SustAI-SCM: Intelligent Supply Chain Process Automation with Agentic AI for Sustainability and Cost Efficiency

Abstract

Sustainable supply chain management (SCM) demands efficiency while minimizing environmental impact, yet conventional automation lacks adaptability. This paper presents SustAI-SCM, an AI-powered framework integrating agentic intelligence to automate supply chain tasks with sustainability in focus. Unlike static rule-based systems, it leverages a transformer model that continuously learns from operations, refining procurement, logistics, and inventory decisions. A diverse dataset comprising procurement records, logistics data, and carbon footprint metrics trains the model, enabling dynamic adjustments. The experimental results show a 28.4% cost reduction, 30.3% lower emissions, and 21.8% improved warehouse efficiency. While computational overhead and real-time adaptability pose challenges, future enhancements will focus on energy-efficient AI, continuous learning, and explainable decision making. The framework advances sustainable automation, balancing operational optimization with environmental responsibility.

1. Introduction

The need for sustainability in operations has fueled the rapid evolution of supply chain management [1]. The modern market demands sustainability along with efficiency and cost reduction. Today’s customer base is more aware of the environmental and societal impact of the manufacturing process of the products they use. The supply chain is a major part of the process, and there is a growing pressure to integrate sustainability into the core of supply chain processes [2]. Industries and consumers worldwide have begun recognizing that supply chain practices directly impact environmental sustainability, resource utilization, and long-term economic stability [3]. This is why innovation in redefining supply chain management has become a pressing need. This paper addresses this need and proposes an innovative AI-driven solution.

Global trade has expanded more than ever in the history of mankind [4]. This naturally has increased the complexity of the supply chain network. Managing vast networks of suppliers, transportation routes, inventory levels, and customer demand requires meticulous coordination [5]. Although there are advanced and user-friendly management systems that help to manage a complex interdependent process, they cannot work autonomously. They are simply assistive systems that help keep track of everything, record all activities, and help make decisions [6]. These are known as traditional supply chain management systems. While these methods have worked for years, they are insufficient in the face of rising environmental concerns and fluctuating global demand. The operational inefficiencies, delays in decision making, frequent misallocation, and failure to cope with dynamic adjustment requirements are a few of the limitations of this existing traditional system which cause significant financial loss and ecological impact [7,8].

Recent advancements in AI have led to innovation in supply chain management as well. The transformative power of AI-integrated solutions has demonstrated promising outcomes in predicting demand, optimizing logistics, and reducing resource waste [9]. However, conventional AI is not adaptable and does not have the capability to understand context and make decisions by reasoning between multiple interdependent heterogeneous contexts. Standard AI models, while powerful, are typically designed to solve individual problems in isolation—demand forecasting, for instance, may not dynamically influence warehouse distribution strategies, even though the two are interconnected [10]. To overcome this limitation and elevate the application of AI in supply chain management to new heights, this paper proposes SustAI-SCM, a transformer-based model that learns, adapts, and optimizes supply chain processes dynamically, with human-like intelligence. Instead of reacting to the problems after they arise in SCM, the SustAI-SCM has been designed to anticipate and mitigate them in real time, making sustainable supply chain management an intelligent, data-driven process. The core contributions of this study are listed below:

• Sustainable process automation: The SustAI-SCM automates supply chain decision making, which reduces the need for manual intervention and lowers the risk of human errors.
• Agentic AI for dynamic optimization: Going beyond the static rule-based automation and incorporation of transformer model through agentic AI framework to autonomously refine supply chain processes, ensuring optimal resource allocation.
• Carbon footprint reduction: By optimizing logistics and inventory, the model effectively reduces transportation emissions and unnecessary energy consumption, which contributes to carbon footprint reduction.
• Cost efficiency and scalability: The AI-driven approach significantly lowers operational costs while maintaining flexibility to scale across different supply chain networks.

Traditional supply chain management systems rely on static, rule-based automation, which lacks adaptability and fails to dynamically adjust to changing operational constraints. Conventional AI models, while effective for specific tasks such as demand forecasting, operate in isolation and do not account for interdependencies between procurement, logistics, and inventory. This results in inefficiencies, misallocation of resources, and an inability to optimize operations in real time. To address these challenges, the SustAI-SCM framework leverages a transformer-based AI model that continuously learns from supply chain operations. Unlike traditional AI, it dynamically refines decision making by considering multiple interdependent factors, making autonomous adjustments based on real-time feedback. This research aims to demonstrate how agentic AI can enable a more efficient, self-optimizing, and sustainable supply chain management system.

The remainder of this paper is structured as follows: Section 2 presents a comprehensive review of the related literature on AI-driven supply chain automation and sustainability efforts. Section 3 details the dataset preparation, preprocessing steps, and the transformer-based model used in this study. The system’s implementation is outlined in Section 4, while Section 5 provides performance evaluations, including automation efficiency, sustainability impact, and a cost analysis. Section 6 discusses the framework’s limitations and proposes potential future research directions. Finally, the paper concludes with key insights and implications for the sustainable supply chain industry.

2. Literature Review

The application of AI in SCM is not new. As a matter of fact, it has been improving SCM for nearly a decade now [11]. However, the expansion of global trade and the volume of supply–demand have made SCM very complex. To keep pace with its ever-growing complexities, researchers have explored AI-driven solutions for inventory control, demand forecasting, and logistics optimization [12]. However, AI models developed to perform an isolated task for a particular part of the SCM fail to meet the sustainability demand of today’s interconnected and rapidly evolving SCM [13]. This is where the need for interconnected, intelligent automation systems has arisen. This section reviews the major contributions in AI-based supply chain automation, agentic AI applications, and sustainability-driven AI frameworks [14].

2.1. AI in Supply Chain Management

Supply chain management is a process that involves multiple interdependent and independent operations, from procurement to transportation and final distribution [15]. Modern SCM systems have AI-driven solutions integrated within them. However, these solutions are deterministic models that are mainly rule-based or particularly designed for a certain part of the overall supply chain operation [16]. While effective in stable conditions, these methods fail to adapt quickly to unexpected disruptions. As a result, they lead to inefficiencies and increased carbon footprints [17]. Two popular fields of AI, machine learning and deep learning, were used to predict demand patterns, optimize warehouse storage, and automate procurement decisions by Pasupuleti et al. [18]. The systematic review conducted by Shahzadi et al. [19] suggests that AI-driven supply chains can reduce costs by 15% to 30%, primarily through intelligent forecasting and process automation. Despite the tremendous potential for cost reduction and reducing the carbon footprint, its task-specific isolated nature stands as a major barrier to the benefits of AI in the supply chain [20]. Demand prediction models, for instance, do not dynamically communicate with routing systems, resulting in suboptimal transportation planning. Similarly, automated warehousing does not always align with real-time inventory replenishment strategies [21]. This lack of context awareness in AI solutions creates inefficiencies that contradict the very purpose of automation. This is where the innovation of the proposed SustAI-SCM fits it. It is an autonomous agentic AI-based approach that coordinates among multiple entities, understands the context, and automates the SCM process effectively and efficiently.

2.2. Agentic AI and Autonomous Decision Making

The primary difference between agentic AI and conventional AI solutions is how they interact with real-world problems [22]. Unlike traditional AI-driven solutions, agentic AI does not simply perform prediction and classifications; instead, it operates independently and makes context-aware decisions [23]. In an agentic AI framework, in most cases, the system does not need to rely on user input. As it is context-aware and works actively, it can easily adapt to the dynamic and evolving landscape of supply chain management [24]. At the core of the agentic AI, there are transformer-based large language models (LLMs), which continuously learn through reinforcement learning from new data and experience. Because of the reinforcement learning module, the agentic AI is capable of improving its knowledge base and adapting to keep performing autonomously [25]. Recent studies have explored agentic AI in finance, healthcare, and robotic process automation, yet its application in sustainable supply chains remains largely unexplored [26].

2.3. Sustainability Challenges and AI Solutions

Financial efficiency and sustainability are not the same in supply chain management. Achieving sustainability is more complex than financial efficiency. Reducing cost and increasing profit margin is enough for financial gains [27]. However, sustainability involves reducing waste, minimizing emissions, and ensuring responsible sourcing. This is what makes it challenging. According to the study conducted by Plambeck and Erica L [28], supply chain operations account for more than 80% of a company’s total greenhouse gas emissions. Research conducted by Chen et al. [29] addressed this issue and developed AI-driven carbon footprint tracking and logistics emission reduction models. Although this is a promising AI-driven solution, it does not work autonomously and requires human intervention. According to the findings of Brintrup et al. [30], AI-driven sustainable procurement can identify ethical suppliers based on historical records and compliance reports, but real-time supplier adaptability remains an open challenge. Another challenge is in market shift, which makes AI-driven inventory management systems ineffective [31]. The agentic AI with adaptability proposed in this paper is the ultimate solution to this problem; it ensures sustainability in supply chain management by adapting itself to the dynamic landscape of SCM.

2.4. Research Gaps and Objectives

The research objectives of this study are the guidelines for methodological development. The objectives have been specified based on the research gaps discovered through the literature presented in this section. The research gaps and corresponding research objectives are summarized in

Table 1. Identified research gaps and proposed solutions.

Research Gap	Proposed Solution
Lack of real-time adaptability in AI-driven supply chain models.	Implement an agentic AI framework capable of autonomous decision-making.
Isolated task automation leads to inefficiencies in sustainability.	Develop an integrated AI system that connects procurement, inventory, and logistics processes dynamically.
Inability to align AI-driven decision making with carbon footprint reduction strategies.	Introduce AI-driven sustainability tracking that actively adjusts supply chain operations to minimize environmental impact.

. This study aimed to achieve these goals, resulting in the development of SustAI-SCM, an intelligent supply chain process automation with agentic AI that ensures sustainability and cost efficiency.

3. Methodology

3.1. Framework Overview

The SustAI-SCM framework, illustrated in Figure 1, has been designed with two constraints at the heart of the design. These two constraints are sustainability and efficiency of supply chain operations using an intelligent AI-powered automation system. That is why this framework has gone beyond rule-based automation tools and uses the agentic AI principle. Because of using the concept of agentic AI, it can perform autonomously, going beyond traditional rule-based AI. It can make autonomous decisions, learn from operational feedback, and refine itself dynamically. The main goal of this framework is to anticipate inefficiencies rather than merely reacting to them [32]. In supply chain management, the core components, logistics, procurement, and inventory interact with each other and are dependent on each other. This framework understands the relation and is capable of ensuring proactive optimization for long-term sustainability and cost-effectiveness. The SustAI-SCM framework is not like conventional AI models, which are dedicated to performing a specific task and operate in isolation. This framework integrates multiple data sources, seamlessly enables multiple agents, and actively automates supply chain processes within a single decision-making ecosystem. This framework, as shown in Figure 1, consists of three core modules. The following list presents the modules and their roles and responsibilities:

• Perception module: This module has been designed to ingest both structured and unstructured data from various sources, including real-time sources. These data are related to logistics, supplier performance records, and demand fluctuation patterns. This module processes the data for the core agentic AI module.
• Agentic AI core: The agentic AI core module is the decision-making engine of the framework. It has been built using a transformer model with the facility for transfer learning and reinforcement learning for real-time adaptability. Unlike traditional static optimization, this module refines operational strategies based on evolving constraints.
• Action execution module: The third module is the action execution module. The decisions made by the agents are translated into actions in this module. That means this module translates the AI-generated insights into actionable supply chain tasks—whether it is adjusting procurement schedules, rerouting logistics for lower carbon emissions, or dynamically managing warehouse inventory.

One of the unique characteristics of the proposed SustAI-SCM is its ability to self-correct inefficiencies. That means even if it fails to ensure efficiency, it corrects itself and enhances efficiency. If the framework detects excessive emissions in a certain transportation route, it does not simply report it. Along with reporting, the SustAI-SCM actively adjusts logistics planning in real time. This adaptive intelligence aligns with the sustainability goals of minimizing waste, reducing carbon footprints, and ensuring ethical procurement. This is how the proposed SustAI-SCM framework makes the difference. The optimization tactics of the proposed SustAI-SCM are presented in Equation (1). It defines the total cost $\Phi$ of a supply chain as a function of logistics $\Lambda$, procurement efficiency $\Gamma$, and environmental impact $\Delta$. The AI system continuously updates this function to minimize costs while maximizing sustainability.

$$\Phi =\sum _{\xi =1}^{\Omega }\left(\alpha {\Lambda }_{\xi }+\beta {\Gamma }_{\xi }+\gamma {\Delta }_{\xi }\right)-\psi \sum _{\rho =1}^{\Theta }\left({\displaystyle \frac{\partial {{\rm Y}}_{\rho }}{\partial \tau }}\right)$$

(1)

In Equation (1), $\alpha$, $\beta$, and $\gamma$ are dynamic weight parameters. These parameters are updated and adjusted based on system feedback. This feedback mechanism ensures that cost reduction and sustainability objectives are balanced. The term $\psi$ represents the adaptive correction factor. It allows the framework to fine-tune its optimization strategy in real time. The derivative term ${\displaystyle \frac{\partial {{\rm Y}}_{\rho }}{\partial \tau }}$ models the rate of change in sustainable performance metrics over time. It ensures a continuously improving system. This framework enables supply chain networks to transition from static, inefficient workflows to an intelligent, responsive system that not only predicts issues but also autonomously resolves them.

Definition of Agentic AI in Supply Chain Management

The term “agentic AI” refers to artificial intelligence systems that possess autonomy, adaptability, and proactive decision-making capabilities. Unlike traditional rule-based AI, agentic AI systems are designed to function with minimal human intervention by dynamically assessing their environment, optimizing actions, and continuously learning from new data. These systems are capable of reasoning across multiple interdependent variables, executing decisions in real time, and refining their strategies based on operational feedback.

In the context of the proposed SustAI-SCM framework, the agentic AI paradigm is realized through a transformer-based model that continuously optimizes procurement, logistics, and inventory management. The agentic AI operates through the following core components:

• Design: The system is structured to autonomously adapt its decision-making logic, leveraging reinforcement learning to refine optimization strategies.
• Operation: The AI actively processes real-time supply chain data, assessing multiple competing objectives such as cost efficiency, sustainability, and operational resilience.
• Communication: The model integrates with various supply chain modules, allowing seamless information exchange between procurement, logistics, and warehouse management.
• Reasoning: Unlike static AI models, the agentic AI in SustAI-SCM evaluates multiple interdependent factors simultaneously, making context-aware adjustments without requiring explicit rule-based programming.

By implementing these properties, the proposed framework embodies the key attributes of agentic AI, making it capable of managing complex supply chain dynamics autonomously.

3.2. Dataset Preparation and Processing

Developing the SustAI-SCM framework has been possible because of the success of developing a unique dataset consisting of real-time logistics records, procurement transactions, supplier sustainability scores, and warehouse inventory logs. It also contains hundreds of case studies and research papers related to sustainable supply chains. Unlike traditional supply chain datasets that focus primarily on cost efficiency, this dataset is designed to balance financial and sustainability metrics simultaneously. The transformer model of the SustAI-SCM framework learns from this dataset and becomes effective in decision making to ensure a sustainable supply chain. However, the raw dataset is not ready to train the transformer model directly. Algorithm 1 demonstrates the dataset preprocessing method used in this paper.

Algorithm 1: Data Preprocessing Pipeline for SustAI-SCM

Raw dataset $\Xi =\{\Pi ,\Lambda ,{\rm Y}\}$ Processed dataset ${\Xi }^{*}$ Step 1: Data Collection Retrieve $\Xi$ from $\mathrm{DataSources}\left(\right)$ Step 2: Cleaning and Normalization $x\in \Xi$   if $x=⌀$ then $x\leftarrow \mathrm{Impute}\left(x\right)$        if x is duplicate or inconsistent then $\mathrm{Remove}\left(x\right)$ $x\leftarrow \mathrm{Normalize}\left(x\right)$ Step 3: Tokenization and Embedding $T\in \Xi$$T\leftarrow \mathrm{Tokenize}\left(T\right)$$T\leftarrow \mathrm{Embed}\left(T\right)$ Step 4: Sequence Standardization $T\in \Xi$$T\leftarrow \mathrm{PadTruncate}(T,\Theta )$ Step 5: Dataset Splitting ${\Xi }^{*}\leftarrow \mathrm{SplitDataset}(\Xi ,70\%,15\%,15\%)$ return${\Xi }^{*}$

The dataset used in this study was obtained from a single multinational logistics company that operates across 30 countries. This dataset includes procurement transactions, logistics records, and sustainability reports collected over a period of six months. All data were anonymized to comply with privacy regulations, ensuring confidentiality while maintaining the integrity of supply chain operations. The use of real-world data from an industry leader strengthens the practical applicability of the SustAI-SCM framework in large-scale supply chain environments.

The final dataset is categorized into three key groups: Procurement data ($\Pi$), logistics data ($\Lambda$), and sustainability indicators (${\rm Y}$).

Table 2. Dataset categories and corresponding statistics.

Category	Description	Records	Attributes
$\Pi$	Procurement transactions	350,000	12
$\Lambda$	Logistics and transportation	1,200,000	15
${\rm Y}$	Carbon footprint and sustainability	900,000	10
Total	-	2,450,000	-

summarizes the dataset components.

The dataset is mathematically structured as a multi-source data aggregation model, where each record is linked across multiple dimensions. Equation (2) formally represents the integration of different dataset components.

$$\Xi =\bigcup _{\zeta =1}^{\Omega }\left({\Pi }_{\zeta },{\Lambda }_{\zeta },{{\rm Y}}_{\zeta }\right),\forall \zeta \in {\mathbb{Z}}^{+}$$

(2)

Here, $\Xi$ represents the full dataset containing multiple interconnected attributes. The index $\zeta$ refers to an individual supply chain event, ensuring that procurement, logistics, and sustainability records are mapped together for each transaction. The use of ${\mathbb{Z}}^{+}$ guarantees that only positive-indexed supply chain events are considered, eliminating zero or negative-value placeholders.

3.2.1. Data Cleaning and Normalization

There were numerous inconsistencies in the dataset including unit mismatches, redundant timestamps, missing values, irrelevant characters, URLs, and many other different types of strings. To address these, first of all, the dataset was cleaned manually, and after that, normalization was performed to bring the documents into a generalized form. The general form of the normalization function is given in Equation (3) [33].

$$\Psi \left(x\right)={\displaystyle \frac{x-{\mu }_x}{{\sigma }_x}},x\in \Xi$$

(3)

In Equation (3), $\Psi \left(x\right)$ represents the normalized value of attribute x, ${\mu }_x$ is the mean value of the attribute across all records, and ${\sigma }_x$ is the standard deviation. This transformation ensures that all features are centered around zero with a variance of one, preventing model bias towards larger numerical scales.

3.2.2. Tokenization and Embedding

The normalized dataset is not yet ready to train the SustAI-SCM transformer model. In the next phase, the data are tokenized. In this process, the texts are divided into small tokens, as defined in Equation (4), where $\vartheta \left(T\right)$ represents the set of tokenized sequences from text corpus T, and $\mathcal{V}$ is the vocabulary containing all recognized tokens [34].

$$\vartheta \left(T\right)=\bigcup _{\eta =1}^{\Theta }\left\{t_{\eta }\mid t_{\eta }\in \mathcal{V}\right\}$$

(4)

After tokenization, the data are transformed into an embedding vector. This is the final form of the data that the transformer model can directly utilize to learn from. Each tokenized record is embedded using a trainable matrix that learns semantic relationships between words, as shown in Equation (5), where $\mathbf{E}$ is the embedding matrix, $e_i$ represents the embedding vector for the ith token, and d is the embedding dimension [35].

$$\mathbf{E}={\left[e_1,e_2,\dots ,e_n\right]}^T,e_i\in {\mathbb{R}}^d$$

(5)

3.2.3. Sequence Standardization

The transformer model used in the SustAI-SCM framework can handle a fixed number of input tokens. A lower or higher number of tokens creates processing anomalies. That is why it is essential to standardize the sequencing. Since supply chain records vary in length, all sequences must be padded or truncated to ensure uniform input size. This step prevents inconsistencies when training the AI model. Padding is applied based on Equation (6), where ${\vartheta }^{\prime }\left(T\right)$ is the processed sequence, and $\left[\mathtt{PAD}\right]$ is a unique token that is added to ensure that all inputs meet the fixed sequence length $\Theta$ [36].

$${\vartheta }^{\prime }\left(T\right)=\left\{\begin{array}{cc}{t}_{\eta },\hfill & \mathrm{if}\eta \le \Theta ,\hfill \\ \left[\mathrm{PAD}\right],\hfill & \mathrm{if}\eta >\Theta \hfill \end{array}\right.$$

(6)

3.2.4. Final Dataset Split

In the last phase of the dataset preprocessing, the dataset was split into training, validation, and testing subsets. In this paper, a stratified sampling approach was used to ensure data quality in every subset. The split ratio is shown in

Table 3. Dataset splitting strategy.

Subset	Percentage (%)	Total Records
Training set	70	1,715,000
Validation set	15	367,500
Testing set	15	367,500

[37].

3.3. Transformer Model Design

The core decision-making engine of the proposed SustAI-SCM framework is a transformer model. In this paper, the design of the transformer model has been customized to optimally perform the designated tasks. The structure of the transformer model is presented in

Table 4. Transformer model architecture for SustAI-SCM.

Layer	Type	Heads	Dimensions	Description
1	Input Embedding	-	60,000 $\times 512$	Maps tokenized supply chain data into dense vectors
2	Positional Encoding	-	$128\times 512$	Encodes word order information for sequence modeling
3–10	Encoder Blocks	-	-	Consists of multi-head attention and feed-forward layers
Details of Transformer Encoder Block (3–10)
3a–10a	Multi-Head Attention	12	$512\times 512$	Computes attention across sequence positions
3b–10b	Layer Normalization	-	512	Stabilizes activations and accelerates convergence
3c–10c	Feed-Forward Network	-	$512\to 2048\to 512$	Applies non-linear transformations for feature extraction
3d–10d	Dropout	-	-	Prevents overfitting by randomly deactivating neurons
11	Output Projection	-	$512\times$ 60,000	Converts encoder output to vocabulary logits
12	Softmax Layer	-	60,000	Outputs probability distribution over tokens

.

3.3.1. Architectural Components

As presented in Table 4, the transformer model designed for this paper is composed of three primary layers. These layers are the input embedding layer, multi-head self-attention module, and feed-forward network (FFN). The embedding layer is responsible for converting raw data into a vectorized representation, ensuring compatibility with the transformer architecture. The multi-head self-attention layer is responsible for capturing long-range dependencies across supply chain variables, ensuring that optimization decisions account for multiple interconnected factors. The third layer, the feed-forward network (FFN), is responsible for processing the attention-weighted inputs to refine predictions, optimize sustainability-driven decision making, and reduce computational noise.

3.3.2. Embedding Layer

The concept of the working principle of the embedding layer of the transformer model considers an input sequence $\mathbf{\Theta }$ consisting of N supply chain events. Each event ${\mathit{\theta }}_i$ is first mapped into an embedding space using a trainable matrix $\mathcal{E}$, as defined in Equation (7), where ${\mathit{\xi }}_i$ represents the embedded feature vector, and $\mathcal{E}$ is the learnable embedding matrix which is passed to the multi-head attention layer [38].

$${\mathit{\xi }}_i=\mathcal{E}·{\mathit{\theta }}_i^T,\forall i\in \{1,2,\dots ,N\}$$

(7)

3.3.3. Multi-Head Attention Mechanism

The multi-head self-attention mechanism enables the transformer model of the SustAI-SCM framework to understand contextual dependencies; unlike traditional recurrent neural networks, which process input sequentially, self-attention allows the model to weigh the importance of each supply chain event relative to all others. This is one of the critical properties of the proposed SustAI-SCM framework’s transformer model in sustainable supply chain management, where past supplier behaviors, carbon footprint patterns, and transportation emissions must all be evaluated dynamically. The conceptual design of each attention head is that the attention score ${\mathcal{A}}_{ij}$ between two input embeddings ${\mathit{\xi }}_i$ and ${\mathit{\xi }}_j$ is computed using the scaled dot-product mechanism presented in Equation (8) [39].

$${\mathcal{A}}_{ij}={\displaystyle \frac{\left(\mathcal{Q}·{\mathit{\xi }}_i\right)·{\left(\mathcal{K}·{\mathit{\xi }}_j\right)}^T}{\sqrt{d}}}$$

(8)

In Equation (8), $\mathcal{Q}$ and $\mathcal{K}$ are trainable weight matrices for the query and key vectors and d represents the dimensionality of the embeddings, ensuring numerical stability during training. After computing attention scores for all input elements, the final context-aware representation is obtained by applying a softmax function, which is defined in Equation (9), where $\mathcal{V}$ is the trainable weight matrix for the value vectors. This equation ensures that the most relevant supply chain features influence each decision [40].

$${\chi }_i=\sum _{j=1}^N\mathrm{softmax}\left({\mathcal{A}}_{ij}\right)·\left(\mathcal{V}·{\mathit{\xi }}_j\right)$$

(9)

3.3.4. Feed-Forward Network and Output Layer

The attention mechanism refines contextual relationships. After that, the signal flows into the feed-forward network (FFN). The FFN performs a two-layer transformation to improve decision accuracy by following the mathematical principle presented in Equation (10) [41].

$${\mathit{\psi }}_i=\sigma \left({\mathcal{W}}_1·{\chi }_i+{\mathcal{B}}_1\right)·{\mathcal{W}}_2+{\mathcal{B}}_2$$

(10)

In Equation (10), ${\mathcal{W}}_1,{\mathcal{W}}_2$ are weight matrices for linear transformations, ${\mathcal{B}}_1,{\mathcal{B}}_2$ are bias terms, and $\sigma (·)$ represents a non-linear activation function (ReLU). The final output $\mathsf{\Psi }$ is generated using a probability distribution over potential optimization actions, as presented in Equation (11).

$$\mathbf{\Psi }=\mathrm{softmax}\left(\mathcal{O}·{\mathit{\psi }}_i\right)$$

(11)

In Equation (11), $\mathcal{O}$ is the output weight matrix. This final step ensures that supply chain decisions, such as optimal procurement adjustments or transportation route selections, are made based on real-time environmental and operational data.

3.3.5. Model Adaptability and Learning

The proposed SustAI-SCM transformer is not like conventional rule-based supply chain systems, which operate on predefined static rules. Instead, it continuously refines its decision-making strategies. Conceptually, it dynamically updates its parameters. The dynamic parameters are $\mathcal{E},\mathcal{Q},\mathcal{K},\mathcal{V},{\mathcal{W}}_1,{\mathcal{W}}_2,\mathcal{O}$. The backpropagation with an adaptive loss function [42], defined in Equation (12), is used to update the weights of the parameters.

$$\mathcal{L}=\sum _{\tau =1}^T\left(\ell ({\mathbf{\Psi }}_{\tau },{\mathbf{{\rm Y}}}_{\tau })+\lambda ·{\displaystyle \frac{\partial }{\partial \tau }}\mathcal{S}\left(\tau \right)\right)$$

(12)

In Equation (12), $\ell ({\mathbf{\Psi }}_{\tau },{\mathbf{{\rm Y}}}_{\tau })$ is the cross-entropy loss between the predicted and actual outcomes, $\lambda$ is a regularization factor that adjusts for sustainability constraints, and $\mathcal{S}\left(\tau \right)$ represents the cumulative sustainability performance over time. By incorporating sustainability-aware learning, the model not only improves supply chain efficiency but also actively reduces environmental impact over iterative updates.

3.4. Training and Hyperparameter Optimization

The transformer model of the proposed SustAI-SCM framework is fine-tuned through the transfer learning method in both supervised and unsupervised ways. The learning curve is illustrated in Figure 2. The primary training objective is to minimize decision errors while simultaneously optimizing sustainability and cost efficiency across procurement, logistics, and inventory processes.

3.4.1. Loss Function and Objective

To ensure that the model balances efficiency and sustainability, a hybrid loss function is used. The total loss function ${\mathcal{L}}_{\mathrm{total}}$ consists of a supervised learning error component and a sustainability-aware regularization term, as expressed in Equation (13) [43].

$${\mathcal{L}}_{\mathrm{total}}=\sum _{\tau =1}^T\left({\mathcal{L}}_{\sup }+\lambda ·{\displaystyle \frac{\partial }{\partial \tau }}\mathcal{S}\left(\tau \right)\right)$$

(13)

In Equation (13), ${\mathcal{L}}_{\sup }$ represents the supervised loss, computed using cross-entropy between the predicted and actual supply chain optimization decisions, $\lambda$ is a regularization factor that dynamically adjusts based on sustainability constraints, and $\mathcal{S}\left(\tau \right)$ tracks the cumulative sustainability score over time.

3.4.2. Optimization Algorithm

It is essential to ensure that the weights of the hidden nodes of the feed-forward neural network of the transformer model of the SustAI-SCM framework are optimally updated. To ensure this, an adaptive gradient optimization approach is employed. The Adaptive Moment Estimation (ADAMW) optimizer is used as the optimization algorithm to update the weights optimally and prevent overfitting. It updates the weight by following the mathematical principle expressed in Equation (14) [44].

$${\mathit{\theta }}_{t+1}={\mathit{\theta }}_t-\eta ·{\displaystyle \frac{{\widehat{\mathbf{m}}}_t}{\sqrt{{\widehat{\mathbf{v}}}_t}+ϵ}}$$

(14)

In Equation (14), ${\mathit{\theta }}_t$ represents the model parameters at time step t, $\eta$ is the learning rate, ${\widehat{\mathbf{m}}}_t$ and ${\widehat{\mathbf{v}}}_t$ are the bias-corrected first- and second-moment estimates of the gradient, and $ϵ$ is a small constant to prevent division by zero.

3.4.3. Learning Rate Scheduling

The learning rate is a crucial factor during network training. If it is too high, it will speed up the process. However, the probability of overshooting is high and the learning process may miss the optimal solution. If it is set to too small a value, then there is a chance of suboptimal convergence. To balance these concerns, a warm-up-and-decay learning rate schedule is applied by following the mathematical principle expressed in Equation (15) [45].

$${\eta }_t={\eta }_{max}·min\left(1,{\displaystyle \frac{t}{t_{\mathrm{warmup}}}}\right)·exp\left(-{\displaystyle \frac{t}{\tau }}\right)$$

(15)

In Equation (15), ${\eta }_t$ is the learning rate at step t, ${\eta }_{max}$ is the peak learning rate, $t_{\mathrm{warmup}}$ is the number of warm-up steps, and $\tau$ controls the rate of exponential decay.

3.4.4. Hyperparameter Tuning

One of the valuable contributions of this study is the discovery of the optimal hyperparameters, presented in

Table 5. Optimized hyperparameters for SustAI-SCM.

Hyperparameter	Optimized Value
Batch Size	32
Learning Rate (Peak)	$5\times {10}^{-5}$
Warm-up Steps	500
Epochs	15
Weight Decay	$0.01$
Gradient Clipping Threshold	$1.0$
Dropout Rate	$0.1$
Attention Heads	8
Optimizer	AdamW
Scheduler	Linear Warm-up with Exponential Decay

. This facilitates the reproducibility of the transformer model trained in this paper. The SustAI-SCM framework is trained using a grid search approach, evaluating different configurations of batch size, dropout rates, learning rates, and attention heads. Table 5 summarizes the best-performing hyperparameter values.

3.4.5. Computational Setup

The transformer model of the SustAI-SCM framework was trained on a massive dataset. The model itself is a heavy model. That is why it required a high-performance computing environment to train the model. To ensure smooth training, the model was deployed on a high-performance computing cluster whose configuration is presented in

Table 6. Computational environment for training.

Component	Specification	Manufacturer	Model/Version	Location (City, Country)
Processor	Intel Xeon Gold 6238R (28 cores, 2.2 GHz)	Intel Corporation	Xeon Gold 6238R	Santa Clara, CA, USA
GPU	NVIDIA A100 Tensor Core (40 GB memory)	NVIDIA Corporation	A100	Santa Clara, CA, USA
RAM	256 GB DDR4	Samsung Electronics	DDR4	Suwon, South Korea
Storage	2 TB NVMe SSD	Western Digital	NVMe SSD	San Jose, CA, USA
Operating System	Ubuntu 22.04 LTS	Canonical Ltd.	22.04 LTS	London, UK
Frameworks	PyTorch 2.0, CUDA 11.7, Transformers 4.31.0	–	–	–

. The entire training process required approximately 8 days, 14 h, and 39 min to converge on a stable model. This investment in computational resources ensured that the framework delivers highly accurate and sustainability-aware supply chain automation.

4. Implementation

Supply chain data sometimes reflect the properties of big data [46]. Even on a regular day, the volume of supply chain-related data is massive. The proposed SustAI-SCM framework needs to perform computationally intensive operations to process these data. That is why it was implemented using a high-performance computing environment. This section provides an overview of the hardware infrastructure, software stack, and deployment considerations.

4.1. Hardware Infrastructure

The hardware infrastructure used to develop the proposed SustAI-SCM framework is listed in

Table 7. Hardware infrastructure for SustAI-SCM.

Component	Specifications	Manufacturer	Model/Version	Location (City, Country)
Processor	Intel Xeon Gold 6248R (24 cores, 3.0 GHz)	Intel Corporation	Xeon Gold 6248R	Santa Clara, CA, USA
GPU	NVIDIA A100 Tensor Core (40 GB HBM2)	NVIDIA Corporation	A100	Santa Clara, CA, USA
RAM	512 GB DDR4	Samsung Electronics	DDR4	Suwon, South Korea
Storage	4 TB NVMe SSD	Western Digital	NVMe SSD	San Jose, CA, USA
Network Interface	10 Gbps Ethernet	Broadcom Inc.	NetXtreme 10G	San Jose, CA, USA
Operating System	Ubuntu 22.04 LTS	Canonical Ltd.	22.04 LTS	London, UK

. The transformer model used in the SustAI-SCM framework is not a standalone system. It evolves with time through reinforcement learning. The hardware listed in Table 7 was carefully selected to ensure a smooth experience even during the learning in progress.

4.2. Software Stack

The design principle of the proposed SustAI-SCM framework follows three principles: modularity, scalability, and efficient AI integration. To ensure that these features are reflected in the implementation architecture, the SustAI-SCM was built using a combination of deep learning frameworks, cloud computing tools, and API-based communication protocols. A list of the core software used in the development process is given in

Table 8. Software stack for SustAI-SCM implementation.

Component	Technology Used (Version)	Developer/Vendor	Location (City, Country)
Programming Language	Python 3.9.13	Python Software Foundation	Wilmington, NC, USA
Machine Learning Framework	PyTorch 2.0.1, Transformers 4.31.0	Meta AI, Hugging Face	Menlo Park, CA, USA; Paris, France
Data Processing	Pandas 1.5.3, NumPy 1.23.5, Spark 3.3.2	NumFOCUS, Apache Foundation	Austin, TX, USA; Forest Hill, IN, USA
Database	PostgreSQL 15.1, MongoDB 6.0	PostgreSQL Dev Group, MongoDB Inc.	Berkeley, CA, USA; New York, NY, USA
API Deployment	Flask 2.2.5, FastAPI 0.95.2	Pallets, FastAPI Community	San Francisco, CA, USA
Cloud Integration	AWS Lambda, Google Cloud Functions	Amazon AWS, Google Cloud	Seattle, DC, USA; Mountain View, CA, USA
Containerization	Docker 24.0.2, Kubernetes 1.26.3	Docker Inc., CNCF	Palo Alto, CA, USA; San Francisco, CA, USA
Monitoring	Prometheus 2.45.0, Grafana 9.5.2	CNCF, Grafana Labs	San Francisco, CA, USA

.

The proposed SustAI-SCM framework uses Docker and Kubernetes, which ensure scalability in a distributed environment. As it runs in the Docker container, it can be operated from both on-premise and cloud computers. The use of PyTorch with GPU acceleration enabled efficient Transformer-based processing for real-time supply chain optimization.

4.3. Case Study: Real-World Deployment in a Supply Chain Network

SustAI-SCM was deployed in a multinational logistics company that has distribution centers in more than 30 countries. This company performs over 500,000 shipments per month. The proposed SustAI-SCM was allowed to be applied in a newly launched experimental distribution center, which is not yet commercially operational. In addition, the experiment had access to the data of the entire company for the past six months, which have been used to replicate the real-world deployment environment. After applying the proposed SustAI-SCM, the fuel consumption was lowered by 22.8%, leading to significant carbon footprint reduction. In addition, the procurement adjustments were much faster, enhancing the overall efficiency by more than 31.4%. Moreover, it significantly reduced inventory waste by reducing stockpiling inefficiencies by 19.6%. At the same time, the AI-driven dynamic route selection improved delivery efficiency by 27.2%.

Table 9. Performance gains from SustAI-SCM deployment.

Metric	Before Deployment	After Deployment	Improvement (%)
Fuel consumption (liters)	1,200,000	927,600	22.8
Supplier selection time (days)	14.2	9.8	31.4
Warehouse overstocking (tons)	8500	6837	19.6
Average delivery efficiency (%)	72.3	91.9	27.2

presents a before-and-after comparison of key operational metrics, and they are visually represented in Figure 3.

Selection of Operational Metrics

The selection of parameters presented in Table 9 was based on their direct impact on supply chain efficiency, sustainability, and cost reduction. These metrics—fuel consumption, supplier selection time, warehouse overstocking, and average delivery efficiency—were identified as key performance indicators (KPIs) that reflect both operational effectiveness and environmental impact. The choice of these specific parameters was guided by industry best practices and prior studies on AI-driven supply chain management. Fuel consumption and warehouse overstocking are critical factors influencing both cost and sustainability, while supplier selection time directly affects procurement efficiency. Similarly, delivery efficiency represents a fundamental measure of logistical performance. The SustAI-SCM framework is designed to be flexible, allowing for the integration of additional parameters based on the needs of specific supply chain operations. Future research can explore the inclusion of real-time demand fluctuations, supplier reliability scores, or energy consumption metrics to further enhance decision-making capabilities.

4.4. Real-Time Sustainability Integration

One of the core features of the proposed SustAI-SCM is the real-time sustainability monitoring module integration. This feature extends the ability of this framework to track the live data related to carbon emissions, sustainability compliance, and dynamic route reconfiguration. AI-driven predictions on fuel consumption and CO₂ impact were used to track carbon emissions. The agent of the agentic AI framework explores various updated sources, learns recent regulations along with existing active ones, and ensures procurement choices follow green sourcing regulations. The global live traffic update is utilized by the AI agency to adjust transportation paths to reduce environmental impact automatically.

5. Performance Evaluation and Results

The effectiveness of the SustAI-SCM framework was assessed through extensive experimentation, ensuring that it met its core objectives of supply chain optimization, cost efficiency, and sustainability integration. This section presents the evaluation methodology, experimental results, and impact analysis.

5.1. Confusion Matrix Analysis

Initially, the ability of the transformer model of the proposed SustAI-SCM framework to classify tasks from context was evaluated using a confusion matrix, illustrated in Figure 4. In the confusion matrix analysis, 650 small paragraphs from the testing dataset were used. These paragraphs refer to six different types of tasks. The experimental result shows that the proposed SustAI-SCM framework can identify the tasks with 97.11% accuracy, 96.98% precision, 96.44% recall, and an F1 score of 96.56%. This indicates that the transformer model is well trained and capable of understanding the context of supply chain management.

5.2. Scalability and Robustness

The volume of supply varies significantly depending on the time and trends of the year. That is why any supply chain management system must be scalable and robust. To test the scalability of the SustAI-SCM framework, experiments were conducted across varying workloads, from 50,000 to 2,500,000 supply chain events.

Table 10. Scalability and robustness analysis with processing time.

Events Processed	Automation Rate (%)	Error Rate (%)	CPU Utilization (%)	Processing Time (s)
50,000	91.5	2.1	45	1.2
200,000	90.2	2.5	57	4.8
500,000	88.7	3.1	71	12.3
1,000,000	87.2	3.8	85	24.7
2,500,000	85.5	4.4	92	62.5

summarizes the key findings.

The model maintained high accuracy and low error rates, even under heavy workloads. However, beyond 2,000,000 supply chain events, CPU usage exceeded 85%, indicating potential hardware constraints. Figure 5 demonstrates the scalability and robustness of the proposed SustAI-SCM.

5.3. Inference Time Analysis

The inference time is crucial for any system developed using a transformer model. At the same time, depending on the volume of queries, a longer inference time can significantly impact the supply chain management system. To analyze the inference time of the proposed SustAI-SCM, an experiment on various tasks, along with multiple simultaneous queries, was performed. The experimental data are listed in

Table 11. Inference time for supply chain optimization tasks.

Task	Single Query (ms)	100 Simultaneous Queries (ms)
Supplier selection	180	245
Route optimization	140	210
Inventory prediction	160	230
Demand forecasting	130	195
Carbon emission estimation	170	235

. According to the experimental performance, the proposed SustAI-SCM is fast in responding to single and multiple queries simultaneously.

Figure 6 shows a visual comparison of the inference time with single and multiple simultaneous queries. According to this comparison, the system maintains under 200 ms inference time in both cases. That means it is considered a real-time system that can respond and perform in real time.

5.4. Sustainability Impact Assessment

One of the primary objectives of SustAI-SCM is to enhance sustainability. The impact was measured using CO₂ emission reductions, fuel efficiency improvements, and waste minimization. The sustainability gains from deploying the system are outlined in

Table 12. Sustainability impact assessment.

Metric	Before Deployment	After Deployment	Reduction
CO2 emissions (kg/month)	15,600	10,870	30.3%
Fuel consumption (liters/month)	1,250,000	925,000	26.0%
Warehouse overstock (tons)	8300	6490	21.8%
Transportation costs (USD)	240,000	188,500	21.4%

The difference in warehouse overstock values between Table 9 and Table 12 arises due to different measurement periods. Table 9 reflects the experimental setup, while Table 12 reports long-term sustainability impact over a 12-month period.

.

Figure 7 illustrates the sustainability improvements over 12 months following the deployment of the SustAI-SCM framework. The plots show a gradual reduction in CO₂ emissions, fuel consumption, warehouse overstock, and transportation costs, indicating enhanced supply chain efficiency and lower environmental impact. The downward trend across all metrics highlights the system’s ability to drive sustainable logistics, optimize procurement, and minimize waste in real-time supply chain operations.

5.5. Process Optimization and Cost Savings

At the heart of the all-positive impact of the proposed SustAI-SCM are process optimization and cost savings. Because of process automation through agentic AI, the framework performs much more efficiently and significantly reduces costs. The efficient operation facilitates sustainability by reducing carbon emissions, fuel cost consumption, overstock, and transport costs. The overall process improvement is measured using the Supply Chain Optimization Index (${\mathcal{O}}_{SCM}$), which is defined in Equation (16).

$${\mathcal{O}}_{SCM}={\displaystyle \frac{{\sum }_{\tau =1}^T\left({\eta }_{\tau }{\mathcal{R}}_{\tau }+{\xi }_{\tau }{\mathcal{S}}_{\tau }-{\zeta }_{\tau }{\mathcal{C}}_{\tau }\right)}{T}}$$

(16)

In Equation (16), ${\mathcal{R}}_{\tau }$ represents cost reductions at time $\tau$, ${\mathcal{S}}_{\tau }$ represents sustainability improvements, ${\mathcal{C}}_{\tau }$ represents computational overhead, and ${\eta }_{\tau },{\xi }_{\tau }$, and ${\zeta }_{\tau }$ are weight factors based on business priorities. The cost savings and process optimizations obtained using SustAI-SCM are presented in

Table 13. Detailed operational cost reduction analysis.

Cost Component	Before (USD)	After (USD)	Savings (%)
Manual labor	80,000	38,500	51.9
Transportation costs	240,000	188,500	21.4
Overstocking losses	90,000	68,200	24.2
Procurement inefficiencies	60,000	42,500	29.2
Energy consumption	35,000	25,600	26.9
Inventory holding costs	45,000	31,700	29.6
Returns and refund processing	55,000	38,300	30.3
Total Cost	605,000	433,300	28.4

.

Figure 8 illustrates the operational cost reduction achieved through the deployment of the SustAI-SCM framework. It compares before-and-after costs across key supply chain components, including manual labor, transportation, overstocking, procurement inefficiencies, energy consumption, inventory holding, and returns processing. The blue line represents costs before deployment, while the green dashed line shows the reduced costs post-deployment. The red dotted line highlights the percentage of savings for each component. The total cost reduction of 28.4% demonstrates the framework’s effectiveness in optimizing financial and operational efficiency while promoting sustainability.

6. Limitations and Future Directions

While the SustAI-SCM framework offers significant advancements in supply chain automation, sustainability integration, and cost reduction, it is not without limitations. Identifying these challenges provides insights into areas for future research and development.

6.1. Limitations

6.1.1. Limited Generalization Across Industries

The dataset used for training SustAI-SCM was derived from a specific set of supply chain operations, specifically within the logistics and procurement domains. While the model generalizes well within these contexts, its performance in other industries (e.g., pharmaceutical or perishable goods supply chains) remains untested. Variations in business policies, regulatory constraints, and inventory structures may impact the model’s adaptability.

6.1.2. Computational Overhead and Energy Consumption

Despite its efficiency in optimizing supply chain operations, training and running the transformer model incurs significant computational costs. The real-time inference process, while fast, requires high-performance GPUs and cloud computing resources, which may not be feasible for small and medium-sized enterprises (SMEs). Additionally, the energy consumption of AI-driven systems raises concerns regarding sustainability in computing itself, which may offset some of the environmental benefits achieved through process optimization.

6.1.3. High Training Time but Efficient Inference

One of the notable constraints of the SustAI-SCM framework is its high training time due to the complexity of transformer-based optimization. However, it is important to note that, once trained, the system performs real-time inference with minimal latency, ensuring that supply chain decisions are executed without delay. This makes the framework highly scalable and practical for real-world industry applications, despite the initial computational overhead during model training.

6.2. Future Directions

6.2.1. Expanding Industry-Specific Training Datasets

To improve adaptability across industries, future iterations of SustAI-SCM will incorporate datasets from pharmaceuticals, food logistics, and industrial manufacturing. This expansion will enable the framework to capture domain-specific constraints, refining its decision making for diverse supply chain environments.

6.2.2. Implementing Energy-Efficient AI Optimization

AI-driven supply chain models demand high computational resources, necessitating energy-efficient optimization. Future research will focus on lightweight transformer architectures and quantization techniques to reduce computational overhead while preserving accuracy. This will make SustAI-SCM more sustainable and accessible for organizations with limited resources.

6.2.3. User-Customizable Optimization Weights

One of the potential enhancements for the SustAI-SCM framework is incorporating user-defined customization for parameter adjustments. In its current implementation, the model dynamically optimizes supply chain operations based on predefined weights assigned to cost, efficiency, and sustainability factors. However, future iterations will introduce an interactive interface that allows project managers to assign different weight parameters based on their specific priorities. For example, a logistics-focused project may prioritize minimizing transportation costs, while another project may emphasize reducing warehouse overstock. By enabling flexible weight assignments, SustAI-SCM can cater to diverse project needs, enhancing its adaptability across different industries and operational constraints.

6.2.4. Reducing Training Time Through Efficient Architectures

To further enhance the feasibility of SustAI-SCM for industry-wide adoption, future research will focus on optimizing the training efficiency of the framework. Potential improvements include leveraging lightweight transformer architectures, model distillation techniques and distributed training approaches. These optimizations will reduce computational requirements while maintaining model accuracy, ensuring that organizations with limited resources can also deploy and benefit from the system.

7. Conclusions

The increasing complexity of modern supply chain management necessitates innovative solutions that can balance efficiency, cost reduction, and sustainability. This paper introduced SustAI-SCM, an agentic AI-driven framework that leverages transformer-based intelligence to automate key supply chain processes while ensuring environmental impact minimization. By integrating real-time decision making, adaptive logistics optimization, and sustainable procurement strategies, the proposed system addresses critical inefficiencies in conventional supply chain models.

A comprehensive experimental evaluation demonstrated the system’s effectiveness in reducing operational costs, optimizing logistics, and minimizing carbon emissions. The results indicate a significant improvement in logistics efficiency, where dynamic route optimization and real-time decision adjustments enhanced delivery precision by 27.2%. Manual interventions and operational waste were reduced, leading to an overall 28% cost savings. Additionally, sustainability gains were evident, as carbon emissions decreased by 30.3%, fuel consumption was lowered by 26%, and warehouse overstocking inefficiencies were reduced by 21.8%. These findings reinforce the potential of AI-driven supply chain automation to create sustainable and economically viable business operations.

As industries worldwide shift toward sustainable business operations, frameworks like SustAI-SCM can pave the way for a new era of AI-driven supply chain management. By merging intelligent automation with environmental responsibility, organizations can not only achieve operational excellence but also contribute to global sustainability efforts. The integration of AI in supply chain management should not be solely about efficiency gains—it must also be about ensuring a sustainable future where technology-driven optimizations actively reduce the environmental footprint of global logistics networks.