Modern supply chains operate within a highly dynamic and complex global environment. The optimization of on-time delivery is a critical objective across diverse operational contexts, including Full Truckload (FTL), Less-Than-Truckload (LTL), parcel, air, and intermodal freight. This challenge spans numerous industry verticals, from retail and manufacturing to specialized sectors like pharmaceuticals, which demand stringent cold chain logistics. To manage this complexity, organizations are increasingly turning to Artificial Intelligence (AI) to automate and enhance decision-making, particularly in the crucial process of carrier selection. AI models can analyze vast datasets to predict carrier performance and recommend optimal choices, promising significant gains in efficiency and reliability.
Despite their potential, the adoption of sophisticated AI models is often hampered by their “black box” nature. The complex, opaque algorithms of deep learning and ensemble models make it difficult for stakeholders to understand how they arrive at specific conclusions. This lack of transparency fosters a trust deficit and presents a significant barrier to adoption, particularly in high-stakes applications.1,2 The problem is compounded by an evolving regulatory landscape, exemplified by the European Union’s AI Act, which classifies AI-driven transport automation as a high-risk domain requiring rigorous explainability and robustness.3
This paper addresses the critical gap between the predictive power of AI and the practical need for transparency in logistics. It aims to demonstrate how eXplainable AI (XAI) can be leveraged to interpret black-box models used for on-time delivery optimization. The primary objectives are: to review the state-of-the-art AI models applied in logistics; to analyze the function and utility of XAI methods in providing decision transparency; to frame how these explanations can be tailored to the distinct needs of technical, business, and operational users; and to discuss the inherent limitations of current XAI techniques, most notably the challenge of distinguishing correlation from causation.
The application of quantitative methods to logistics optimization has a long history, with early work focusing on mathematical programming and heuristics to solve problems like the Vehicle Routing Problem (VRP) and design distribution networks.4 While foundational, these methods often lack the capacity to handle the immense scale and complexity of contemporary logistics data. The field has since evolved toward data-driven AI and machine learning (ML) models that offer superior predictive performance.
Recent studies consistently identify ensemble learning methods, specifically XGBoost and LightGBM, as top-performing models for logistics-related prediction tasks, such as Travel Time Prediction (TTP).5 For more complex optimization and selection problems, such as choosing a healthcare supply chain mode, deep reinforcement learning (DRL) algorithms like the Deep Deterministic Policy Gradient (DDPG) have demonstrated a significant advantage over traditional methods by effectively balancing economic, social, and environmental objectives.6 These advanced models, however, are characteristically opaque, forming the basis of the “black box” problem that impedes trust and adoption in critical sectors.1,2
To address this challenge, the field of eXplainable AI (XAI) has emerged, offering post-hoc techniques to interpret model behavior. Methods such as Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) are widely used to generate plausible and intuitive explanations by attributing a model’s output to its input features.5 For instance, in TTP models, these techniques have confirmed that temporal features like departure hour and spatial features like travel distance are the most significant predictors.5 More advanced frameworks like the Explainable Spatio-Temporal Inference Network (eX-STIN) integrate SHAP directly into the modeling pipeline to provide transparency for demand predictions.7
A crucial limitation of many XAI applications, however, is their reliance on correlation rather than causation.3,8 Standard XAI tools can struggle to differentiate spurious correlations from genuine causal mechanisms, which limits the reliability of their explanations for informing real-world interventions. Recognizing this gap, recent research has focused on integrating causal discovery into ML frameworks. Methods like the Improved Iterative Causal Discovery (IICD) algorithm and ReX, which uses SHAP to help infer causal graphs, aim to build models on a foundation of causal relationships rather than mere statistical patterns.9,10 Such causal-based models have been shown to achieve significant improvements in prediction accuracy, underscoring the value of moving beyond correlation.11
This paper employs a conceptual analysis and synthesis of existing literature to construct a framework for applying XAI to on-time delivery optimization and carrier selection. The methodology is grounded in a review of peer-reviewed studies, technical reports, and industry analyses focused on AI, XAI, and logistics management. The analytical approach is centered on interpreting and integrating findings to address the practical challenges of implementing explainable AI in a complex operational domain.
The core of the analysis is structured around a role-based framework that considers the distinct explanatory needs of different user personas within a logistics organization. This framework categorizes audiences into three primary groups:
The scope of this analysis covers a wide range of operational contexts as defined by freight type (FTL, LTL, parcel, air, intermodal), industry vertical (retail, manufacturing, pharmaceuticals, consumer packaged goods), and geographic scale (domestic, cross-border, international). The study focuses on the application of post-hoc XAI techniques to existing high-performance “black-box” models, such as deep neural networks and gradient boosting machines, as this reflects the predominant approach in the current state of practice.
The synthesis of existing research reveals that XAI can effectively bridge the gap between the predictive power of AI models and the need for human-centric understanding in logistics. The findings are organized around the identification of key predictive factors, the tailoring of explanations for different stakeholders, and the performance of explainable models.
XAI techniques like SHAP and LIME are instrumental in uncovering the primary drivers behind model predictions. In studies on travel time prediction, a direct proxy for on-time delivery, XAI methods consistently highlight the importance of spatio-temporal features. Factors such as the departure hour, day of the week, travel distance, and origin-destination pair are identified as the most significant variables influencing predictions.5 In more specialized contexts, such as healthcare supply chain optimization, the key factors for mode selection extend beyond simple logistics metrics to include a balanced scorecard of economic indicators (e.g., inventory turnover rate), social benefits (e.g., pharmaceutical qualification rate), and environmental impacts (e.g., total waste emissions).6 By making these drivers transparent, XAI allows organizations to understand what criteria the AI is prioritizing in its carrier selection logic.
A critical finding is that the utility of an explanation depends on its audience. For data scientists, XAI provides detailed, technical outputs. For example, SHAP interaction values can be computed to reveal not only the individual contribution of each feature but also the synergistic or antagonistic effects between pairs of features, offering a deeper level of model diagnostics.13 For logistics managers, these technical details must be aggregated and translated into strategic insights. The value lies in high-level summaries, such as those presented in role-based dashboards, that connect model outputs to business outcomes (e.g., “Carrier A’s on-time performance score is declining due to repeated delays on the Chicago-to-Dallas lane”).12 For operational users like dispatchers, explanations must be immediate and actionable to build trust and justify real-time decisions. A simple message like, “Assign to Carrier B: Highest on-time probability for this lane and available capacity,” provides the necessary confidence to act on an AI recommendation.1
The adoption of AI is justified by its performance, and studies show that the models amenable to explanation are highly effective. Ensemble models like XGBoost and LightGBM not only provide top-tier accuracy in prediction tasks5 but are also computationally more efficient than more complex hybrid deep learning models.5 In optimization tasks, DRL algorithms have been shown to outperform conventional methods by finding solutions that more closely align with predefined optimal targets.6 Furthermore, integrated AI systems, such as the AI Strategy Agent for Airline Logistics (AISAL), have demonstrated tangible business value, achieving a 22% improvement in forecast accuracy and an 11% reduction in cost per available seat kilometer, while embedding explainability as a core component.14
The findings confirm that XAI is not merely a technical feature but a critical enabler of AI adoption in logistics. By translating opaque model logic into human-understandable insights, XAI directly addresses the pervasive challenge of trust and transparency. The ability to tailor explanations to different roles—from a granular SHAP plot for a data scientist to a concise justification for a dispatcher—is paramount for integrating AI into the complex socio-technical fabric of a logistics organization.
The practical implications are significant. Logistics companies can use XAI-driven insights to move beyond simple carrier selection and engage in more strategic carrier relationship management, providing partners with specific, data-backed feedback for performance improvement. Moreover, the implementation of XAI is becoming a matter of regulatory necessity. The EU AI Act’s classification of transport AI as “high-risk” mandates a level of transparency and auditability that XAI is uniquely positioned to provide.3 As one report notes, this has already compelled companies to undertake significant engineering efforts, such as adding extensive code for “decision audit” modules, to comply with explainability requirements.15
Despite its benefits, current XAI practice has a fundamental limitation: it primarily exposes correlations, not causation.8 An AI model might learn that a carrier is frequently late on rainy days, and an XAI tool will highlight “weather” as an important feature. However, it cannot distinguish whether the rain causes the delay (a causal link) or if both are correlated with a third, unobserved factor like increased city-wide traffic (a spurious correlation). This distinction is critical, as interventions based on correlational insights may be ineffective or even counterproductive.3 This causality gap remains the most significant challenge for the robust and reliable application of AI in system dynamics. Additional limitations include the computational overhead of running XAI analyses on top of already complex models5 and the substantial implementation effort required to integrate meaningful explanation capabilities into existing platforms.15
Addressing the causality gap is the next frontier for research in this domain. Future work should focus on integrating causal discovery algorithms, such as IICD or ReX, into logistics optimization frameworks to build models that learn from causal relationships.9,10 This would enhance the robustness of predictions and provide a more reliable basis for intervention. Another promising avenue is the development of inherently interpretable models that do not require post-hoc explanation, such as the proposed Interpretable Neural System Dynamics (INSD) pipeline, which aims to combine the performance of deep learning with the transparency of system dynamics equations.3 Further research is also needed to standardize frameworks for generating and validating role-specific explanations to ensure they are not only technically sound but also genuinely useful and trustworthy for end-users.
The optimization of on-time delivery through intelligent carrier selection is a cornerstone of modern logistics management. While powerful AI models offer unprecedented predictive capabilities, their “black box” nature has been a persistent obstacle to widespread adoption. This paper has synthesized evidence demonstrating that eXplainable AI provides a viable path toward resolving this challenge by rendering model decision-making transparent, interpretable, and trustworthy.
The key contribution of XAI lies in its ability to identify the most influential factors in a prediction and, crucially, to tailor these explanations to the specific needs of data scientists, logistics managers, and operational staff. This role-based approach fosters confidence and empowers personnel at all levels to leverage AI as a collaborative tool. However, practitioners must remain cognizant of the limitations of current techniques, particularly the challenge of inferring causality from correlation. The future of AI in logistics will be defined not only by its predictive accuracy but by its interpretability. Moving from correlational patterns to causal understanding represents the next critical step in building truly intelligent and reliable supply chain systems.
