Supply-demand mismatches constitute a major source of overproduction, inventory redundancy, and reverse logistics for returns, generating unnecessary emissions in procurement, manufacturing, and logistics ([1]). Traditional forecasting relies heavily on linear extrapolation of historical sales data and is poorly equipped to perceive abrupt shifts in consumer demand. By integrating multi-source data-such as historical sales records, social media sentiment, meteorological information, and macroeconomic indicators-AI constructs demand sensing models that transmit front-end consumption signals backward to manufacturing and procurement. This enables enterprises to implement demand-driven purchasing, demand-driven production scheduling, and dynamic regional inventory allocation, thereby reducing uncertainty and conserving resource inputs at the source ([6]). Taking JD Logistics’ effective case practice as an example, its self-developed MRV-T (Monitoring, Reporting, Verification, and Tracking) system links demand signals arising from order fluctuations with various carbon emission factor datasets, uses AI models to analyze energy consumption in real time, and dynamically adjusts capacity allocation, thereby promoting end-to-end supply-demand matching.

Logistics and transportation represent the most visible and proportionally significant carbon-emitting segment of supply chains. Traditional scheduling often takes the shortest distance or the lowest cost as a single objective, making it difficult to incorporate energy and carbon constraints ([2]). By introducing real-time traffic, weather, time-of-use electricity tariffs, and other multi-dimensional variables, and by simulating warehouse node location choices and multimodal transport combinations, AI can support more balanced decision making among transportation cost, delivery time, energy consumption, and carbon emissions. As an illustration, Schneider Electric decomposes its emission reduction targets into a set of executable measures and employs AI models for intelligent scheduling, optimizing substantial reductions in both transportation costs and energy-and carbon-related consumption.

A prerequisite for the green and low-carbon transition of supply chains is the traceability of emission sources. Traditional accounting methods rely on industry-average emission factors and cannot support precise, batch-by-batch or node-by-node decision-making. This is especially true for Scope 3 emissions, which account for the overwhelming majority of a firm’s carbon footprint. Because supplier tiers are numerous, data remain persistently opaque, rendering most emissions effectively untraceable ([9]). AI is beginning to break this deadlock. By leveraging technologies such as natural language processing and computer vision, it can automatically extract and parse energy consumption and carbon emission data from unstructured information such as suppliers’ environmental compliance reports and transportation energy records, thereby helping to dismantle information barriers inside and outside enterprise boundaries and laying a foundation for accurately defining carbon footprint management responsibilities.

The green transition of a supply chain cannot be accomplished by a single focal firm alone; its effectiveness depends to a large degree on the collective response of the supplier network. The critical challenges lie in effectively identifying green performers among a vast number of suppliers and in incentivizing and empowering numerous small and medium-sized enterprises to invest in energy saving and carbon reduction ([4]). Using industrial knowledge graphs and multi-dimensional machine learning, AI can construct supplier evaluation models that integrate indicators such as energy intensity, carbon footprint management maturity, and the share of renewable energy, thereby giving green preference in procurement sourcing and tendering. Taking Alibaba Cloud’s “Energy Expert” (Nenghaobao) as an illustrative case of an AI-based sustainability tool, this tool can automatically calculate carbon emissions for organizations, including small and medium-sized suppliers, and provide optimization recommendations, thereby supporting the green and low-carbon coordinated development of the supply chain network.

The existing pathways are highly dependent on the supply of high-quality, standardized data, which is precisely the scarcest resource in supply chains. Current levels of enterprise digitalization are uneven, and inconsistencies in data formats, accounting scopes, and boundary definitions create intrinsic deficiencies in model training data. Scope 3 carbon emissions involve multi-tier suppliers; the further upstream one goes, the lower the availability and reliability of data become ([8]). A large number of small and medium-sized enterprises have yet to adopt automated data management systems, thereby constraining the analytical precision of models.

The application of AI in carbon management is transitioning from descriptive analysis to prescriptive decision-making, yet the internal reasoning processes of some deep learning models remain highly opaque to managers. When models output green, energy-saving, and carbon-reduction recommendations in a “black-box” manner, decision-makers tend to be cautious. This contradiction is particularly pronounced in the carbon emissions domain: carbon data are already subject to considerable uncertainty, and when this is compounded by the uninterpretability of algorithmic reasoning, managers’ willingness to adopt model outputs is doubly suppressed.

The green transition of supply chains faces a classic collective action dilemma: investments in energy saving and carbon reduction are concentrated in focal lead firms or first-tier suppliers, whereas the benefits are dispersed across the entire supply chain and society at large ([4]). The introduction of AI not only fails to automatically resolve this dilemma but may actually exacerbate incentive misalignments. Focal lead firms possess advantages in capital, data, and talent for deploying energy and carbon management systems, while the vast majority of small and medium-sized suppliers lack both technical capability and the prospect of recouping the costs of green retrofitting within short-term commercial returns. In fragmented supply chain structures that lack a dominant focal firm, the transaction costs of collaborative emission reduction will continue to rise.

AI technology itself is not carbon-neutral. Large language models consume substantial computing power during both training and inference phases; embedding them into supply chain management systems inevitably incurs additional energy consumption. Locally achieved emission reductions driven by AI may, at the system level, be partially or even fully offset by the technology’s own carbon footprint, producing a rebound effect. If this additional emission burden is not incorporated into the overall carbon accounting framework of the supply chain, the net environmental contribution of AI-driven carbon reduction cannot be accurately measured. The high energy consumption of certain models, together with limited data interoperability and insufficient scalability, constitute major obstacles to the current technological pathways. In the absence of a full-life-cycle carbon performance evaluation mechanism, the effectiveness of the AI-driven green and low-carbon transition of supply chains will be exposed to rebound risks.

The key to overcoming data fragmentation lies in institutional arrangements. Promoting the standardization of carbon emission data enables AI models to support data interoperability across different supply chain contexts, thereby laying a foundation for large-scale application. By adopting privacy-preserving computation techniques such as secure multi-party computation, data ownership can be empowered and collaborative modeling can be realized, protecting the commercial secrets of all parties while guaranteeing coordinated green development for energy conservation and carbon reduction. Simultaneously, cloud-based deployment and lightweight tools can lower the access threshold for small and medium-sized enterprises, preventing technological barriers from creating a new green digital divide.

To address the algorithmic credibility dilemma, explainability should be established as a rigid constraint in model design. In terms of methodological choices, structurally transparent approaches such as decision trees, rule learning, and explainable boosting tree models should be prioritized for identifying key carbon emission factors and generating strategies. For scenarios that necessitate high-precision complex models, post-hoc explanation techniques should be supplemented, so that every emission-reduction recommendation is accompanied by a traceable reasoning rationale. By deeply coupling the algorithmic output logic with existing supply chain management processes such as audit trails, exception reviews, and compliance approvals each emission-reduction recommendation carries an interpretable business rule explanation. Only then can AI be transformed from an auxiliary technical tool into a trusted decision-making partner for managers, thereby genuinely mitigating the negative resistance and skepticism arising from “black-box” decisions.

Resolving incentive misalignments requires internalizing the externalities of energy saving, emission reduction, and carbon mitigation through institutional design and intelligent execution. Focal lead firms can leverage the real-time performance tracking capabilities of AI to offer tiered incentive schemes such as preferential procurement quotas, priority purchasing, or favorable payment terms to suppliers that consistently improve their energy and carbon performance, thereby directly converting emission-reduction behavior into tangible economic returns. In the future, smart contracts can be introduced to upload verified emission-reduction performance from each node onto a blockchain for immutable record-keeping, automatically executing carbon credit transfers or financial compensation according to pre-defined rules. This decentralized automatic settlement can substantially reduce the transaction costs of multi-party verification and repeated negotiation, while also rendering the relationship between emission-reduction contributions and corresponding rewards transparent and clear, thus effectively curbing free-riding behavior. In this way, the efficiency gains generated by technology and the equity pursued by institutions reinforce each other, forming a synergistic development pattern that is complementary rather than conflicting.

To guard against rebound risks, the carbon footprint of the AI system itself must be incorporated within the accounting boundary. During the system planning stage, energy and carbon consumption evaluation indicators should be introduced; lightweight models and green data centers should be prioritized to contain the technology’s carbon cost at the source. A scientifically rigorous accounting methodology should be constructed, deducting the carbon cost of model training and inference from the driven emission reductions and using the net contribution as the ultimate yardstick of effectiveness. Taking a full-life-cycle perspective to calculate the net contribution of the baseline prevents local optimizations from imposing new environmental costs at the system boundary, thereby ensuring that AI technology achieves a net reduction in energy and carbon, rather than engaging in a covert transfer of carbon burden between computing resources and the supply chain.