overcoming challenges related to data quality, system interoperability, and data silos that remain common across supply chain networks. Various organizations have used AI to understand the potential needs of each customer and the allocation of required resources in different locations to potentially optimize their operations and maintain competitiveness in the market. For instance, Amazon uses AI to optimize delivery routes, improve warehouse robotics, and improve forecasting, which helps them fulfill more same-day deliveries [8]. Starbucks implemented an AI-driven inventory counting system across over 11,000 stores in North America to track inventory, improve availability, and thus reduce stockouts [9].

Recently, generative artificial intelligence (GenAI) has emerged as one of the most transformative advancements in the field of artificial intelligence. Unlike traditional AI models that focus primarily on prediction, GenAI models, such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Large Language Models (LLMs) can generate new data, patterns, and insights based on learned representations [10]. The rapid evolution of GenAI technologies, such as GPT-5 has demonstrated significant capabilities in Natural Language Processing, image synthesis, simulation, and decision support. According to recent industry analyses, generative AI is being increasingly adopted across sectors to accelerate innovation, automate content generation, and enhance analytical reasoning [11,12]. The adaptability and creative potential of these models make them particularly relevant for complex and data-intensive domains, such as supply chain management, where decision-making often requires processing vast amounts of structured and unstructured data to respond dynamically to uncertainty and change [11,12,13]. Research on GenAI has revealed various recent advancements. This research and development has led to several popular GenAI algorithms, including Large Language Models (LLMs), such as Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer 3 (GPT-3), Reinforcement Learning from Human Feedback (RLHF), Sequence-to-Sequence (seq2seq) models, and Variational Autoencoders (VAEs) enabling advanced content generation, language understanding, and complex task solving [14].

While GenAI has rapidly advanced in various domains, such as image synthesis, text generation, and design automation, its application in SCM is growing. The AI in supply chain market is expected to grow at over 30% annually, reaching around US$22.7 billion by 2030. This may reflect both rising demand and the scaling deployment of AI tools [15]. According to a report by Gartner, 70% of large organizations will adopt AI-based supply chain forecasting by 2030 due to the need for more adaptive demand planning in volatile markets [16]. However, only 23% of supply chain organizations currently possess a formal AI strategy, indicating a significant gap between potential and structured practice [17]. Recent studies have explored the use of GenAI for various supply chain tasks, such as demand forecasting, inventory optimization, production scheduling, and supply chain risk management. For instance, GANs have been used to generate synthetic demand data to improve forecasting models under data scarcity conditions [18]. In addition, VAEs have been applied for scenario modeling to improve supply chain resilience and sustainability [19]. Transformers and LLMs, such as GPT-based architectures, have shown potential in improving decision support systems, automating supplier communications, and extracting insights from unstructured logistics data [19].

GenAI has the potential to enhance supply chains in various ways, for example, by improving demand forecasting, vendor negotiations, product development, and logistics optimization due to its unique abilities, such as generating data, automating predictions, engaging various automated agents for efficient interaction with clients, and understanding potential keywords and image synthesis [20]. Therefore, it has gained extensive popularity across various industries. A recent study shows that approximately 40% of organizations that invest in GenAI aim to use it to improve efficiency and resilience [21]. Some retailers, such as Walmart, use generative AI to enhance search results for customers, resulting in improved shopping experiences and operational efficiency [22]. In addition, Walmart’s Pactum AI chatbot reduced vendor negotiation times, resulting in 20% cost savings. Similarly, Amazon’s Project P.I. AI inspects products for defects before shipping, reducing returns by 30% [23]. IKEA employs generative AI, ensuring ethical implementation through workforce training to boost customer service and sales [24]. In the fashion supply chain, GenAI enhances efficiency through data-driven inventory management and forecasting [25]. Retailers, such as Perry Ellis and H&M use AI to reduce return rates by 15% and increase profits by 16% through improved product descriptions and targeted advertising [26]. In the food industry, General Mills has developed AI tools to improve supply chain efficiency with initiatives, such as the ELF logistics product and COD Pod application, resulting in reduced waste and a faster and more accurate decision-making process [27]. In the transportation sector, Maersk utilizes AI in various ways, such as using AI robotics to enhance logistics, tripling sorting speed and boosting inventory pickup by 33%, optimizing routes to reduce costs, and using ChatGPT to streamline contract negotiations for greater efficiency [28]. FedEx utilizes AI for delivery time estimates and for optimizing logistics through machine learning to enhance operational efficiency, forecasting accuracy, and customer service [29].

Despite these advances, most research focuses on narrow applications rather than integrated decision-making frameworks across supply chain functions. Furthermore, existing studies often emphasize technical model development over a comprehensive analysis of their practical business value, interpretability, or scalability in real-world supply chain environments. The literature does not yet offer an integrated perspective that connects specific GenAI techniques with their corresponding decision-making roles across the entire supply chain lifecycle. Moreover, existing review articles typically focus on specific GenAI architectures, selected SCM functions, or descriptive summaries of applications, without systematically comparing multiple GenAI models across the full range of supply chain decision-making tasks. Jackson et al. [30] discuss the potential of digitalization and AI in enhancing supply chain efficiency. However, it does not provide a unified framework that encompasses all SCM areas, including reverse logistics and sustainability considerations. Similarly, Wamba et al. [31] explore the transformative impacts of GenAI and ChatGPT on operations and SCM, yet its analysis is limited to specific use cases without addressing the broader implications for supply chain frameworks. Although these studies provide valuable insights into the potential of GenAI, they do not combine review procedures with bibliometric analyses, Supervised Embedding Visualization, and a unified SCM decision-making framework to examine how different Generative AI models are applied across supply chain tasks. Furthermore, while studies have examined the use of GANs for specific applications, for instance, demand forecasting, they often neglect the ethical and sustainability dimensions critical to modern SCM practices [32]. The existing research predominantly focuses on isolated applications of GenAI, leaving a gap in understanding how these technologies can be holistically applied to enhance overall supply chain resilience and sustainability. This gap highlights the need for this research, which integrates multiple GenAI architectures within a unified supply chain decision-making framework and examines their applications using bibliometric analysis, Supervised Embedding Visualization, and qualitative synthesis.

This study provides a structured and comprehensive review of GenAI uses in SCM. We use a six-level decision-oriented framework that categorizes key SCM decisions to systematically analyze how different GenAI models, such as GANs, Transformers, VAEs, and flow-based models, are applied across these decision levels. After reviewing the major categories of GenAI and recent advancements, we use the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) approach and Supervised Embedding Visualization on over 692 collected peer-reviewed publications to identify current research trends, practical applications, and how each GenAI model supports specific supply chain tasks. This study simultaneously compares multiple GenAI architectures, maps their relationships with SCM decision-making tasks, and identifies research trends through bibliometric, temporal, density, and Supervised Embedding Visualization. The paper highlights critical research gaps at the intersection of GenAI models and SCM decision areas and recommends directions for future research. The main research questions of this paper are:

• What are the major research trends in the application of GenAI to SCM?
• How are different GenAI models applied across the various supply chain decision-making tasks?
• How has research on Generative AI in supply chain management evolved over time, and what patterns can be identified through bibliometric and Supervised Embedding Visualization analyses?
• What research gaps and future opportunities exist for the application of GenAI in SCM?

This work synthesizes the state of knowledge and provides a unified perspective that links GenAI models, SCM decision-making tasks, bibliometric evidence, and future research opportunities. The proposed framework can also serve as a reference for future reviews, bibliometric studies, and empirical investigations of GenAI adoption in supply chain management.

The rest of the paper is organized as follows. Section 2 presents the research methodology. Section 3 provides a brief review of various GenAI architectures. Section 4 presents a framework that highlights the key tasks involved in the overall closed-loop supply chain. Section 5 explores the potential uses of Generative AI in SCM. Section 6 evaluates the existing literature on GenAI in the supply chain using bibliometric analysis. Section 7 evaluates the applications of GenAI in SCM tasks using Supervised Embedding Visualization. Finally, Section 8 concludes the paper with directions for future research.

2. Research Methodology

We use a literature review approach to assess and synthesize the research on SCM and GenAI. The approach follows the PRISMA framework [33] to provide a transparent, reproducible review process (The checklist is available in the Supplementary Materials.). The structured guidelines of this framework for literature identification, screening, eligibility assessment, and study selection help reduce selection bias and improve the credibility and reliability of the findings [34].

2.1. Sources of Information and Search Strategy

We obtained data from the core collection of the Web of Science, which is known for high-quality peer-reviewed publications in the English language. The database was searched, and data was downloaded on 25 April 2025, and it includes peer-reviewed papers from January 2020–April 2025 to align our research with the rapid post-2020 increase in Generative AI research, with a focus on the intersection of Generative AI architecture and SCM decision-making tasks. Research that emphasizes the usage of Generative AI related to circular supply chain management was considered using the following Boolean logic for the search:

(“Generative AI” OR “Generative Artificial Intelligence” OR “Generative Adversarial Network*” OR “GAN*” OR “Transformer*” OR “Variational Autoencoder*” OR “VAE*” OR “flow-based model*” OR “Large Language Model*” OR “LLM*”) AND (“Supply Chain Management” OR “SCM” OR “Logistics” OR “Inventory Management” OR “Demand Forecasting” OR “Demand Planning” OR “Demand Sensing” OR “Demand Identification” OR “Procurement” OR “Sourcing” OR “Warehousing” OR “Transportation” OR “Scheduling” OR “Manufacturing” OR “Network Design” OR “Reverse Logistics” OR “Product Returns” OR “Repair” OR “Warranty” OR “Customer Services” OR “Refurbishment” OR “Remanufacturing” OR “Order Management” OR “Asset Management” OR “Optimization” OR “Data Analysis” OR “Product Life Cycle Management” OR “Quality Management” OR “Risk Management” OR “Resilience” OR “Sustainability” OR “Ethics”).

These search criteria were used to restrict our search to the identified frameworks of Generative AI, which include Generative Adversarial networks, Variational Auto-Encoders, Transformers, And flow-based models, as well as general “Generative AI” terminology to include a broader range of focused data to reduce bias. Although digital infrastructure, such as cloud computing, the Internet of Things, cybersecurity, and edge computing are relevant, if studies discuss these technologies without explicitly referring to Generative AI and SCM, they do not meet the criteria.

Since several search terms, for instance, Transformers, optimization, and data analysis, are widely used across multiple disciplines, the initial search intentionally prioritized sensitivity over specificity to reduce the likelihood of overlooking relevant studies. Consequently, all retrieved publications underwent a screening process following the PRISMA guidelines. This screening process ensured that the final dataset consisted solely of studies with direct relevance to supply chain management.

2.2. Eligibility Criteria

To ensure methodological rigor and relevance, a predefined set of inclusion and exclusion criteria was established. The inclusion criteria required studies to be (1) peer-reviewed journal articles published between 2020 and 2025. This captures the rapid growth and adoption of Generative AI research during this period; (2) articles written in English; (3) focused on Generative AI, either through explicit discussion of Generative AI or specific generative architectures, including Generative Adversarial Networks, Variational Autoencoders, Transformers, and flow-based generative models; (4) directly related to at least one supply chain management function or associated process; and (5) published in academic journals to ensure the quality and reliability of the reviewed literature. Studies were excluded if they were review articles or other non-original publications, conference proceedings, or book chapters.

2.3. Study Selection and Screening Process

A multi-staged, iterative screening method was used according to the PRISMA 2020 standards [35]. During screening, we manually studied the abstracts to understand the context and select papers that focus on Generative AI and supply chain management. Each article was independently assessed by two reviewers. Titles and abstracts were initially screened against the predefined inclusion and exclusion criteria. Studies that did not meet the intersection of Generative AI and SCM, lacked methodological transparency, or fell outside the above-defined scope were manually eliminated. Publications that satisfied the screening criteria or for which eligibility was uncertain were subsequently evaluated through a full-text review. Disagreements between reviewers were resolved through discussion until consensus was reached. The final inclusion decisions were based on mutual agreement between the reviewers. The process is shown in Figure 1.

2.4. Study Selection Results

The initial database search based on the search and exclusion criteria in Web of Science resulted in 8124 records. After screening, 7327 records were excluded as they did not meet the intersection of Generative AI and SCM, based on abstract and review. We also analyzed the remaining 797 articles for further discrepancies in more detail using full text. This resulted in the exclusion of 105 articles, and the remaining 692 papers were retained for analysis. Backward citation tracking was performed to identify potentially relevant publications that were not retrieved during the database search. However, no additional eligible studies meeting the inclusion criteria were identified.

2.5. Research Methods and Analysis

After finalizing the list of articles, the analysis phase employs bibliometric tools to quantify research trends and scholarly influence, including trend topic evaluation, keyword co-occurrence mapping, temporal analysis, and citation network analysis. In addition, a qualitative content review complements the bibliometric findings to delve deeper into emerging research themes, conceptual linkages and co-occurrence mapping, and existing gaps in the literature. This integrative methodological approach enables both a macro-level understanding of the field’s evolution and a micro-level insight into its conceptual development and future research trajectories.

3. Generative AI Models

GenAI includes various sub-categories, such as Generative Adversarial Networks, Variational Autoencoders, Transformers-based, and Flow-based generative models. In this section, we briefly review each model and its recent advancements.

Table 1. Summary of GenAI models.

Model	References	Strengths	Limitations	Applications
Generative Adversarial Networks	[36,37,38,39,40,41,42,43]	Learning with minimal labeled data, generation of high-quality, diverse outputs, adaptability to multimodal inputs, effectiveness for representation learning, and applicability in imitation learning without explicit rewards	Training instability, mode collapse, sensitivity to hyperparameters, difficulty in achieving convergence, challenges with multimodal or manifold data	Image generation, enhancement, and translation; text-to-image synthesis; video- and audio-based face generation; and imitation learning in robotics and autonomous driving
Variational Autoencoders	[44,45,46,47]	Scalable probabilistic modeling, efficient inference, ability to capture complex latent patterns, smooth interpolation, effectiveness for collaborative filtering	Approximate posterior may miss complexity, limited applicability for discrete variables, potential overfitting, unstable deep hierarchical training	Image and text generation, data denoising, inpainting, super-resolution, and recommender systems
Transformer-based Models	[48,49,50,51]	State-of-the-art performance in NLP, capturing long-range dependencies, enabling parallel training, flexible personalization, high performance across diverse domains	High data and computational requirements, opaque decision-making processes, struggles with complex reasoning, large architecture complexity	Machine translation, summarization, question answering, chatbots, protein structure prediction, and recommendation systems
Flow-based Generative Models	[52,53,54]	Exact log-likelihood computation, invertible transformations that allow tractable inference, efficient parallelizable synthesis, and an interpretable latent space	High computational weight, sensitive to initialization, complexity for manifold-valued data, and optimization challenges	Image generation, inpainting, denoising, speech synthesis, brain image reconstruction, and texture generation

presents a summary of their applications, advantages, and limitations.

3.1. Generative Adversarial Networks

Generative Adversarial Networks (GANs) introduced by Ian Goodfellow in 2014, are AI algorithms used in unsupervised machine learning [55,56]. They consist of two neural networks [36]: a generator creating data and a discriminator evaluating it, competing in a zero-sum game [37,38]. The generator produces random data initially and improves over time to mimic real data, learning from discriminator feedback without direct access to real data [38,57]. GANs are known for generating realistic images, videos, and audio, with applications in image synthesis, style transfer, and super-resolution [39]. Despite their potential, GANs face challenges in terms of training instability and mode collapse [37]. Future research aims to improve network design and explore connections with reinforcement learning [58].

3.2. Variational Autoencoders

Variational Autoencoders (VAEs) introduced by Kingma and Welling in 2013, are generative models that learn latent data representations using deep neural networks [44]. The encoder maps input data into a latent space by producing mean and variance parameters for a probabilistic distribution [59], while the decoder reconstructs data from sampled latent points [60,61]. VAEs optimize via a variational lower bound on the marginal likelihood, avoiding complex calculations, and employ reparameterization for efficient gradient-based training [62]. They generate new samples by sampling from the learned distribution, enabling applications such as image synthesis [63]. However, VAEs may produce blurry outputs due to Gaussian priors and lack interpretability in latent spaces [45,64]. Advanced variants, such as Nouveau VAE (NVAE), use hierarchical structures and stabilizations like spectral regularization to improve high-resolution image generation [46,65]. Another major extension uses collaborative filtering for recommendation systems [47].

3.3. Transformers-Based Models

The Transformer architecture, introduced by [66], forms the foundation of most state-of-the-art language models. It employs an encoder–decoder structure with self-attention to weigh contextual relevance in sequential data, enabling efficient parallel processing and faster training [66]. Positional encoding preserves sequence order without recurrence, supporting tasks such as translation and text generation [67]. Its scalability has enabled large language models such as GPT-3 and more advanced models, widely used in generative AI [68]. Transformer-based models excel in knowledge-intensive tasks, including summarization, insight extraction, and creative output [69], yet remain vulnerable to adversarial perturbations [48]. Variants include Temporal Fusion Transformers for time series [70], T5 for enhanced Natural Language Processing tasks [71], and Vision Transformers for image classification [72].

3.4. Flow-Based Generative Models

Flow-based generative models transform a simple distribution into a complex one through a sequence of invertible transformations [73], enabling high-quality sample generation and interpretable representations [74]. These models operate by applying efficient, reversible mappings to a base Gaussian distribution, allowing the generation of diverse, realistic samples with precise latent space control [52]. They support exact latent variable inference and log-likelihood optimization without approximation [52]. In this category, Real-Valued Non-Volume Preserving (Real NVP) models employ affine coupling layers for efficient modeling, but they may yield improbable samples when prioritizing diversity over quality [53]. Generative Flow (Glow) enhances scalability and expressiveness using invertible convolutions and permutations, excelling particularly in image generation [52]. However, for video tasks, high dimensionality and temporal dependencies pose computational challenges, requiring careful modeling of dynamics [75].

4. Supply Chain Tasks

There have been a few frameworks for supply chain tasks in the literature. The original development of the Supply Chain Operations Reference (SCOR) architecture was based on a linear supply chain structure, and it was later upgraded to incorporate circular supply chain operations [76]. Current models, including the updated SCOR Digital Standard (SCOR-DS), have intrinsic limitations when applied, especially at the junction of the circular economy and Generative AI. Although these models offer a thorough basis, they may not be able to fully capture the contextual and causal elements that contribute to performance gaps. As a result, academics and professionals have argued that these models should be combined with other analytical techniques to increase their efficacy [77].

We use a modified version of the SCOR model accordingly. In contrast to SCOR, which views AI as a general enabler [78], this classification has six levels of tasks as shown in Figure 2. The framework was developed using a concept-driven approach based on the SCOR model and was refined through an iterative review of the supply chain and Generative AI literature. Specifically, SCOR served as the conceptual foundation, while the reviewed literature was used to identify increasingly important decision-making activities that were not explicitly represented within the traditional SCOR structure following the emergence of GenAI.

The framework was developed iteratively. The supply chain core activities in Level 2 were first established based on the SCOR operational processes. Subsequently, Level 1 was introduced to represent planning activities that drive operational decisions, while Level 0 was added to capture analytical capabilities, such as data analytics and optimization, that support all supply chain functions. Level 3 was introduced to represent cross-functional ethics and sustainability decisions, Level 4 to capture operational management and product lifecycle decisions, and Level 5 to represent reverse logistics and post-sale activities identified through the literature on circular supply chains.

This decision-oriented perspective provides a more suitable framework for classifying GenAI applications than a purely process-oriented structure. Additionally, the framework allows for a more focused, managerially usable architecture by considering a separate level for ethics and sustainability tasks (Level 3), as well as defining a level for product recovery mechanisms (Level 5). This provides specific guidance that is frequently implicit or necessitates significant customization within the broader, traditional structure of the SCOR standard. Since the framework is independent of a particular GenAI architecture, it can also be extended to incorporate emerging technologies and future developments in AI, which can provide a foundation for subsequent empirical and conceptual research. It is worth mentioning that the proposed framework classifies supply chain decision-making tasks rather than replacing the SCOR model. While SCOR provides a process-oriented representation of supply chain operations, the objective of this study is to categorize how GenAI supports different types of managerial decisions throughout the supply chain. The six levels in the framework represent conceptually distinct categories of supply chain decision-making; hence, they are not intended to be mutually exclusive. Instead, they reflect the interconnected nature of supply chain activities, where a single GenAI application may support multiple managerial decisions simultaneously. The framework and its six levels are described below:

• Level 0 serves as the analytical foundation of the supply chain tasks. It converts raw, multi-source data into actionable insights using advanced analytics and optimization techniques. This level focuses on finding patterns, predicting outcomes, and recommending optimal decisions that balance cost, quality, and performance. Therefore, it enables proactive, data-driven decision-making and ensures operational efficiency across interconnected supply chain processes using historical data, predictive analytics, and optimization.
• Level 1 focuses on supply chain planning and analysis. It involves assessing market demand, forecasting future needs, and aligning supply and capacity accordingly. This level also addresses network planning, structuring suppliers, warehouses, and distribution points to balance service levels and costs. This level ensures adaptation to changing market conditions and data to achieve efficient and effective supply chain performance.
• Level 2 encompasses sourcing and selecting suppliers, procuring materials, and managing warehousing operations. This level ensures the efficient flow and storage of goods, supports order fulfillment, and manages inventory and capacity buffers to address demand variability for suppliers, manufacturers, and end consumers.
• Level 3 emphasizes the ethical and sustainable responsibilities of the supply chain. This level impacts performance, stakeholder perception, and community impact. The goal is to minimize environmental footprint, ensure ethical practices, and integrate sustainability in both forward and reverse supply chains to build trust, corporate responsibility, and long-term brand value.
• Level 4 focuses on operational management and product lifecycle oversight. It includes product development, order management, asset management, risk mitigation, and quality management to ensure efficient production and delivery. This level aims to maximize profitability, minimize costs, and maintain resilience in the supply chain while aligning operations with ethical and sustainable practices highlighted in Level 3.
• Level 5 addresses reverse logistics and after-sales services. It focuses on product returns, refurbishment, remanufacturing, repairs, and warranty management to recover value from products while ensuring customer satisfaction. This level emphasizes sustainable practices, efficient transportation, and effective documentation to support circular supply chains, enhance consumer trust, and extend the lifecycle of products.

Level 0 includes foundational analytical capabilities, such as data analytics and optimization, that support decision-making across all supply chain functions, underpinning all other activities and providing a clear blueprint for applying GenAI to optimize uncertain return streams and scenario simulations. Levels 1–5 represent progressively more specific operational, strategic, governance, and after-sales decision areas.

Each article was reviewed using its title, abstract, keywords, and the full text (if necessary) to identify the SCM tasks addressed in the study. Because many GenAI applications may involve multiple supply chain tasks, an article may be assigned to more than one task or level according to the coding table in Appendix A. Consequently, the proposed framework should be interpreted as a multi-label classification of SCM tasks rather than a mutually exclusive categorization of articles. This reflects the interconnected nature of modern supply chains and enables a more comprehensive representation of GenAI applications.

5. Generative AI in SCM

Table 2. An illustrative mapping of papers between GenAI models and supply chain decision-making areas.

Generative AI	Level	Task	Description	Reference
Generative Adversarial Network	0	Optimization	Genetic Algorithms are improved by GAN-driven market simulations.	[79]
			Produces realistic datasets to mitigate data scarcity.	[80]
		Data Analytics	Multimodal continuous and imbalanced discrete problems in tabular data are addressed by CTGAN.	[81]
			CTAB-GAN+ produces synthetic tabular data with high utility and privacy preservation.	[82]
	1	Identification of Demand	M-GAN-XGBoost forecasts product sales from historical data by combining LSTMs, GANs, and XGBoost.	[83]
			EPR-GAIL utilizes purchase, exploration, and preference decisions to represent user consumption.	[84]
		Forecasting	TimeGAN improves synthetic time-series forecasting by maintaining temporal dynamics.	[85]
		Network Design	Creates robust and effective supply chain networks by learning supply chain trends.	[86]
		Demand Planning	Learns latent distributions to forecast future demand patterns by considering spatiotemporal uncertainty and external factors.	[87]
	2	Sourcing	Creates realistic demand scenarios to enhance planning and supplier negotiations.	[88]
		Procurement	Creates synthetic, realistic demand situations for new products and uses simulated customer-order patterns to improve procurement strategies.	[88]
		Inventory Management	Adaptive multivariate GANs optimize replenishment and stress-test systems by simulating realistic inventory demand.	[89]
		Warehousing	QGANs produce realistic warehouse data to maximize layout, slotting, picking, and labor–robot collaboration.	[90]
		Transportation and Logistics	Improves marine ship-speed forecasts by capturing spatiotemporal dynamics and guaranteeing reliable, distributionally faithful predictions.	[91]
	3	Ethics	CGANs-KHO with blockchain enables demand forecasting and transparent, auditable, and equitable distribution.	[92]
		Sustainability	Generates synthetic data to optimize inventory, logistics, and risks for effective, sustainable supply chains.	[93]
	4	Order Management	An e-commerce conditional GAN produces realistic orders, allowing for data-driven order management and analysis.	[88]
		Product Lifecycle Management	WGANs use topology-optimized data from the Product Lifecycle Management design phase to produce innovative, high-performance automotive wheel designs.	[94]
		Resiliency and Risk Management	Improves resilience, contingency planning, and anomaly detection by simulating demand shocks, supplier disruptions, and cyberattacks.	[90]
		Asset Management	Enables reliable, transferable fault diagnosis under a variety of circumstances by synthesizing realistic fault-condition sensor data.	[95]
		Quality Management	Enhances visual inspection and inline quality control by augmenting defect-free datasets and supporting pixel-level unsupervised anomaly detection.	[96]
	5	Product Returns	DE-GANs create images with a variety of defects, allowing for reliable automated triage in reverse logistics under data scarcity.	[97]
		Remanufacturing and Refurbishment	DE-GANs allow CNNs to evaluate the acceptability of reuse and automate inspection decisions by producing realistic worn-part images.	[97]
		Repair	Root cause localization and targeted repair decisions are made possible by the GAN-VAE hybrid’s learning of causal latent relationships.	[98]
VAE	0	Optimization	IoT data is analyzed by VAEs to optimize demand, identify abnormalities, inform reuse choices, and guarantee safe, open supply chains.	[99]
	1	Demand Planning	Extracts latent demand features, creates scenarios, models uncertainty, and enhances planning under volatile conditions.	[100]
	2	Inventory Management	VAEs learn latent patterns to detect uncontrolled inventory anomalies and prevent stockouts and overstocking.	[101]
		Transportation and Logistics	Efficiency and analytical performance are increased by compressing high-dimensional route, traffic, and delivery data into the latent space.	[102]
			VAEs train probabilistic latent models for effective route optimization from structured and unstructured transportation data.	[103]
	3	Sustainability	Pareto-optimal cost-emission sourcing trade-offs are made possible by sustainable scenarios produced by probabilistic latent demand models.	[104]
		Ethics	Facilitates data-driven, rule-free anomaly identification, enabling more equitable, objective, and fair monitoring.	[105]
	4	Product Lifecyle Management	Provides quick prototyping, validation, and design space exploration by producing several variations in product design.	[106]
		Resiliency and Risk Management	Creates scenarios by learning demand distributions with sustainability aspects, enabling flexible and reliable supply chain decisions.	[19]
		Quality Management	Learns structured 2D latent sensor representations to provide real-time part quality prediction and monitoring.	[104]
	5	Customer Service	Utilizes brand interactions to model consumers’ latent emotional states and identifies factors that drive customer happiness to enhance service.	[107]
	5	Repair	TCN-VAE reconstructs features from multivariate sensor data while investigating latent variables to facilitate causal fault analysis.	[108]
Transformer-based Models	0	Data Analytics	Transformer’s NLP capabilities enhance data quality and minimize labor by automating extraction from unstructured documents and analyzing sentiment.	[109]
		Optimization	Transformers enable context-aware, comprehensive network optimization by using self-attention to balance supply chain components.	[110]
	1	Identification of Demand	Utilizes self-attention to extract temporal patterns to determine demand trends, spikes, and anomalies.	[111]
		Forecasting	Utilizes an encoder–decoder architecture with probabilistic decoding to forecast, quantify uncertainty, and facilitate accurate decision-making.	[111]
		Network Design	Employs self-attention to effectively optimize supply chain network designs, assess choices, and identify dependencies.	[112]
		Demand Planning	Uses self-attention to identify long-range dependencies to make better supply chain and demand planning decisions.	[113]
	2	Sourcing	GPT-4 generates expert-aligned judgments from natural language by automating supplier ranking using AHP-weighted synthesis.	[114]
		Procurement	Uses self-attention to align solutions with procurement requirements by integrating supplier capabilities and documentation.	[115]
		Manufacturing and Scheduling	Manufacturing signals are fused by self-attention and multi-head attention, allowing for precise scheduling, process optimization, and decision support.	[109]
		Inventory Management	Uses self-attention to enable precise, context-aware forecasting for inventory optimization by modeling multimodal demand.	[116]
		Warehousing	Transformers dynamically prioritize retrieval jobs in warehouses using GNN-encoded item and layout embeddings.	[117]
		Transportation and Logistics	Self-attention on GNN-encoded supply networks enables adaptive, globally optimal route planning with meta-reinforcement learning.	[110]
		Transportation and Logistics	With GAN refinement, globally optimized, resource-aware path planning is made possible by self-attention on GNN-processed robot and cargo data.	[118]
	3	Sustainability	Multi-head self-attention enables real-time ethical supply chain optimization, tracking origin, emissions, and unsustainable behaviors.	[119]
		Ethics	The SustAI-SCM transformer enables ethical, transparent, and compliance-aware procurement by using multi-head self-attention on supplier data.	[119]
	4	Order Management	Transformers automate order management and enable precise, real-time operations by integrating unstructured and ERP data.	[120]
		Product Lifecycle Management	Uses Natural Language Processing (NLP) to comprehend textual data in order to analyze market research reports, consumer reviews, and social media.	[121,122]
			Contributes to the development of novel concepts and ideas for original concept design.	[123]
		Resiliency and Risk Management	Sequence modeling and knowledge-graph grounding are used to extract context-aware, real-time data from unstructured text.	[124]
		Asset Management	An encoder–decoder Transformer forecasts multi-step volatility, enhancing portfolio risk-adjusted performance, while the encoder predicts returns for the following week.	[125]
		Quality Management	Self-attention can detect unusual patterns that might point to shortcomings or inefficiencies.	[126]
	5	Product Returns	OpenTransformer can forecast the likelihood of a product return by extracting acoustic cues from anchor speech in live feeds.	[127]
	5	Customer Service	Uses chatbots and virtual assistants to provide accurate answers 24/7, automate repetitive queries, and provide customized suggestions.	[128]
	5	Remanufacturing and Refurbishment	Hybrid transformer-attention guides Laser Stock Peening remanufacturing for maximal lifespan by predicting the bearing’s remaining useful life.	[129]
	5	Warranty	CRFormer enables early, data-driven reliability and warranty management by forecasting long-term automotive failures from short-term claims.	[130]
	5	Repair	A transformer uses self-attention and temporal encoding for adaptive, real-time repair prediction after ingesting multimodal maintenance data.	[131]
Flow-Based	1	Identification of Demand	By capturing variability and uncertainty, these models provide accurate demand estimates and guarantee effective training across product lines.	[132]
	1	Forecasting	Conditional flows combine digital traffic, IoT data, and transactions to identify market abnormalities and estimate demand.	[132]
	1	Demand Planning	Learns calibrated multivariate demand distributions to effectively make risk-aware safety stock and service-level decisions.	[133]
	2	Warehousing	Complex demand distributions are modeled by exact likelihood computation, which helps with warehouse and inventory management decisions.	[132]
	3	Sustainability	Normalizing Temporally Flows enable efficient hardware reuse or resale and minimize e-waste by precisely modeling decommissioning dates in reverse cloud supply chains.	[134]
	3	Order Management	Temporal Normalizing Flow models are used in flow-based generative AI to model order-event sequences for precise fulfillment times.	[134]
	4	Resiliency and Risk Management	Risk-aware SCM disruption modeling, early warning, and resilient planning are made possible by integrating NKF and ACNet flows.	[122,135]
	4	Quality Management	Normalizing Flows describe multivariate process data for sensitive anomaly detection and real-time quality assurance.	[136]

provides an illustrative mapping between GenAI models and supply chain decision-making areas based on some highly cited studies. The table illustrates how different GenAI models have been applied to a wide range of functions, including forecasting, optimization, sourcing, inventory management, transportation, sustainability, and resilience. This table also highlights the diversity of methodological approaches and decision contexts addressed in the existing literature.

6. Bibliometric Analysis

In this section, we conduct various analyses on the collected articles, including keyword co-occurrence analysis, temporal analysis, density analysis, and Supervised Embedding Visualization.

Figure 3 illustrates the number of publications per year. The trend highlights gradual growth from 2020 to 2023, followed by a surge in subsequent years. This indicates an increasing interest in GenAI since 2023 and the importance of using GenAI to address SCM challenges, such as resilience and efficiency amid global disruptions, the introduction of advanced language models, and the increased importance of sustainability. Additionally, the significant increase in publications after 2023 reflects the rapid evolution and growth of a wide variety of GenAI models, regardless of specific applications within SCM [137]. For instance, the public release of LLMs, particularly ChatGPT in 2022, substantially increased academic and industrial interest in GenAI, leading researchers to explore its applications across numerous business domains [138]. Thereafter, other factors, such as the availability of open-source models and increased computational accessibility and efficiency, facilitated the growth of these models and their applications.

The op countries based on the number of publications and total citations are given in

Table 3. Top ten leading countries for research on Generative AI in SCM by the number of publications and total citations.

Rank	Country	Number of Publications	Total Citation	Average Citations
1	China	370	6378	17.24
2	USA	62	1719	27.73
3	Korea	26	558	21.46
4	India	26	318	12.23
5	England	18	558	31.00
6	Canada	16	412	25.75
7	Australia	14	390	27.86
8	Germany	13	289	22.23
9	Taiwan	13	150	11.54
10	Spain	10	347	34.70

. In addition, the top four cited articles for each country are shown in

Table 4. Top four cited articles for each country.

Country	Paper	Article Title	Publication Year	Citation
China	[139]	Deep generative modeling for mechanistic-based learning and design of metamaterial systems	2020	365
	[140]	Urban ride-hailing demand prediction with multiple spatiotemporal information fusion network	2020	153
	[141]	A temporal fusion transformer for short-term freeway traffic speed multistep prediction	2022	133
	[142]	Defect-aware transformer network for intelligent visual surface defect detection	2023	122
USA	[143]	A Trustworthy Privacy Preserving Framework for Machine Learning in Industrial IoT Systems	2020	284
	[144]	Predicting field production rates for waterflooding using a machine learning-based proxy model	2020	124
	[145]	Cooperative lane control application for fully connected and automated vehicles at multilane freeways	2020	93
	[146]	Bayesian Spatiotemporal grAph tRansformer network (B-STAR) for multi-aircraft trajectory prediction	2022	80
Korea	[147]	Automated defect inspection system for metal surfaces based on deep learning and data augmentation	2020	285
	[148]	Adversarial Defect Detection in Semiconductor Manufacturing Process	2021	50
	[149]	Transformer-Based Reinforcement Learning for Scalable Multi-UAV Area Coverage	2024	37
	[150]	Federated PCA on Grassmann Manifold for IoT Anomaly Detection	2024	24
India	[151]	Assessing the nexus of Generative AI adoption, ethical considerations and organizational performance	2024	100
	[152]	Vehicular Trajectory Classification and Traffic Anomaly Detection in Videos Using a Hybrid CNN-VAE Architecture	2022	81
	[153]	Global-Local Attention-Based Butterfly Vision Transformer for Visualization-Based Malware Classification	2023	26
	[154]	PDSMV3-DCRNN: A novel ensemble deep learning framework for enhancing phishing detection and URL extraction	2025	14
England	[155]	Blockchain-enabled supply chain traceability-How wide? How deep?	2023	133
	[156]	A combined machine learning algorithms and DEA method for measuring and predicting the efficiency of Chinese manufacturing listed companies	2021	125
	[157]	A Vision Transformer Approach for Traffic Congestion Prediction in Urban Areas	2023	110
	[158]	Deep Transfer Learning With Self-Attention for Industry Sensor Fusion Tasks	2022	38
Canada	[159]	An attention-based multiscale transformer network for remote sensing image change detection	2023	139
	[160]	On the application of machine learning for defect detection in L-PBF additive manufacturing	2021	100
	[161]	V2VFormer++: Multimodal Vehicle-to-Vehicle Cooperative Perception via Global-Local Transformer	2024	47
	[162]	Automated Defect-Detection System for Water Pipelines Based on CCTV Inspection Videos of Autonomous Robotic Platforms	2024	34
Australia	[163]	Bidirectional Spatial–Temporal Adaptive Transformer for Urban Traffic Flow Forecasting	2023	163
	[164]	Artificial Intelligence-Based Power Transformer Health Index for Handling Data Uncertainty	2021	69
	[165]	Vision Transformer Inspired Automated Vulnerability Repair	2024	45
	[166]	An unsupervised defect detection model for a dry carbon fiber textile	2022	33
Germany	[167]	ChatGPT and generative artificial intelligence: an exploratory study of key benefits and challenges in operations and supply chain management	2024	197
	[168]	Unsupervised pre-training of graph transformers on patient population graphs	2023	21
	[169]	Designing flexibility procurement markets for congestion management investigating two-stage procurement auctions	2022	16
	[170]	Design and evaluation of an Autonomous Cyber Defense agent using DRL and an augmented LLM	2025	13
Taiwan	[171]	Improving Generalization in Reinforcement Learning-Based Trading by Using a Generative Adversarial Market Model	2021	48
	[172]	VR-enabled engineering consultation chatbot for integrated and intelligent manufacturing services	2022	44
	[173]	A new ViT-Based augmentation framework for wafer map defect classification to enhance the resilience of semiconductor supply chains	2024	13
	[174]	Application of retrieval-augmented generation for interactive industrial knowledge management via a large language model	2025	12
Spain	[175]	Automated Road Damage Detection Using UAV Images and Deep Learning Techniques	2023	84
	[176]	Multi-Transformer: A New Neural Network-Based Architecture for Forecasting S&P Volatility	2021	75
	[177]	Transformer-Based Models for Automatic Identification of Argument Relations: A Cross-Domain Evaluation	2021	66
	[178]	Integrated Multi-Head Self-Attention Transformer model for electricity demand prediction incorporating local climate variables	2023	33

.

6.1. Keywords Co-Occurrence Analysis

To illustrate the relationships among various keywords/terms used in the collected papers and analyze the relationships among them, we use the widely used VOSviewer 1.6.20. This tool enables us to create the co-occurrence network of keywords to reveal any potential clusters existing in the literature based on collected keywords. If a keyword appears in different papers’ titles and/or abstracts, it is defined as co-occurring [179]. In the network, links indicate co-occurrence associations, whereas nodes represent keywords, with a larger node size indicating a higher frequency of occurrence.

To prepare for this analysis, we manually reviewed author keywords to improve consistency. To this end, all keywords were converted to a consistent format by standardizing capitalization, singular and plural forms, and abbreviations. We also merged all synonymous terms into a single keyword and removed duplicates.

To improve the interpretability of the network, only keywords appearing in at least three publications were included in the analysis. This resulted in a final set of 68 keywords. The keyword co-occurrence network was constructed using the Full Counting method, indicating that each occurrence of a keyword within a publication contributed equally to the co-occurrence network. Then, VOSviewer automatically generated the network layout and identified keyword clusters using its visualization of similarities (VOS) mapping. Figure 4 illustrates the resulting keyword co-occurrence network and the relationships among keywords identified from the 692 reviewed articles.

Figure 4 illustrates the clusters that are formed, as well as the relationships between the keywords generated by VOSviewer based on the 692 papers. The figure shows six main clusters:

• The yellow cluster includes publications primarily related to the applications of Transformer-based models designed for Natural Language Processing. For transportation, several studies utilize the Transformer architecture to model intricate traffic flow dynamics. Some examples of these models include the Adaptive Spatial–Temporal Transformer Network (ASTTN) [180], Routeformer [181], the Graph-Enhanced Spatial–Temporal Transformer (GE-STT) [182], and the Cross-dimensional global interactive transformer (CDGIT) [183] specifically designed to capture dynamic spatial–temporal features and inter-series dependencies for improved traffic forecasting. Other models, such as the Motion-Inspired Spatial–Temporal Transformer (MSTFormer) [184] apply a motion-inspired Transformer for vessel trajectory prediction. A Heterogeneous Hypergraph Transformer network with Cross-modal Future Interaction (HHT-CFI) proposed by X. Zhou et al., 2025 [185] uses a hypergraph Transformer for multi-agent trajectory prediction. Transformer-based models are also employed for energy consumption forecasting using the Temporal Kolmogorov-Arnold Transformer (TKAT) [186], oil production forecasting via the Informer model [187], and lithium-ion battery State of Health (SOH) prediction through an ADTC-Transformer framework [188].
• The green cluster demonstrates the growing application of Transformer-based models to various optimization problems. The cluster links keywords such as costs, analysis, decision-making, data, scheduling, and cloud computing, among others, and illustrates how AI-driven optimization strategies can enhance productivity in a variety of computing contexts. For example, in the Hybrid-Prediction integrated Planning (HPP) proposed by [189], the authors use the Transformer architecture for interactive open- and closed-loop planning. The algorithm outperforms current algorithms on the Waymo Open Motion Dataset (WOMD) and Car Learning to Act (CARLA) benchmarks with greater prediction consistency. Transformer-based architectures are also used by [190] to develop reliable AI systems for vital IoT infrastructures, guaranteeing transparent decision-making in dynamic sensor network situations without sacrificing efficiency.
• Recent developments in Transformer- and autoencoder-based models in transportation, agriculture, and environmental monitoring are illustrated in the purple cluster. This cluster focuses on the application of artificial intelligence to improve accuracy across domains. For example, ref. [191] integrates the Transformer-Photovoltaic (TransPV) model with Vision Transformers to improve photovoltaic panel detection in a variety of structures to support sustainable energy mapping.
• The red cluster represents the advancement of research on Generative AI driven by LLMs and GANs across a variety of fields. These developments improve supply chain performance and green innovation in small- and medium-sized e-commerce enterprises [192]. Another major advancement is agent-driven generative semantic communication in 6G networks, improving the accuracy and efficiency of GenAI in the development of digital infrastructure and societal change [192,193].
• The orange cluster highlights the roles of attention mechanisms, self-attention, and Markov decision processes in various industrial processes, such as manufacturing and packaging. For example, ref. [194] uses a diffusion-based Transformer (DIFFormer) in conjunction with other models, such as deep reinforcement learning, to model the dynamic flexible job-shop scheduling problem as a Markov decision process. The model uses graph-based encoding of operations and machines, integrates selective rescheduling strategies, and applies the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm for efficient policy learning to overcome the drawbacks of metaheuristic methods.
• The blue cluster encompasses data science, risk management, decision-making, quality, blockchain, collaboration, and neural networks, highlighting the intersection of GenAI applications across various domains. For instance, ref. [195] uses GAN-based data augmentation for imbalanced spectrum data in the field of materials science to improve classification accuracy. In other research conducted by [196], the authors integrate blockchain and formal verification to incorporate verifiable ethical limitations in AI agents for healthcare supply chain processes.

The keyword co-occurrence network indicates that Transformer-based models have become the central research topic by connecting multiple SCM task levels. Unlike earlier GenAI models that were primarily developed for specific analytical tasks, Transformer-based models appear across transportation, forecasting, optimization, inspection, logistics, and decision-making, indicating that recent research is shifting toward more general-purpose GenAI models that are capable of supporting multiple supply chain functions. They have strong links to prediction, temporal analysis, and convolutional neural networks, which highlights that the algorithms are primarily used for advanced deep learning methods for time-dependent and predictive tasks. This is also observed in

Table 5. Top 10 keywords by occurrences.

Rank	Keyword	Occurrences	Total Link Strength	Cluster
1	transformers	218	424	Yellow
2	data science	119	392	Blue
3	inspection and defect detection	73	195	Blue
4	computer science	37	156	Purple
5	prediction	43	117	Yellow
6	transportation	47	109	Blue
7	temporal analysis	33	105	Yellow
8	convolutional neural network	35	101	Yellow
9	optimization	37	97	Green
10	image detection	26	85	Purple

, which summarizes the top 10 keywords in the co-occurrence network. The prominence of data science, inspection and defect detection, and transportation indicates that GenAI is being actively applied to practical operational problems, particularly in monitoring and logistics-related contexts, while the presence of the optimization keyword reflects a growing emphasis on decision-oriented applications. Additionally, the extensive connections among clusters suggest that GenAI research is becoming increasingly interdisciplinary. Rather than addressing isolated SCM problems, recent studies combine forecasting, optimization, transportation, quality management, and data analytics via integrated decision-support systems. On the other hand, topics such as cybersecurity, ethics, and responsible AI do not explicitly appear in the network. This suggests that future research should place greater emphasis on the security, resilience, and sustainable deployment of GenAI in supply chains. Considering the framework and keyword co-occurrence network reveals that the literature remains concentrated on Levels 0 and 2 tasks, whereas comparatively limited attention has been devoted to post-sale and reverse logistics activities represented in Level 5.

6.2. Temporal Analysis

The analysis identifies three main phases of research since 2020, as shown in Figure 5.

• The blue phase focuses on research conducted between approximately 2020 and 2021, during which traditional machine learning techniques and Generative AI were incorporated into SCM for structured data-based demand forecasting, inventory control, and logistics optimization [164,197]. Improving automation and operational efficiency enabled researchers to create data pipelines and a predictive analytics foundation for the adoption of GenAI [160]. Initial works around 2020 also emphasized foundational generative techniques for data scarcity in SCM, for example, utilizing VAEs to enable interpolation between microstructures and encoding meaningful patterns of variation in geometries and properties [139]. Another example is the Causal-Aware GAN method to batch incomplete traffic tensors in consecutively missing cases with a fast recovery process [198]. In 2021, research focused on the increased usage of GANs, VAEs, multi-Transformers, and hybrid ML models to enhance efficiency in defect detection, demand forecasting, scheduling, and anomaly management in the manufacturing, transportation and energy sectors [160,176,199].
• The green phase includes studies conducted in approximately 2022 and 2023, with a focus on the development and usage of various Transformer-based models. Among different models, ref. [200] proposed the Multi-Feature Integration Network (Swin-MFINet) to provide precise pixel-level manufacturing defect identification, which is helpful for quality management. Ref. [130] proposed the Complementary Reliability perspective Transformer (CRFormer), which reduces warranty and production expenses while improving automobile reliability prediction using short-term data. In 2022, research focused on Transformer-based and hybrid deep-learning models to integrate attention mechanisms and federated learning to address data scarcity, privacy, and real-time optimization in manufacturing, transportation, and energy systems. An example is the FedAnomaly framework for privacy-preserving anomaly detection in cloud manufacturing [201]. Research conducted in 2023 focused on the experimental and solution-oriented usage of Transformer-based deep learning models for infrastructure and framework development. An example is the Variational Mode Decomposition Sample Entropy Transformer (VMDSE-Tformer) to predict resource demand in cloud data centers, with implications for more effective resource management and service delivery [202].
• The yellow phase describes research done in 2024 and 2025 that highlights how Generative AI may solve problems in various domains based on improving Transformer-based models. For example, ref. [110] integrates Graph Neural Networks, self-attention mechanisms, and meta-reinforcement learning (Meta-RL) to handle route optimization path planning in both static and dynamic contexts. A Transformer-based deep reinforcement learning (T-DRL) approach, a so-called two-stage encoder–decoder architecture with multi-head attention, is proposed by [203]. The algorithm is used for multi-objective, multi-hydropower reservoir operation optimization by extracting complex spatial–temporal dependencies and producing context-aware optimal decisions for water supply, power generation, and ecological protection. Ref. [149] developed a Transformer-based deep multi-agent reinforcement learning (T-MARL) framework that is used for cooperative, scalable decision-making for the multi-unmanned aerial vehicle area coverage problem in intelligent transportation systems. Research conducted in 2024 focused on the problem-solving ability of Generative AI to accomplish tasks by leveraging Generative AI models to enhance efficiency, interpretability, and adaptability in complex systems such as transportation and drug delivery [204,205]. In 2025, some studies focused on a solution-oriented approach to Transformers, emphasizing practical advancements in AI-driven frameworks for data sharing and anomaly detection [154,206,207].

The temporal analysis reveals interesting observations. Beyond the chronological development of GenAI, this analysis illustrates a clear transition in research. Early studies primarily focused on adapting existing machine learning and generative models to individual supply chain tasks, while more recent work increasingly emphasizes foundation models capable of supporting multiple decision-making tasks. This advancement suggests that GenAI research is evolving from algorithm development toward integrated intelligent decision-making systems. Furthermore, the analysis demonstrates the rapid emergence of Transformer-based architectures after 2022. Their crucial position with strong connections to forecasting, transportation, inspection, manufacturing, and data science indicates that Transformer-based models have become the dominant foundation for recent GenAI research in SCM. This trend reflects the growing preference for GenAI models that can handle diverse data types and support multiple supply chain decision-making tasks. The appearance of large language models and prompt engineering also suggests that research is beginning to move beyond predictive analytics toward natural language reasoning and knowledge extraction. Although these topics remain relatively new, they are likely to become increasingly important for procurement and supplier collaboration. Finally, the analysis indicates that recent research is shifting from improving predictive accuracy toward context-aware decision-making to integrate reasoning, attention mechanisms, and reinforcement learning to address complex operational decisions.

6.3. Density Analysis

Authors have largely focused on topics related to Transformers, data science, prediction, computer science, machine learning, artificial intelligence, large language models, neural networks, production, and transportation, as shown in Figure 6. Transformer-based models utilize temporal analysis and spatiotemporal modeling to improve operations in a variety of fields. For example, in transportation, logistics, urban planning, and traffic management, models such as the Sandwich Transformer are used for urban traffic [208], and the Deep Multi-View Spatiotemporal Virtual Graph Neural Network (DMVST-VGNN) is used for urban vehicle scheduling [209].

Transformer-based models can also be used to optimize processes in quality, scheduling, decision-making, and teamwork. The CAME-Transformer with compressed attention improves machinery dependability through multimodal signal fusion [210]. The Swin Transformer-based multi-feature integration network (Swin-MFINet) facilitates quality control operations in textile and steel plants [200]. The Incomplete Multimodal Transformer with Deep Autoencoder (IMTDAE) fortifies cognitive digital twins with integrated GPT used to automate manufacturing decisions [211,212]. Virtual inspections are effectively performed by Quality Transformers with federated learning [213]. The zero-shot class knowledge graph (ZS-CKG) can be used to recognize surface defects for quality control in manufacturing systems [214]. The algorithm obtained an improvement of over 33% compared to the best competing methods in the literature.

Transformer-based models can be used to improve sustainability, planning, and energy efficiency as well. In sustainable energy planning, the Long Short-Term Memory (LSTM) Transformer fusion with the Particle Swarm Optimization algorithm lowers the Mean Absolute Error of prediction by 30% [215]. Other examples are the Hybrid Transformer-LSTM for flood avoidance in mining [216], the Transformer-based deep reinforcement learning (T-DRL) for enhancing power generation [203], the 3DPECP-Net for enhancing low-energy additive manufacturing [217] and the Bidirectional Encoder Representations from Transformers for Batteries (BERTtery) for preventing EV battery thermal runaway [218].

Transformer-based models are essential for anomaly detection, intrusion detection, and cybersecurity. The Unsupervised Federated Hypernetwork Method for Distributed Multivariate Time Series Anomaly Detection and Diagnosis (uFedHy-DisMTSADD) was proposed by [219] for multivariate time series anomaly detection. For intrusion detection, a transformer-based model combining a Convolutional Neural Network (CNN) and a Recurrent Neural Network (RNN) developed by [220] attained 100% accuracy in the diagnosis of ensemble traffic [221].

The reviewed literature shows that the success of LLMs has encouraged researchers to explore their use in supply chain applications, particularly by using large amounts of textual information. To this end, recent research has shifted toward Transformer-based approaches primarily because they are effective at analyzing documents, reports, contracts, customer feedback, and other textual information, while also supporting forecasting, quality management, and data analytics tasks, which often require analyzing large volumes of data and textual information. In contrast, models such as GANs and VAEs are more suited for generating synthetic data, anomaly detection, and scenario simulation, which have received comparatively less attention in the current literature. Based on the keyword analysis, the 10 most cited papers for the application of Transformers are summarized in

Table 6. Ten most cited papers on Transformer-based models.

No	Authors	Article Title	Article Keywords	Description	Country	Citation
1	[163]	Bidirectional Spatial–Temporal Adaptive Transformer for Urban Traffic Flow Forecasting	Dynamic halting mechanism; spatial-temporal; transformer; urban traffic forecasting	Proposes bidirectional spatial–temporal adaptive transformer (Bi-STAT), which uses adaptive modules and dual decoders for past recollection to improve generalization for traffic forecasting.	Australia	163
2	[159]	An attention-based multiscale transformer network for remote sensing image change detection	Change detection; attention mechanism; Transformer; multiscale	Introduces the attention-based multiscale transformer network (AMTNet) for bi-temporal change detection in remote sensing, which uses feature exchange between Siamese branches.	Canada	139
3	[222]	Light-weight federated learning-based anomaly detection for time-series data in industrial control systems	Anomaly detection; ICS; federated learning; autoencoder; Transformer; Fourier	Proposes a federated learning, autoencoder, and Transformer architecture with Fourier mixing for robust, lightweight, and fast anomaly detection in industrial control systems.	France	134
4	[141]	A temporal fusion transformer for short-term freeway traffic speed multistep prediction	Deep learning; temporal fusion transformer; traffic speed; multistep prediction	Adopts the Temporal Fusion Transformer (TFT) for accurate short-term freeway speed prediction for intelligent transportation management to plan travel routes.	China	133
5	[157]	A Vision Transformer Approach for Traffic Congestion Prediction in Urban Areas	Convolutional neural networks; Transformers; transportation; Roads; feature extraction; deep learning; computational modeling; vision transformers; deep learning; intelligent transportation system; long-short-term-memory (LSTM); traffic congestion prediction	Utilizes a Vision Transformer with Convolutional Nerual Networks for citywide traffic congestion prediction by tokenizing features for LSTM input.	England	110
6	[218]	Battery fault diagnosis and failure prognosis for electric vehicles using spatiotemporal transformer networks	lithium-ion battery; fault; failure; diagnosis and prognosis; Transformer; field data	Introduces BERTtery (Bidirectional Encoder Representations from Transformers for Batteries), a specialized Transformer for predicting EV battery faults from early-cycle charging data by capturing multiscale spatiotemporal signals.	China	99
7	[223]	LSTTN: A Long-Short Term Transformer-based spatiotemporal neural network for traffic flow	Traffic forecasting; spatiotemporal modeling; long-short term forecasting; Transformer; Mask Subseries Strategy	Proposes Long-Short Term Transformer-Based Network (LSTTN) framework for traffic flow using masked subseries pretraining and dilated convolutions to extract long/short-term trends.	China	85
8	[146]	Bayesian Spatiotemporal grAph tRansformer network (B-STAR) for multi-aircraft trajectory prediction	Multi-agent trajectory prediction; graph neural network; Transformer; air traffic management	Introduces a Bayesian Spatiotemporal grAph tRansformer (B-STAR) for multi-agent trajectory prediction under uncertainty.	USA	80
9	[224]	Pavement crack detection with hybrid-window attentive Vision Transformers	Crack detection; pavement distress; attentive vision transformer; feature attention; high-resolution network; deep learning	Proposes CrackFormer, a hybrid-window attentive Vision Transformer for pavement crack detection, enhancing semantic details and saliency via weighted multi-head attention.	China	76
10	[225]	Multi-attribute adaptive aggregation transformer for vehicle re-identification	Vehicle re-identification; transformer; multi-attribute adaptive aggregation; multi-sample dispersion	Introduces vehicle attribute transformer (VAT) for re-identification, embedding color/model/viewpoint features and using adaptive aggregation for multi-attribute weighting useful for intelligent transportation systems	China	76

.

7. Supervised Embedding Visualization of Generative AI for Supply Chain Tasks

In this section, we use Uniform Manifold Approximation and Projection (UMAP) to conduct a Supervised Embedding Visualization based on the collected papers. UMAP is a dimensionality reduction technique often used to visualize high-dimensional text embeddings [226]. Using this method, we can observe how various AI models are positioned in supply chain research and which GenAI models are most strongly associated with specific supply-chain tasks. UMAP can maintain both local and global data structures, which facilitates the presentation of theme patterns in keyword clustering. The process of Supervised Embedding Visualization is shown in Figure 7. All preprocessing, Supervised Embedding Visualization, and UMAP analyses were implemented in Python 3.13. We group the collected papers based on two criteria: (1) Generative AI model type: GAN, VAE, Transformers, and Flow-Based GenAI, (2) Supply Chain Level: Levels 0–5. Before Supervised Embedding Visualization, the abstracts were preprocessed using the Natural Language Toolkit (NLTK) in Python. In the preprocessing pipeline, we converted all text to lowercase, removed punctuation, and eliminated English stop words. A predefined vocabulary consisting of Generative AI model names, their common variants and synonyms, and supply chain decision-making keywords was constructed manually based on the literature (see Appendix B). Duplicate keywords were removed, and synonymous terms referring to the same GenAI architecture were merged into a single representation to improve consistency.

To represent the semantic content of each abstract, sentence embeddings were generated using the all-MiniLM-L6-v2 model from the SentenceTransformers library. Then, these embeddings were projected into a three-dimensional space using UMAP implemented in Python through the umap-learn package. UMAP was implemented using the Python umap package with the following parameters: n_neighbors = 15, min_dist = 0.1, metric = cosine, and n_components = 3. We used these parameter values to ensure local neighborhood relationships while producing stable and interpretable low-dimensional representations. Each article was assigned a dominant GenAI model and a dominant supply chain decision category based on the frequency of predefined keywords in its abstract. The resulting visualization provides an exploratory representation of the semantic proximity and separation among predefined GenAI model–SCM level categories. Considering four GenAI models and six levels for SCM tasks, it creates various combined labels (GenAI Model—SCM Level) as shown in Figure 8. The chart includes all the labels with at least one matching article found for them. For this visualization, one model and one SCM-level label were required for each article. Therefore, a dominant display label was assigned based on the study’s primary research objective, principal methodological contribution, and main reported application. For three articles, no dominant category could be determined; therefore, they were excluded from this visualization.

The UMAP chart in Figure 8 shows three main observed groups:

Group 1: This group highlights the use of Transformer-based models for SCM Level 2 tasks, such as logistics, manufacturing, and scheduling. Reference [227] examines the use of Transformer-based models in logistics and transportation, demonstrating that a deep reinforcement learning framework combined with a Transformer architecture outperforms traditional methods in optimizing unmanned aerial vehicle routing for cooperative missions. In [228], the authors show that digital twins enhance the processing of steel hull parts and the distribution of resources in shipbuilding during the manufacturing and scheduling phases. Recurrent neural networks, such as the attention-based TFT model, improve restaurant labor planning for scheduling by identifying temporal dependencies [229]. To handle complex job scheduling in multi-core systems, ref. [230] combines Transformer-based models (TransformX) with reinforcement learning in a federated setting. Ref. [110] suggests a technique for route optimization in supply chains that combines Graph Neural Networks (GNNs), self-attention mechanisms, and meta-reinforcement learning (Meta-RL). This group also includes a few articles on the use of VAEs for SCM Level 0 and Level 2 tasks. To improve and optimize the safety and efficiency of autonomous vehicles by enabling socially responsive, context-aware decisions like yielding or merging, ref. [231] investigates a VAE-based Driving Prior Learning model to infer human-like behaviors from actual driving data. A complementary Transformer-based policy network is also used to capture spatial–temporal interactions with surrounding traffic.

Group 2: The second group focuses on using Transformer-based models for SCM Level 0 tasks. This group illustrates the growing usage of Transformer-based models for data analytics and optimization. They improve demand forecasting when combined with graph networks [232], improve congestion prediction via hybrid CNN–Vision Transformer models [157], and enable accurate spatiotemporal traffic forecasting by capturing short- and long-term dependencies [223]. They demonstrate their versatility in providing scalable, comprehensible solutions across transportation, manufacturing, and energy domains by enabling early battery defect detection [218], improving vehicle re-identification through multi-attribute attention [225], and refining surface defect identification using global and local features [200,233]. For example, a variant of the Series Conversion Normalization Transformer (SC Nor-Transformer) is used to tackle timing bias and heterogeneity to enhance distributed multivariate time-series anomaly detection [219]. This allows for precise fault localization and detection in financial and industrial systems. Utilizing LLMs, the authors propose the xTP-LLM model that converts multimodal traffic data into natural language, producing precise and comprehensible traffic flow forecasts that support traffic management optimization [234]. Additionally, LLMs are used to automatically create knowledge graphs and ontologies from technical documentation, enabling optimization tasks like urban choice analysis and multimodal freight, and facilitating data integration [235].

Group 3: The third group includes several subgroups. The first focus is on using GANs for SCM Level 2 tasks. Sourcing, warehousing, and procurement are examples of potential GAN utilization. GANs can help improve supplier provenance, compliance, and authenticity through blockchain-enabled traceability [155]. They can also be used to enhance demand forecasting in conjunction with procurement planning [229]. Ref. [228] uses a digital twin to support smart picking and processing in shipbuilding and warehousing operations to optimize part picking and transport scheduling. Meanwhile, defect detection and vision-based defect classification [147] link manufacturing quality control with sourcing and procurement decisions by assessing supplier qualifications and quality of received parts at warehouses. Ref. [236] uses a conditional GAN (ICGAN) to handle data imbalance in wind turbine gearbox predictive maintenance. The second focus in this group is on using GANs for SCM Level 0 tasks. GANs exhibit moderate yet promising applicability in optimization and data analytics. For instance, ref. [205] proposes a large-scale mobile traffic simulation that facilitates multiscale temporal data creation for telecom resource planning and optimization by capturing both micro- and macro-behavior patterns. In reservoir waterflooding prediction in petroleum engineering, the Conditional Deep Convolutional GAN (cDC-GAN) model lowers computational costs, enhances fluid saturation forecasts, and permits optimization under uncertainty [144]. Ref. [198] presents a variant of GANs that learns spatial and temporal traffic patterns to generate realistic estimations for incomplete traffic measurements, thereby improving the reliability of downstream analytics and decision support. The third focus is on using Transformer-based models for SCM Levels 4 and 5 tasks. For example, ref. [130] presents CRFormer, a Transformer encoder-based model that predicts long-term component failure patterns by combining time- and mileage-based reliability perspectives from short-term automotive claim data. This allows for quality assurance planning, warranty cost optimization, and well-informed production decision-making.

The Supervised Embedding Visualization results show that Transformer-based models are primarily concentrated in SCM Level 2, indicating that recent research has focused on applying foundation models to core operational activities, such as transportation, logistics, manufacturing, and inventory management. In contrast, GAN models are predominantly associated with SCM Level 0, which indicates that their widespread use for analytical tasks includes synthetic data generation, optimization, and data augmentation. Other applications that involve VAEs and flow-based models appear more dispersed across the remaining SCM levels and continue to evolve.

Literature Analysis Based on SCM Levels and Tasks

Figure 9 illustrates the number of publications for each SCM task at each level. The majority of the articles considered GenAI for Level 0 tasks (Data Analytics and Optimization), illustrating Generative AI’s strong adoption for solving complex, computationally intensive problems. For Level 1, most of the research is focused on forecasting and network design problems, while GenAI’s potential for Demand planning has been considered in fewer articles. This indicates applications of GenAI for prediction, facility location, and network design optimization for strategic planning [209,232,237]. In Level 2, manufacturing, scheduling, transportation, and logistics tasks have been primarily considered in research. In Manufacturing and Scheduling, GANs are often used for fatigue prediction and defect diagnosis, VAEs are mainly used for defect detection, and Transformer-based models are used for scheduling optimization and prediction of a part’s useful life [160,238,239]. Sourcing, procurement, inventory management and warehousing remain potential research areas for the future. In transportation and logistics, GANS are used for traffic data imputation, VAEs are applied for vehicle trajectory prediction, and Transformer-based models are used for traffic flow prediction, delivery time estimation, port congestion prediction, autonomous vehicle routing, and vehicle arrival time estimation [152,157,240].

For Level 3, Sustainability in SCM is a growing research area, and authors have considered potential usages of GenAI for sustainability-related decisions [155]. However, ethics has received limited research, despite its increased importance for ethical sourcing options. LLMs can be helpful in minimizing overproduction and material waste, which is accomplished by optimizing inventory and by utilizing Natural Language Processing to analyze historical data [1]. In a study by [241], it was shown that the BERT model can be used for text classification to prioritize risk and provide a weighted risk ranking, which can enable organizations to quantify and mitigate environmental, social and governance risk in the supplier network. Similarly, the self-attention mechanism in Transformer-based models enables the analysis of text related to legal audits, such as forced labor, labor exploitation, and human rights, to automatically classify hidden risks for non-regulatory compliance [242]. In Level 4, authors have conducted significant research on quality management, while Asset Management, Resiliency, and Risk Management have been covered moderately, indicating growing areas of interest for applications of GenAI models. In quality management, GANs are mainly used for complaint, corrective, and preventive action management, VAEs are used for audits and inspections, Transformer-based models are used for fault diagnosis, and flow-based models are used for deviation and non-conformance parameter tuning [147,148,243,244]. On the other hand, order management and Product Lifecycle Management demands overall planning, and are potential areas for further research.

Level 5, compared to other levels, requires more attention for research. Even though reverse logistics and return management are essential to circular supply chain models, and they can increase customer satisfaction and save operating costs, daily tasks such as order administration, customer support, and product returns are still mostly unexplored.

Table 7. Summary of common SCM tasks, focused research areas, predominant GenAI techniques, existing research gaps, and managerial outcomes.

Generative AI	Common SCM Task	Focused Research Areas	Predominant GenAI Techniques	Existing Research Gaps	Managerial Outcomes
Generative Adversarial Networks	Data Analytics, Optimization, Manufacturing and Scheduling, Transportation and Logistics, Quality Management	Smart Manufacturing and Industrial Defect Detection, Intelligent Transportation Systems, Traffic and Autonomous Vehicles, Cybersecurity and Network Intrusion Detection	DCGAN (Deep Convolutional GAN), Conditional GANs (CGAN/ACGAN/CTGAN), WGAN-GP (Wasserstein GAN with Gradient Penalty), CycleGAN and Pix2Pix, TimeGAN/Sequence GANs	Research in Sourcing, Procurement, Warehousing, Product Lifecycle Management, Reverse Supply chains	Analyze image-based tasks, such as manufacturing defect images. CGANs are used for generating specific datasets for a specific type of cyberattack, a specific manufacturing defect, or tabular financial/credit risk data WGAN are used for network security and IoT intrusion detection because they provide more stable training and prevent “mode collapse” when generating complex network traffic patterns. TimeGANs are used for time-series data, such as generating synthetic vehicle trajectories, mobile user traffic patterns, or energy load scenarios.
Variational AutoEncoders	Data Analytics, Optimization, Manufacturing and Scheduling, Quality Management	Spatiotemporal modeling, topology and inverse design optimization, Circular Supply Chain IoT integration, and visual/signal quality inspection under constrained data conditions	Conditional VAEs, hybrid CNN/TCN-VAEs, surrogate VAE-ANN models, denoising and stacked VAEs, and VAE combined with One-Class SVM	Research in required in Demand/Capacity Planning, Network Design, Sourcing, Procurement, Warehousing, Ethics, Order Management, Asset Management, Repair, Customer Service, Warranty	Millisecond-level real-time predictions, significant energy and material savings, high-accuracy defect detection and enhanced operational sustainability.
Transformers	Forecasting, Data Analytics, Optimization, Manufacturing and Scheduling, Transportation and Logistics, Quality Management	Spatiotemporal prediction, text/knowledge extraction, network design, smart manufacturing, autonomous logistics, and visual inspection	TFT, PatchTST, Graph Transformers, BERT, LLMs/RAG, Decision Transformers, ViT, Swin, RT-DETR, and Hybrid CNN-Transformers.	Research in Product Returns, Warranty, Sourcing, Procurement, Warehousing, Refurbishment, Remanufacturing and Warranty	End-to-end supply chain improvement by enhancing forecasting accuracy, automating analytics and compliance monitoring. Optimizes operations and logistics. Improves manufacturing efficiency, strengthens transportation reliability and drives quality improvement while reducing costs and risks.
Flow Based	Forecasting, Demand Planning, Resiliency and Risk, Quality Management	Generates synthetic relational graph data and simulates systemic failures by generating full, high-fidelity synthetic copies of a logistics graph. Normalizing Flows capture deep temporal dependencies, seasonality and the conditional noise of historical data.	Normalizing Flows, Time Series forecasting	Flow-based generative AI has not been extensively researched. Further research will benefit from investigating how temporal analysis can help in predicting reverse supply chains.	Using flow-based, operations can transition from reactive scheduling to resilient, proactive demand mitigation.

summarizes the focused research, techniques, research gaps, and managerial outcomes for each GenAI method.

GenAI can also enhance resilience in the supply chain. With the integration of GenAI with various decision-making tasks in the supply chain, it can provide real-time visibility, data-driven insights, and seamless information sharing across external and internal sources. This enables synchronized partners to collaborate effectively and adapt swiftly to operational disruptions [245]. According to [246], Transformer-based models use self-attention to capture complex, non-linear patterns in time-series data, enabling supply chain resilience through rapid, real-time forecast adjustments during unexpected disruptions to prevent inventory imbalances and operational bottlenecks. In another study, a high-capacity generative Transformer-based model was used to simulate disruption scenarios for edge case stress testing using Transformer-based models to assist decision-making for dynamic negotiation and inventory reallocation [247]. An algorithm, the so-called Hyb-KAN, integrates Transformer-based models with deep learning architectures to improve the detection and classification of online cyber threats by analyzing contextual textual information [248]. Some Transformer-based models use self-attention mechanisms to analyze unstructured cyber threat intelligence and accurately classify different categories of cybersecurity threats [249], and Transformer-based models, such as BERT, analyze network traffic and system logs to detect sophisticated cybersecurity threats, including network intrusions, malware, and phishing attacks [250].

GenAI also relies on large volumes of timely data generated by digital supply chain technologies, and therefore, the reliability of GenAI depends not only on model performance but also on the security and resilience of the underlying digital infrastructure. IoT devices continuously collect operational information from sensors, vehicles, warehouses, production systems, and assets. In this regard, Sorour et al. [251] proposed a privacy-preserving intrusion detection framework for federated IoT networks that combines LSTMs with optimization and federated learning to detect cyberattacks while keeping sensitive data decentralized to improve the cybersecurity and resilience of IoT-enabled systems. Other infrastructures, such as cloud platforms, provide scalable storage and computational resources required for training and deploying large generative models, with edge computing supporting real-time inference for time-sensitive applications. Regarding this, Zhou et al. [252] developed the EC-TRL framework, which integrates Transformer-enhanced reinforcement learning with edge-cloud computing to optimize dynamic resource scheduling under changing workloads. The framework improves energy efficiency, load balancing, and task allocation to support cloud and edge infrastructures for GenAI applications in decision-making. In addition, blockchain has been recognized as a mechanism for improving data traceability, and trust among supply chain partners, thereby supporting secure data sharing and regulatory compliance. Goushlavandani et al. [253] proposed an intrusion detection framework for IoT networks based on blockchains, which secures the exchange of feature vectors and updates threat information across distributed nodes. The study demonstrates that integrating blockchain with AI-based intrusion detection improves data integrity, privacy, and resilience against cyberattacks in distributed digital infrastructures.

8. Conclusions and Future Research

In this paper, we used the PRISMA framework to conduct a comprehensive review of the uses of various GenAI models for SCM tasks, structured into six levels. Since the framework was developed conceptually from the literature, its validity was assessed through its application to classify the reviewed studies. The framework was able to accommodate all identified SCM decision-making tasks while allowing studies to be associated with multiple levels when appropriate, reflecting the interconnected nature of modern supply chain processes.

We applied analytical approaches, such as bibliometric analysis (keyword co-occurrence, temporal, and density analyses) and Supervised Embedding Visualization, to identify the focus of existing research and gaps for future research. The review examines the potential uses of GANs to generate synthetic data and discover patterns that enable organizations to better understand data characteristics and support analytical tasks. It also considers Transformer-based models, which have been widely used to analyze textual content and support data-driven decision-making across supply chain functions, leading to improved identification of discrepancies in textual information. Additionally, the applications of VAEs and flow-based models are discussed to provide a comprehensive understanding of their use in supply chain processes, such as hypothetical analyses, scenario generation, route optimization, and inventory management.

The keyword co-occurrence analysis identified six major research clusters, including Transformer-based models and their applications, recent developments in VAEs and Transformer-based models, advancements in GANs, developments in LLMs, and the role of attention mechanisms in GenAI. GenAI research in SCM is evolving from isolated algorithmic developments toward integrated decision-support systems. Transformer-based models now serve as the central technology linking forecasting, optimization, logistics, transportation, manufacturing, and quality management. In contrast, comparatively limited attention has been devoted to responsible AI, cybersecurity, sustainability, and reverse supply chain activities. These findings suggest that future research should focus on integrating GenAI across multiple SCM decision-making levels while simultaneously addressing secure, resilient, and sustainable deployments.

The temporal analysis revealed three descriptive temporal patterns in the introduction and development of these models. These patterns include the initial introduction of ML and GenAI models for SCM tasks, the development of various Transformer-based models, and the subsequent improvements and rapid expansion of advanced GenAI models for different SCM applications. The density analysis indicated that the primary focus of existing research is concentrated on Transformer-based models for various SCM tasks, such as data analytics, intelligent transportation, logistics, and spatiotemporal modeling. The Supervised Embedding Visualization identified three main clusters of GenAI applications for SCM tasks, particularly highlighting the use of Transformer-based models and GAN models for SCM Level 2 and Level 0 tasks. Finally, a task-level analysis demonstrated that the majority of the existing research has focused on the application of GenAI models in data analytics, optimization, forecasting, manufacturing and scheduling, transportation and logistics, and quality management, whereas comparatively fewer studies address areas such as cybersecurity, resilience, ethics, and sustainability.

Overall, the comprehensive review demonstrates that GenAI has shown promising capabilities for supporting several supply chain decision-making tasks, particularly in forecasting, data analytics, manufacturing, transportation, and logistics, where an increasing number of empirical studies have reported positive outcomes. In contrast, applications related to supply chain resilience, sustainability, and cybersecurity remain relatively limited and conceptual, with only a small number of case studies. This indicates that while the current literature suggests that GenAI has considerable potential to improve decision support and enhance supply chain performance, many proposed applications have yet to be validated through large-scale industrial implementations.

Considering the limitations of this research, the findings are subject to the scope of the database, search terms, screening rules, and coding procedures used in the review. Broad search terms were applied to improve retrieval coverage, while relevance was assessed through title, abstract, keyword, and full-text screening. Nevertheless, classification uncertainty may remain, particularly for studies spanning multiple GenAI models or SCM functions. The Supervised Embedding Visualization is exploratory and should not be interpreted as independently generated topic clusters. These limitations restrict the generalizability of the reported patterns and reinforce the need for future reviews using expanded databases and additional empirical evidence. Our study has identified several avenues for future research:

• Analysis of Figure 9 reveals that few studies have been conducted for Service and Returns (Level 5). The ability of Transformer-based models to analyze unstructured data for sentiment analysis is a potential area for research, enabling better understanding of clients’ perspectives. This can be used to predict product returns and improve customer service. Additionally, Transformer-based models can be studied to understand and decode warranty claims for a better understanding of traceability to enable trust among consumers and financiers. Other GenAI models can be integrated with GANs to generate high-fidelity synthetic defect datasets to automate inspection. VAEs can also be used to generate repair geometries that enable precision additive manufacturing of worn components.
• According to Figure 9, GenAI models can be used for SCM level 2 tasks to generate realistic demand and supply scenarios, simulate supplier behavior under uncertainty, and create synthetic transaction and operational data to improve decision-making in environments characterized by data sparsity or volatility. These models could be used to evaluate alternative sourcing strategies, support supplier selection and negotiation, and assess inventory policies. Similarly, in warehousing and inventory management, they have high potential to be used to generate realistic order streams, layout configurations, and replenishment scenarios to optimize storage, picking, and stock allocation decisions.
• Despite their low adoption in supply chain research, flow-based generative models have the potential for tasks related to probabilistic forecasting and inventory optimization. Future work could focus on adapting flow-based models to better capture the dynamic and interdependent nature of supply chains, particularly in environments with high uncertainty and variability. These models can be used for scenario generation to evaluate a wide range of potential outcomes, as well as for dynamic demand forecasting and inventory prediction considering seasonality, trends, supplier reliability, lead times, and external factors for higher accuracy.
• Generative AI models with complex architectures may pose challenges in interpretability. In SCM, the black box nature of AI models can create trust issues. This may result in hesitation to rely on the generated predictions and recommendations. Future research could investigate approaches to improve model explainability in GenAI, such as incorporating interpretable layers or designing visual tools that allow users to observe influencing factors on outputs. Future research could focus on developing ethical guidelines and privacy-preserving techniques for generative AI in SCM, such as differential privacy, federated learning, or blockchain integration.
• As organizations increasingly integrate GenAI into various SCM tasks, such as procurement, logistics, and forecasting with large volumes of sensitive operational, financial, and customer data, it creates concerns regarding data privacy, information leakage, and regulatory compliance. At the same time, several recent studies demonstrate that Transformer-based models can strengthen cybersecurity by detecting cyber threats, identifying anomalous network activities, and analyzing unstructured cyber threat intelligence using contextual language representations [248,249,250]. Therefore, while GenAI offers opportunities to improve both operational decision-making and cybersecurity, future research should place greater emphasis on developing secure, privacy-preserving, and trustworthy GenAI frameworks that protect sensitive supply chain information while maintaining the benefits of intelligent decision support.
• One direction of research could focus on the uses of GenAI for enhancing decision-making in ethical supply chains. This includes areas such as supplier compliance monitoring, fair labor practices, and environmental sustainability. GenAI can generate synthetic datasets for simulating scenarios or predicting outcomes to address ethical challenges and advance sustainable practices across global supply chain networks. For instance, GANs can simulate potential violations of labor laws or environmental standards, allowing organizations to preemptively mitigate risks, and flow-based models could provide probabilistic insights into potential non-compliance events based on historical trends. VAEs can be utilized to compress and analyze massive datasets on labor practices to uncover hidden patterns of exploitation. In sustainability, for example, GenAI models could simulate the effects of adopting renewable energy sources or switching to eco-friendly packaging, providing actionable recommendations for improving sustainability metrics.
• The existing literature largely concentrates on the use of GenAI for individual SCM tasks, such as inventory management, demand forecasting, and logistics optimization, often treating these as separate functions rather than interconnected components within a continuous process. Future research could enhance this by developing models that enable seamless connections across each phase. This can create a circular flow of information that optimizes efficiency and adaptability throughout the entire SCM process. This could also be potentially supported by agentic AI systems capable of autonomously coordinating and executing decisions across multiple supply chain functions.
• A promising area for research is the application of GenAI in Product Lifecycle Management. Particularly, the research could examine how GenAI models can optimize design processes, predict product failure rates, and enhance decision-making around product disposal, refurbishment, or recycling. Such applications could significantly reduce waste and extend the usability and durability of products in a circular economy. GenAI models can analyze past product designs and customer feedback to generate optimized prototypes that meet functional and aesthetic criteria. Similarly, VAEs or flow-based models can analyze historical performance data and generate synthetic scenarios that predict product failure rates under different usage conditions. These models should also be able to enhance end-of-life decision-making related to recycling, refurbishment, and remanufacturing, as well as optimize lifecycle costs.
• Generative AI models are often trained on static datasets, limiting their effectiveness in dynamic supply chain environments. Future research should focus on developing models that can process and learn from real-time data streams to use them to adapt to rapidly changing conditions in different areas, including transportation, warehousing, and customer service. This can be supplemented by real-time traffic data and sensor data from vehicles to generate optimized transportation routes that minimize delays and costs.
• The search strategy in this paper primarily focused on Generative AI models and SCM tasks; hence, it did not explicitly include enabling technologies, such as cloud computing, the Internet of Things, cybersecurity, and edge computing as primary search terms. Future reviews could broaden the search strategy to investigate the integration of GenAI with the digital infrastructure that supports modern supply chains to evaluate the effects of these infrastructures on the outcome of GenAI and its impacts. With the increasing growth of GenAI integration with cloud platforms and IoT devices, protecting sensitive data, preventing unauthorized access, and ensuring secure information sharing become essential challenges. Future research should, therefore, investigate the privacy rules of GenAI and robust cybersecurity frameworks for supply chain applications.