You’re listening to “Semantic-Driven Internet of Behaviours for Enhancing Supply Chain ESG Capabilities Through Generative AI,” by Y.P. Tsang and colleagues. Published in 2025.

Abstract.

Pursuing sustainable development goals requires enterprises to enhance their environmental, social, and governance (ESG) capabilities. In logistics and supply chain management, where small and medium enterprises dominate, integrating ESG practices is challenging and often favors larger companies with established frameworks. This study introduces an ESG recommendation system based on generative artificial intelligence (GERS) to provide accessible, tailored ESG guidance. Leveraging large language models and an ESG knowledge base, GERS offers actionable recommendations, particularly benefiting small and medium enterprises. Evaluated through a case study with a Hong Kong Logistics Association ESG assessment programme, expert panels confirmed the quality of its recommendations. Results demonstrate the GERS’s ability to generate ESG improvement plans, enhancing capabilities efficiently.

This research highlights the transformative potential of generative artificial intelligence in fostering sustainability, showcasing its role in creating adaptive, context-aware services that drive collaborative learning and sustainable practices in supply chains.

INTRODUCTION.

In recent decades, the global focus on sustainable development has intensified. Numerous countries and regions have adopted policies aimed at achieving societal sustainability, including efforts to achieve carbon neutrality and foster free trade agreements. The sustainable development goals (SDGs), defined by the United Nations, are particularly prominent. In business, concepts such as corporate sustainability, corporate social responsibility, and health, safety, and environment have converged and ushered in an era dominated by environmental, social, and governance (ESG) considerations. This ESG paradigm prioritises corporate factors to enhance corporate sustainability. Moreover, research indicates that improvements in ESG performance can bolster supply chain resilience.

Within contemporary supply chain management, emerging business models are designed to meet customer demands while also addressing environmental and social objectives. A broad spectrum of ESG assessment frameworks has emerged to evaluate corporate practices in supplier relationship management, business ethics, and waste management. This evolution has culminated in a dynamic community dedicated to continuous ESG improvement, increasingly showcasing its expertise through sustainability reports and digital disclosures. In parallel with these developments, the digital era has ushered in the expansive Internet of Behaviors (IoB) concept. Initially used to analyse individual online activities, the IoB framework is now being extended to capture and decode corporate ESG behaviours from various digital sources, including social media, public disclosures, and electronic sustainability reports.

By embedding IoB into ESG analysis, stakeholders can access a richer, data-driven understanding of how companies interact with their environment and society. Recent advances in generative artificial intelligence (GenAI), as exemplified by large language models (LLMs) such as Llama and Deepseek, have further facilitated this approach by aggregating and interpreting vast amounts of unstructured online data. Enhanced further by retrieval-augmented generation (RAG) techniques, these models can transform diverse ESG-related digital traces into actionable insights. Given the continuously expanding volume of unstructured ESG data, LLMs enable real-time analysis and immediate actionable insights that are unattainable through conventional manual methods.

This capability reduces operational burdens and financial barriers, especially for small and medium-sized enterprises (SMEs), and ensures that recommendations remain current in a fast-paced, data-rich environment. By leveraging these advanced natural language processing tools, organizations can make informed decisions swiftly while maintaining high standards of accuracy and relevance. In this way, integrating the IoB with AI-driven analytics represents a promising pathway to bridge knowledge gaps and drive sustainable practices across the supply chain ecosystem.

This study proposes a GenAI-based ESG recommendation system (GERS), which enhances various pre-trained LLMs through prompt engineering. Additionally, a knowledge base incorporating existing sustainability reports and ESG solution frameworks is integrated to provide domain-specific recommendations, as illustrated in Figure 1. The system’s effectiveness and performance are assessed through a case study, in which GERS is applied to the ESG Supply Chain Management (ESGSCM) assessment program of the Hong Kong Logistics Association (HKLA) to provide recommendations to companies. An expert panel is established to evaluate the generated recommendations, with various LLMs and temperature settings examined to determine a suitable system design.

With the aid of the proposed system, ESG recommendations can be effectively generated to suggest actionable improvement plans, supporting SMEs in developing their ESG capabilities and contributing to achieving the SDGs in the contemporary business environment.

Note. SME = small and medium enterprises; GenAI = generative artificial intelligence; ESG = environmental, social, and governance.

This study makes several novel contributions to the field of AI-mediated corporate sustainability. First, it extends the IoB paradigm to systematically capture and analyse digital footprints of ESG activities, thereby bridging the data gap between traditional ESG assessments and real-time corporate practices. Second, by leveraging state-of-the-art LLMs, this research introduces a solution that offers tailored sustainability strategies, especially valuable for SMEs with limited resources. Third, the study advances methodological innovation through a comprehensive evaluation framework that combines pairwise comparison consistency thresholds with expert validation, ensuring the reliability and effectiveness of the generated insights.

Collectively, these contributions enhance our theoretical understanding of the interactions between digital behavior and sustainability outcomes, offering practical implications for businesses aiming to achieve the SDGs in an increasingly digitalized economy. For instance, leveraging similar interest behaviors within semantic web applications has been shown to optimise collaboration efficiency in digital networks. Similarly, the integration of semantic web technologies with reinforcement learning models has also proven effective in addressing sustainability challenges.

The remainder of this paper begins with a review of the literature on ESG development and AI-driven IoB, identifying key research gaps. It then presents the design of the proposed GERS, which embeds the RAG process within LLMs to enhance knowledge retrieval. A case study follows, using an ESG evaluation framework to assess the system’s effectiveness. The discussion interprets the findings and considers their implications. The paper concludes with remarks on limitations and directions for future research.

LITERATURE REVIEW.

This section reviews the recent literature on ESG in supply chains, focusing on ESG standards and frameworks, and then explores IoB driven by LLMs. A research gap is identified to motivate the development of GERS.

Contemporary Development of Environmental, Social, and Governance Initiatives in Supply Chains

ESG initiatives, originally proposed in 2005 by the United Nations Environment Programme Finance Initiative (2005), have become pivotal in shaping responsible investment strategies and business practices globally. The concept links sustainability with economic performance, encouraging businesses to adopt broader societal concerns into their strategic decision-making. ESG reflects a shift from voluntary, ethics-based initiatives to more structured, finance-focused criteria that align ethical practices with business strategy and financial performance rooted in corporate social responsibility, corporate sustainability, and health, safety, and environment considerations. It plays a role in promoting sustainable business practices that align with the broader objectives of the SDGs.

The proliferation of ESG has led to the development of various frameworks and standards that provide guidelines and metrics for evaluation. For example, these include Morgan Stanley Capital International Inc. (MSCI) ESG Ratings, S&P Global Corporate Sustainability Assessment (CSA), Morningstar Sustainalytics, and HKLA’s ESGSCM. Table 1 summarises the comparisons between these assessment frameworks regarding the focus and Strengths.

Note: 1 The ESG ratings measured and provided by Morgan Stanley Capital International Inc. (MSCI) 2 S&P Global Corporate Sustainability Assessment (CSA) 3 Hong Kong Logistics Association’s ESG Supply Chain Management program (HKLA’s ESGSCM)

Incorporating ESG into supply chains has become a key objective for businesses seeking to enhance sustainability across their operations. Companies are increasingly held accountable not only for their direct operations but also for their supply chains, which often span multiple countries and involve various tiers of suppliers. The challenge lies in the complexity of monitoring and ensuring compliance across diverse geographical and regulatory landscapes. Recent studies highlight the importance of adopting supplier codes of conduct, integrating ESG criteria into supplier selection and evaluation processes, and utilising digital tools for tracking and reporting ESG performance.

Organizations actively facilitate the sharing of best practices, enhance sustainability training, and improve transparency. Key practices include the development of centralized sustainability knowledge hubs, which provide resources and best practices across an organization, and utilizing sustainability reports as a tool for both internal and external communication. These reports document progress and challenges, engaging stakeholders in dialogue about sustainability efforts and performance. The rise of digital platforms and collaboration tools has further enabled the sharing of ESG knowledge and practices across organizational boundaries. This includes partnerships between businesses and non-governmental organizations, industry collaborations, and multi-stakeholder initiatives that aim to standardize ESG reporting and improve the reliability and comparability of ESG data.

Large Language Models for Environmental, Social, and Governance Development

LLMs, such as OpenAI’s GPT series, have revolutionised various domains of artificial intelligence by demonstrating remarkable capabilities in processing and generating human-like text. These models, built on advanced transformer architectures, excel at understanding complex natural language patterns and are increasingly applied across various domains. In the context of the IoB, LLMs are emerging as pivotal tools to enhance information retrieval, content summarization, and knowledge discovery within organizations. A groundbreaking innovation in this domain is the integration of RAG techniques, which combine the generative capabilities of LLMs with real-time information retrieval methods to provide contextually enriched and dynamically updated responses. This fusion transforms static LLM outputs into interactive, continuously evolving knowledge bases.

De Nicola et al. (2024) leveraged the potential of LLMs to construct a business process knowledge base, which serves as a foundation for automating business process analysis. The knowledge elicitation process can be especially facilitated to extract meaningful information from a vast amount of textual data. Filippo et al. (2024) investigated the fact that LLMs can perform tasks in various domains, including human resources, programming, social media, office automation, search engines, and education. Based on the above technological foundation, the intersection of ESG and AI has gained considerable traction in recent research. Early contributions, such as Mishra (2023), have explored fine-tuning LLMs with strategic prompt engineering to classify ESG-related news articles effectively, thereby enabling proactive monitoring of emerging trends.

Chen et al. (2023) noted that LLMs may assume a more creative role in the realm of co-word analysis. Similarly, Zou et al. (2025) demonstrated the application of RAG-enabled LLMs in dissecting contemporary ESG reports and frameworks across diverse sectors (e.g., consumer staples and healthcare), illustrating the capability of AI to extract nuanced, sector-specific insights from voluminous unstructured data. Building on these foundational studies, more recent efforts can be expanded into predictive and prescriptive applications of AI in ESG contexts.

Position of This Research

Despite these advancements, a notable gap remains in harnessing LLMs for generating strategic, actionable recommendations, particularly within supply chain contexts. Current research emphasises performance and risk prediction, leaving ESG-related knowledge underutilised for guiding ongoing operational improvements. By leveraging state-of-the-art LLMs in combination with a comprehensive and dynamically updated ESG knowledge base, this work is positioned to produce tailored, real-time guidance for companies. This approach advances the application of IoB in corporate monitoring, providing a much-needed tool for empowering SMEs to improve their ESG practices continually.

AN ENVIRONMENTAL, SOCIAL, AND GOVERNANCE RECOMMENDATION SYSTEM BASED ON GENERATIVE ARTIFICIAL INTELLIGENCE

To facilitate the social interactions among enterprises for ESG development, GERS leverages the potential of LLMs and RAG to offer domain-specific recommendations. The proposed system consists of a knowledge acquisition module, a prompt engineering module, a RAG-based LLM module, and a group-based multi-criteria response evaluation module, as illustrated in Figure 2.

Knowledge Acquisition Module

A critical step in developing a domain-specific LLM solution is constructing a knowledge base with a corpus of relevant documents for retrieval and generation. More importantly, the ground of external knowledge can effectively reduce the number of hallucinations in LLM applications, making generated responses more accurate and relevant. Three types of relevant knowledge are suggested to generate ESG-related recommendations for supply chain firms, as shown in Table 2. In addition to the internal knowledge retained in LLMs, external knowledge about ESG assessments, such as EcoVadis, is necessary. Those assessments’ detailed criteria for assessing ESG perspectives are defined and tailored to supply chain management. Subsequently, the key ESG concepts in the supply chain context can be identified rather than considering the entire ESG spectrum.

Second, another important process is sharing sustainability reports, as practised by most large enterprises, which contain actionable proposals for improving corporate sustainability. Large enterprises tend to publicise their ESG efforts in the contemporary business environment to improve perceptions of their capabilities and responsibility. Third, frameworks for preparing and implementing ESG solutions are considered to offer a clear and well-defined solution to address ESG-related challenges, in particular, using the challenge–evaluation–planning–action–review (CEPAR) model proposed by the International Chamber of Sustainable Development.

Note: 1 Hong Kong Logistics Association’s ESG Supply Chain Management program (HKLA’s ESGSCM) 2 Challenge-Evaluation-Planning-Action-Review (CEPAR) model

All the above knowledge, namely a corpus of documents D= {d1,d2...,dn}, is retained in the knowledge base, in which embeddings are generated for the documents to obtain ED= ​[​ ​e​ ​d​ 1​​, ​e​ ​d​ 2​​..., ​e​ ​d​ n​​]​. When a query is received, the cosine similarity between the embedded query and documents can be measured by using the following equation to extract the top-k1 most relevant documents to prioritise the recall:

​sim​(​ ​e​ q​, ​e​ ​d​ i​​)​ = ​ ​e​ q​ ∙ ​e​ ​d​ i​​ ​‖​e​ q​‖​‖​e​ ​d​ i​​‖​ ​

Subsequently, the top-k1 documents are re-ranked to improve the precision, with a cross-encoder model applied to measure relevance scores between the query and top-k1 documents. Next, the top-k2 documents, where k2≤k1, are obtained for the response generation.

PROMPT ENGINEERING MODULE

In addition to providing queries to determine a set of relevant documents, a suitable prompt strategy should be employed to enhance the relevance and precision of generated responses. Based on the fundamentals of prompting, several prompt strategies, such as zero-shot and role-play prompting, have been developed. The proposed system leverages a tailor-made prompt design that carefully integrates multiple prompt strategies, specifically, role-play prompting, chain-of-thought prompting, and self-consistency, to align with the unique demands of ESG assessment in supply chain management. Table 3 provides further explanation of these prompting methods.

Note. ESG = environmental, social, and governance; CEPAR = challenge–evaluation–planning–action–review.

By integrating the above prompting methods, the proposed prompt in this module is designed to assume the role of an expert ESG consultant, ensuring a focused, domain-specific perspective. It employs role-play prompting to simulate real stakeholder scenarios, chain-of-thought prompting to break complex ESG challenges into clear, sequential steps, and self-consistency prompting to validate and stabilise recommendations via iterative generation. This structured approach follows the CEPAR framework, guiding the model to deliver comprehensive, actionable, and contextually relevant responses for ESG logistics and supply chain management improvements.

A Large Language Model Module Based on Retrieval-Augmented Generation

The data flow of the RAG-based LLM module begins with data loading, where information is retrieved from sources such as text files, PDFs, and websites and integrated into the pipeline using data connectors, including those from the Llama Hub. Next, during the indexing stage, the data are represented numerically as vector embeddings, paired with metadata to facilitate accurate retrieval. These embeddings and metadata are then stored in a vector database, such as Qdrant, to avoid repeated indexing. When a query is received, the system initiates a two-stage retrieval process. An initial embedding-based retrieval fetches the top-​k​ 1​documents, prioritising recall, followed by a re-ranking stage to extract the top-​k​ 2​ ​documents, emphasising precision.

The module then uses these re-ranked documents to augment the query context, allowing the pre-trained LLMs to synthesise an accurate and relevant response. With the relevant documents and prompts in place, the information is fed into the pre-trained LLM, such as Llama 3 (8B or 30B). The pre-trained LLM, which has been trained on large amounts of text, leverages its deep learning capabilities to process the input. Consequently, the RAG process synthesises the information from the retrieved documents and the guidance provided by the prompts for response generation, resulting in enhanced quality and reliability.

To optimise the quality of the generated responses, two primary parameters, namely the temperature and the number of top-ranked items after re-ranking ​k​ 2​, can be investigated. The temperature, between 0 and 1, controls the randomness of the predictions. A relatively low temperature value should be considered when performing factual tasks to prohibit more diverse and creative responses. The value ​k​ 2​ controls the quantity of top-ranked information to be considered after re-ranking. When adjusting the value of ​k​ 2​, a balance between relevance and diversity in the response generation process should be made. Karpukhin et al. (2020) suggested that a small value between 5 and 20 is practical to maintain a reasonable recall while ensuring relevance. Thus, the generated responses about ESG recommendations can be used for comparison and evaluation.

For instance, when an ESG analyst submits the query, “What strategies can a logistics company adopt to reduce their carbon footprint in urban transportation?” the RAG-based module initiates the retrieval process. In the embedding-based retrieval phase, top-​k​ 1​ documents such as recent sustainability reports, research papers on urban logistics, and policy guidelines on carbon reduction are identified. During the re-ranking stage, the system prioritizes documents that specifically detail actionable strategies, such as optimizing delivery routes, adopting electric vehicles, and leveraging predictive analytics for energy management. These selected documents are then used to augment the query context, and an LLM model (e.g., Llama 3) processes this enriched input to generate a tailored recommendation.

Group-Based Multi-Criteria Response Evaluation Module

Once the recommendations are generated, a group of qualified, certified ESG planners is invited to evaluate the accuracy, relevance, feasibility, and comprehensiveness of the recommendations. This helps ensure their factual correctness, pertinence, and practicality, which can fully address the weaknesses identified. Since the values of parameters, including ​k​ 2​ and temperature could affect content generation. The Bayesian best–worst method (Bayesian BWM) is applied to compare various sets of recommendations and obtain the most reasonable one. Its technical mechanism is described with pseudo-code in Figure 3. Using the Bayesian BWM, expert judgments are systematically aggregated while accounting for uncertainty. By requiring assessors to make pairwise comparisons between the best and worst criteria, Bayesian BWM converts qualitative insights into quantitative weights.

This is crucial when dealing with ESG evaluations, where outputs can be sensitive to variations in LLM parameters such as temperature. The probabilistic framework, employing multinomial and Dirichlet distributions, enables the seamless integration of prior knowledge with observed data, ensuring that the inherent variability among expert opinions is effectively managed. In doing so, posterior weights are generated to reflect a balanced consensus among certified ESG experts. This approach enhances transparency and interpretability in the decision-making process, ensuring that the final composite ESG recommendations are reliable and contextually relevant.

In order to ensure the quality of the pairwise comparison data, input-based consistency ratio thresholds, as proposed by Liang et al. (2020), are employed. After assessors complete both best-to-others and others-to-worst pairwise comparisons, the resulting data are evaluated by computing the input-based consistency ratio (CRI), as defined in Equation 2. This ratio must meet the predefined threshold; if it does not, assessors are required to revise their pairwise comparisons to enhance internal consistency.

​|​a​ Bj​ ∙ ​a​ jW​ − ​a​ BW​|​ ​CR​ I​ = ​max​ j​ ​ ​a​ BW​ 2​ − ​a​ BW​ ​, where ​ a​ BW​ > 1​

Algorithm 1: Bayesian Best–Worst Method With the Kendall’s ​W​ Evaluation

Inputs: Set of decision criteria C= {c1,c2...,cn} and Set of decision-makers D= {DM1, DM2...,DMm} Outputs: Aggregated weights of criteria W= {w1,w2...,wn}

Procedure: For each DMi in D: Select the best criterion cB and the worst criterion cW from C Perform pairwise comparisons to form the Best-to-Others vector AB and Others-to-Worst vector AW ​A​ B​ = ​{​ ​a​ B1​, ​a​ B2​..., ​a​ Bn​}​ // Comparison of cB with cj ​A​ W​ = ​{​ ​a​ 1W​, ​a​ 2W​..., ​a​ nW​}​ // Comparison of cj with cW End for Model pairwise comparisons using probability distributions Use the multinomial distribution for pairwise comparisons Use the Dirichlet distribution for the final aggregated weights Develop a Bayesian hierarchical model to compute the weights Define prior distributions for the weights based on initial beliefs Use observed data (pairwise comparisons) to update the prior distributions Compute the posterior distributions of the weights Compute the optimal weights W Use the posterior distributions to determine the aggregated final weights W For each decision criterion cj Convert the pairwise comparison data from each DMi into rank orders Compute the cumulative rank Ri= Σ(rank assigned by each DM) for each cj.

Compute Kendall’s W using Equation Perform hypothesis testing based on chi-square statistic If agreement (​W​) is statistically significant, Validate W as the final aggregated weights Otherwise, re-assess the pairwise comparisons Output the verified weights W

To enhance the reliability of our multi-criteria evaluation, Kendall’s coefficient of concordance (Kendall’s ​W​) is integrated into the Bayesian BWM framework. Kendall’s ​W​ quantitatively measures the agreement among assessors by converting their pairwise comparisons into ranked data. This statistical measure verifies that the aggregated weights obtained from Bayesian BWM genuinely reflect a consistent expert consensus. By validating inter-assessor consistency through hypothesis testing (using the chi-square statistic), Kendall’s ​W​ adds a critical layer of scrutiny. Only when a statistically significant agreement among assessors is confirmed and the Bayesian-derived weights deemed reliable. This integration ensures that our final recommendations are both numerically rigorous and qualitatively consistent across expert evaluations.

By converting the pairwise comparison data (i.e., best-to-others and others-to-worst) into ranks, Kendall’s ​W​ statistic is expressed as n ​ ​(​ ​R​ i​ − ​ R ​)​ 2​ 12 ∙ ​∑ i=1​

​W = ​

​ ​m​ 2​(​n​ 3​ − n)​ − m ∙ T where ​R​ i​ = ​∑ j=1​ m ​ ​r​ ij​ and rij is the rank assigned by assessor j to subject i. In addition, ​m​ and ​n​ n ​ ​(​ ​t​ k​ 3​ − t​)​ denote the numbers of assessors and subjects for evaluation, respectively, and ​T = ​∑ k=1​ is the correction factor for tied ranks, in which tk is the maximum number of tied ranks for all the evaluation subjects ​n​. Furthermore, Kendall’s ​W​ can be applied to hypothesis testing based on the chi-square statistic, namely ​χ​ 2​ = m​(​n − 1​)​W​, with ​n − 1​ degrees of freedom. The null hypothesis (H0) is that there is no agreement among the assessors. The alternative hypothesis (Ha) is that there is some agreement among assessors.

Such a method provides a systematic framework suitable for aggregating the opinions of a few key domain experts with specialised knowledge. A hierarchy tree is shown in Figure 4 to illustrate the recommendation generation process. In Level 1, pairwise comparisons are conducted between the four defined criteria to determine the aggregated final weights, namely ​W​ L1​ = ​{​ ​w​ 1​ L1​, ​w​ 2​ L1​..., ​w​ 4​ L1​}​. In Level 2, pairwise comparisons are conducted for a set of generated ESG solutions by using different L2​|​j = 1..., e​}​ for the criterion ​ parameters in LLM settings under the four criteria, namely ​W​ i​ L2​ = ​{​ ​w​ i,j​ i​ and solution j. Therefore, the composite weight of the ESG solution j is expressed as

​ω​ j​ = ​∑ i​ ​w​ i,j​ L2​ ∙ ​w​ i​ L1​

Note. ESG = environmental, social, and governance.

CASE STUDY

To verify the effectiveness of the proposed system for ESG recommendations, a case study is conducted based on the ESGSCM programme of the HKLA. A company in the assessment pool, referred to as ABC Ltd., is selected for in-depth investigation, and a set of recommendations is generated to resolve the most critical ESG problem.

Case Background

HKLA launched the ESGSCM programme in April 2024, accompanied by the development of a comprehensive rating questionnaire designed to evaluate companies’ ESG performance, with a primary focus on supply chain management. The assessment encompasses 119 measurement items, categorised into five domains: (i) business operations, (ii) suppliers and customers, (iii) logistics, (iv) organizational and managerial structures, and (v) waste management in logistics operations. Each item is quantified on a scale from 0 to 10. To appraise the ESG performance, an assessment pool is established for all participants, and the median performance across the five domains is calculated. By employing modified z-scores, both overall and sectional performance can be evaluated to ascertain whether a company’s ESG performance is above or below the industry median.

Additionally, Pearson’s correlation coefficient is used to identify the most influential measurement items concerning the overall rating, such as the top 10 items. By comparing these influential items with the company’s input data, areas of weakness warrant improvement can be identified. These identified areas can serve as inputs to the proposed GERS, facilitating the generation of tailored ESG recommendations. ESG recommendations are prepared in this case study following the CEPAR framework. The target corporation’s business context is outlined in the

Challenge phase, and one or two core ESG challenges are identified as the focal issue to address. The Evaluation step examines whether the challenge is financially material by considering double materiality, which involves assessing both the corporation’s impact on society and the environment, as well as its financial implications. During the Planning phase, policy directions and strategic actions are designed to mitigate ESG risks while enhancing competitive positioning, focusing on stakeholder preferences and relevant performance indicators. The Action stage details the implementation of the proposed solutions, including resource allocation, incentive structures, and stakeholder communication strategies to ensure effective execution.

Finally, the Review phase establishes methods to assess outcomes using both financial and non-financial key performance indicators (KPIs), providing a basis for monitoring progress and adapting strategies if expectations are not met.

In addition to the assessment criteria within the knowledge base, sustainability reports (up to 20 February 2025) published by ten industry leaders, namely SF Holding, Alibaba, JD Logistics, DHL, FedEx, Kerry Logistics, Lalamove, UPS, Maersk, and Amazon, have been reviewed to document best ESG practices. The summaries of their ESG practices are organised and documented as an essential component of the knowledge base. Furthermore, when addressing ESG challenges, the CEPAR model, which is a five-step methodology encompassing Challenge, Evaluation, Planning, Action, and Review, is employed. This model focuses on evaluating the financial materiality of ESG challenges according to the SASB standards. An action plan can be proposed by constructing a stakeholder impact map, aligning with specific SDGs.

With the aid of the above knowledge, the challenges faced by ABC Ltd can be analysed by leveraging the power of an LLM with the RAG mechanism.

Settings and Protocol

For the deployment of the proposed system, the configurations of the RAG and LLM are specified as shown in Table 4. LlamaIndex is chosen for RAG in LLMs due to its scalability and adaptability to various data sources, allowing large and diverse datasets to be effectively handled. In particular, its highly efficient and optimised indexing algorithms are used to achieve fast and accurate document retrieval. One of the FlagEmbedding models, bge-small-en-v1.5, created by the Beijing Academy of Artificial Intelligence (BAAI), is selected to generate embeddings of retained documents and user queries. Its semantic richness and context awareness result in embeddings that describe the context more effectively. The open-source Meta Llama 3-8B LLM is adopted in the proposed system to generate ESG recommendations that industrial practitioners can easily understand.

To evaluate the quality and relevance of the responses, two values of the parameters of k2 and temperature are considered, yielding four sets of solutions per ESG challenge. Then, pairwise comparisons under the Bayesian BWM can be conducted to determine the most appropriate ESG solution based on the responses of the specialised ESG experts.

Based on the settings defined above, the proposed system is applied in a real-life situation, as illustrated in Figure 5, which shows the flow of the entire system implementation. By analysing the enterprise performance in terms of its ESG aspects, a finalised ESG solution based on the CEPAR model is generated for continuous improvements.

Deployment

During the trial run of the HKLA’s ESGSCM programme, 10 companies were invited to complete the questionnaire. A medium-sized third-party logistics company, referred to here as ‘Company A’, is selected from the pool to demonstrate the efficacy of GERS. In addition to assessing ESG performance in five individual sections, as mentioned in Section 4.1, the top 10 most influential items are identified by evaluating the Pearson’s correlation coefficient, such that the top 10 items statistically have the strongest proportional effect on the overall ratings. This can help prioritise the enterprise’s resources to develop critical areas, improving ESG performance. The improvement areas for Company A are listed in Table 5.

Based on the selected prompt methods discussed in Section 3.2, a customised prompt and query, as shown in Appendix A, are applied in the settings of the LLM, which incorporates the CEPAR model to investigate specific ESG challenges. The knowledge base also includes published sustainability reports from 2022 and 2023 from well-known companies, such as Alibaba, Amazon, DHL, FedEx, JD Logistics, Kerry Logistics, Lalamove, Maersk, SF Express, and UPS.

In 2022, these companies drove decarbonization through fleet electrification, renewable energy integration, and carbon offset initiatives while managing risks and ensuring transparent governance. They also invest substantially in employee well-being, diversity, and community resilience, and leverage digital transformation to enhance operational efficiency. Furthermore, these leaders advocate responsible supply chain practices, cutting-edge technology, and collaborative partnerships, setting industry benchmarks for creating long-term economic, environmental, and social value in a globally interconnected marketplace.

In comparison to 2022, more ambitious decarbonization targets are demonstrated in 2023, with more precise roadmaps for achieving carbon neutrality through expanded electric fleets, improved renewable energy sourcing, and advanced digital carbon accounting techniques, such as refined Scope 3 measurement frameworks. Governance structures have been further strengthened, with enhanced risk management and increased transparency in stakeholder engagement. The reports highlight improvements in employee well-being and diversity initiatives, including more comprehensive training programs, flexible work arrangements, and upgraded benefits. Additionally, community inclusion efforts are now supported by digital platforms that connect rural areas to global markets and support local value chains.

To generate recommendations for the above three challenges, a group of five assessors are invited to evaluate a set of alternatives and determine the most appropriate solution. For the Level 1 comparisons, assessors are required to evaluate the four primary criteria using the best–worst pairwise comparison process. Based on the defined parameter settings, four alternatives per challenge are formulated for the Level 2 pairwise comparisons, namely C1 (temperature = 0 and k2 = 3), C2 (temperature = 0 and k2 = 7), C3 (temperature = 0.5 and k2 = 3), and C4 (temperature = 0.5 and k2 = 7). All the collected pairwise comparison data are summarised in Appendix B, where the input-based consistency ratios are also evaluated to ensure that the assessors conduct the comparisons consistently.

By applying the group-based multi-criteria response evaluation module from GERS, it is found that the composite weights for the Level 1 criteria—namely, accuracy, relevance, feasibility, and comprehensiveness—are 0.1989, 0.2445, 0.4480, and 0.1086, respectively. The credal ranking between these four criteria is visualised in Figure 6, which shows the confidence in the superiority of one criterion over another. Feasibility is the most important criterion in preparing the ESG solution, and comprehensiveness is the least important. Therefore, the practicality of recommendations is prioritised in this case study.

The assessors then conduct Level 2 pairwise comparisons to evaluate a set of ESG solutions (C1 to C4) for each criterion and challenge. The weights determined by the Bayesian BWM for problems P1, P2, and P3 are summarised in Tables 6, 7, and 8, respectively. The corresponding rankings of the four ESG solutions are ​C3 ≻ C1 ≻ C2 = C4​, ​C3 ≻ C1 ≻ C4 ≻ C2​, and ​C3 ≻ C1 ≻ C4 ≻ C2​. It is found that methods C1 and C3 can consistently generate the highest-quality responses, with C3 slightly superior to C1. The performance of C2 and C4 is highly similar, with equal performance when solving P1. It is also observed that C1 and C3 generally outperform C2 and C4.

Note. C = challenge.

Note. C = challenge.

Note. C = challenge.

According to the above group-based multi-criteria evaluation process, method C3 should be adopted to generate solutions to improve the ESG capability of Company A. The ESG solutions are summarised in Tables 9, 10, and 11 for P1, P2, and P3, respectively. In Table 9, the proposed solution addresses the challenges of inefficient e-booking and e-ordering systems, which currently result in operational delays, customer dissatisfaction, and limited scalability. Evaluations reveal that poor digital infrastructure increases operational costs and hinders competitiveness compared to advanced 3PL providers, exacerbating sustainability issues by elevating carbon emissions and resource waste. Stakeholders, including customers, employees, investors, regulators, and logistics partners, demand a more transparent, efficient, and secure system.

To address these concerns, the plan advocates for developing an integrated platform that leverages AI-driven optimization, predictive analytics, robust cybersecurity, and comprehensive staff training. The approach is structured into immediate, short-term, and long-term actions, starting with the formation of cross-functional teams, conducting stakeholder interviews, piloting the system, and eventually rolling out a fully integrated solution with real-time tracking and AI-powered recommendations. The success of the solution is monitored through KPIs focused on operational efficiency, customer satisfaction, scalability, and sustainability.

Note. C = challenge; E = evaluation; P = planning; A = action; R = review.

As shown in Table 10, the lack of measurements for waste storage and disposal control presents severe environmental and health risks. Improper waste management can lead to environmental pollution and respiratory diseases, negatively impacting both society and the environment, and resulting in increased costs, reputational damage, and potential fines. The stakeholder impact map identifies the need for a safe working environment for employees, regulatory compliance and cost reductions for management and suppliers, high-quality products and services for customers, and reduced environmental impact for regulators and local communities.

Recommended policies include implementing a data-driven system for waste disposal and tracking, developing regulations for waste management and recycling, encouraging source reduction and recycling, and fostering a culture of sustainability aligned with SDGs 12 (Responsible Consumption and Production), 13 (Climate Action), and 17 (Partnerships for the Goals). To facilitate the implementation of these policies, resources should be allocated to waste management infrastructure, employee training, and partnerships with suppliers and regulators. Incentives such as participation rewards, training, and recognition for environmental responsibility should be provided to motivate employees. The performance metrics encompass both financial indicators, such as cost savings and revenue growth, as well as non-financial indicators, including waste reduction, recycling rates, and employee engagement.

Note. C = challenge; E = evaluation; P = planning; A = action; R = review.

In Table 11, the proposed solution targets operational inefficiencies, escalating costs, and environmental issues in warehouse operations by standardising workflows. With inconsistent processes leading to higher emissions, waste production, and customer dissatisfaction, the initiative responds to growing pressures from customers, regulators, and market forces. It aims to enhance efficiency, reduce costs, and boost sustainability by auditing current practices, developing comprehensive standard operating procedures, integrating technologies such as warehouse management systems and automation tools, and conducting targeted staff training. A pilot phase in select warehouses should precede full-scale implementation, with continuous monitoring and regular audits ensuring adaptive improvements.

The success of this solution can be evaluated through KPIs that capture improvements in operational costs, order fulfilment time, inventory accuracy, and sustainability metrics. This integrated roadmap addresses immediate operational challenges and supports long-term strategic goals in sustainability and customer satisfaction.

Note. C = challenge; E = evaluation; P = planning; A = action; R = review.

RESULTS AND DISCUSSION

The case study demonstrates the successful implementation of our proposed system to address three key challenges identified within the HKLA’s ESGSCM programme. By integrating RAG and prompt engineering techniques, our system effectively consolidates publicly available ESG reports into a comprehensive knowledge base and generates practical solutions. This approach aims to motivate logistics and supply chain SMEs to develop and continuously improve their ESG capabilities. Drawing from the case study results, in this section, we focus our discussion on three critical aspects: the agreement of the assessors, the impact of LLM settings, and the research implications.

Agreement Analysis of Accessors

For the Level 1 pairwise comparison, the agreement of the evaluation criteria is important as it indicates that the assessors have similar ideas regarding the solution evaluation. For best-to-others comparisons, we observed a Kendall’s ​W​ value of 0.676, with a corresponding chi-squared statistic of 10.14 (​p​ = 0.0174). Similarly, others-to-worst comparisons yielded a Kendall’s ​W​ value of 0.556, with a chi-square statistic of 8.34 (​p​ = 0.0395). Given a significance Level of 0.05, we reject the null hypothesis in both cases, indicating statistically significant agreement among assessors. This concordance extends to all primary evaluation criteria: accuracy, relevance, feasibility, and comprehensiveness.

These results suggest a shared understanding among assessors regarding the importance and application of the evaluation criteria, enhancing the validity of our solution assessment process. In other words, the marking rubrics among assessors are mutually agreed upon and similar to those used to judge the performance of the ESG solutions.

Impact of Large Language Model Settings

Through pairwise comparisons, the case study investigates four alternative ESG solutions derived from different LLM settings. It is found that adjusting the temperature value yields marginal improvements in response quality. At a temperature of 0, the LLM produces deterministic outputs, selecting tokens with the highest probability at each step. Introducing a non-zero temperature (0.5 in our study) incorporates a degree of randomness and diversity into the responses. However, the observed improvement is minimal, suggesting that the deterministic outputs at temperature 0 are already of sufficient quality for ESG solution generation. This implies that in this specific context, the added diversity from increased temperature does not substantially enhance the utility of the generated responses.

Conversely, increasing the k2 value, which determines the number of top-ranked items selected after re-ranking retrieved results using a cross-encoder for MS Marco, negatively impacted response quality. Expanding k2 from 3 to 7 was expected to enhance diversity by retaining more potentially relevant items, but it resulted in a significant decline in quality. This counterintuitive outcome can be attributed to the inclusion of less relevant items, which potentially introduces noise that dilutes the overall quality of the information used for response generation. These findings highlight the delicate balance required in parameter tuning for LLM-based ESG solution generation systems. Although some parameter adjustments may theoretically enhance diversity or information richness, their practical impact on response quality can be counterproductive in specific applications.

Research Implications

The innovation of developing ESG capabilities through GenAI is not only a technological breakthrough but also a catalyst for reshaping corporate behavior and decision-making. The proposed GERS offers a transformative approach to global sustainability by democratising access to ESG best practices and accelerating the integration of sustainable initiatives across corporate supply chains. GERS generates positive externalities by aggregating and analysing collective insights from sustainability reports and diverse ESG frameworks, such as reduced environmental degradation, improved labor conditions, and enhanced corporate governance. These benefits ultimately foster environments that promote sustainable behavior and ethical decision-making among corporate actors.

Importantly, given the sensitive nature of the data being processed, GERS adheres to responsible AI principles by prioritising fairness, accountability, transparency, and privacy-by-design. Compliance with frameworks like the General Data Protection Regulation ensure that sensitive corporate data and stakeholder information remain secure and are not misused. Furthermore, ethical considerations are integrated within its design to maximise societal benefits while safeguarding individual rights. This commitment to ethical AI reinforces user trust in AI-mediated ESG guidance. It establishes an inclusive and secure environment for large corporations and SMEs, driving sustainable change in line with global regulatory standards.

Moreover, introducing GERS can potentially create a paradigm shift in how businesses, especially SMEs, approach ESG integration. By offering cost-effective and AI-driven ESG guidance, the system addresses common resource constraints that often hinder SMEs from establishing robust ESG practices. This democratization of ESG insights encourages more inclusive and sustainable supply chains, contributing to a corporate ecosystem where ESG excellence becomes standard practice rather than an exception. GERS’s adaptive, self-learning nature enables businesses to stay agile in their sustainability efforts, ensuring continual responsiveness to emerging ESG trends in a rapidly evolving global market.

Finally, GERS exemplifies the application of GenAI in analysing and shaping corporate behavior, effectively establishing an IoB. By processing extensive ESG-related data and delivering context-aware, actionable recommendations, the system fosters novel AI-mediated social interactions among business professionals, particularly benefiting SMEs. This approach encourages collaborative learning and knowledge sharing within business communities, directing corporate behavior toward more sustainable practices. The continuous learning capability of GERS highlights the dynamic role of AI in capturing and influencing behavioral patterns in corporate settings, thereby demonstrating how adaptive, human-centred technologies can address complex societal challenges while embodying the core principles of the IoB paradigm in corporate sustainability contexts.

CONCLUSION.

In summary, this research presents GERS, a novel application of generative AI designed to tackle critical sustainability challenges in supply chain management. By integrating RAG with LLMs for ESG analysis, our methodological framework formalises the process of capturing digital ESG footprints and converting them into actionable insights. This approach bridges the gap between the extensive ESG expertise of large corporations and the resource constraints faced by SMEs, thereby democratising access to sustainable business practices and supporting technology-mediated decision-making. The research findings demonstrate that GERS efficiently processes and synthesises vast amounts of ESG-related data, delivering tailored recommendations vital for organizations, especially SMEs with limited resources for comprehensive ESG strategy development.

Furthermore, the system’s integration into existing digital platforms provides a versatile decision-support tool for sustainability consultants, corporate strategy teams, and regulatory bodies, enabling them to monitor and respond to evolving ESG standards in real-time.

Ultimately, this research offers a practical solution for advancing ESG development while serving as a proof of concept for incorporating generative AI into complex societal and business challenges. As organizations progress toward a more sustainable and responsible global economy, embedding GERS within existing ESG monitoring frameworks can enhance decision-making processes and drive transformative change across industries. Nonetheless, it is important to acknowledge the limitations and the need for future research, including long-term investigations to assess the impact of GERS on corporate performance and global sustainability metrics.

Building on these findings, future research directions can further enhance the capabilities and impact of GERS, addressing its limitations and exploring new possibilities for advancing sustainable practices and decision-making. First, longitudinal studies are necessary to assess the long-term effects of GERS on corporate performance and its contribution to global sustainability metrics. Second, research could focus on improving the system’s scalability and adaptability across diverse industries and geographic regions, ensuring its relevance to varying ESG standards and regulatory requirements. Additionally, integrating advanced machine learning techniques, such as explainable AI, could provide deeper transparency and trust in the decision-making process, especially for SMEs and regulatory bodies.

Ultimately, interdisciplinary research on the ethical implications and social acceptance of generative AI in ESG management may provide critical insights to refine its deployment and ensure alignment with broader sustainability objectives. These efforts would not only expand GERS’s applicability but also strengthen its role in advancing sustainable practices globally.

FUNDING SOURCES

This research was supported by BDIC (HSUHK) and ISE (PolyU). It is also supported in part by a grant from the Research Grants Council of the Hong Kong Special Administration Region, China (UGC/FDS14/E08/21), a grant from Public Policy Research Funding Scheme of HKSAR Government (Project no.: 2023.A6.225.23B; Project Code: K-QZ3E) and an internal project supported by the Faculty Development Grant (Project Acc. Code: 800005) of the School of Decision Sciences, HSUHK.

CONFLICTS OF INTEREST

We confirm that there are no known conflicts of interest associated with this publication and there has been no significant financial support for this work that could have influenced its outcome.

PROCESS DATES

07. 2025.