AI Supply Chain

• AI Supply Chain is a multi-layered concept that encompasses both AI-driven operational planning and a dependency network of AI components.
• It employs graph models, machine learning, and reinforcement learning to optimize forecasting, routing, and disruption management with measurable results.
• The field also tackles governance, compliance, and sustainability by ensuring data integrity and transparent oversight across interconnected stakeholders.

AI supply chain denotes, in recent research, two related phenomena. In one sense, it refers to the use of artificial intelligence across supply-chain planning, execution, monitoring, and sustainability functions such as demand forecasting, routing, warehousing, procurement, document processing, and disruption response. In the other, it denotes the upstream–downstream network of data, models, programs, services, infrastructure, and labor through which AI systems themselves are produced, adapted, and delivered. The literature therefore treats AI supply chain simultaneously as an operational decision system and as a dependency graph whose structure shapes provenance, explainability, compliance, resilience, and market power (Hasan, 2024, Hopkins et al., 28 Apr 2025, Cesarano et al., 30 Apr 2026).

1. Conceptual scope and formal representations

Research on AI supply chains models modern AI production as a specialized, inter-organizational ecosystem rather than a vertically integrated pipeline. One influential formulation represents an AI supply chain as a directed graph $G=(V,E)$ in which each node $v \in V$ denotes an AI component, such as a model or dataset, and each directed edge $(v_j,v_i)\in E$ denotes a dependency from an upstream component to a downstream one. Each node is associated with an operation $h_v$ that maps parent outputs into the component, and graph properties such as path length, width, reachability, cycles, and centrality are linked to transparency, resilience, information flow, and feedback risk (Hopkins et al., 28 Apr 2025).

Historically, this graph structure emerged from a shift from end-to-end, siloed ML development toward specialization and outsourcing. Public datasets such as WordNet, MNIST, and ImageNet reduced datawork costs; transfer learning and fine-tuning reduced the need to train models from scratch; and, since 2022, general-purpose foundation models such as GPT, Llama, Claude, and Gemini lowered entry barriers and stabilized a common “recipe” of large upstream datasets, base-model training, and downstream fine-tuning into domain-specific tools. Around this pipeline, an ecosystem of base-model providers, dataset curators and labelers, hosting and inference platforms, fine-tuning services, and application integrators has emerged (Hopkins et al., 28 Apr 2025).

A complementary taxonomy decomposes the AI supply chain into four components: Data, Models, Programs, and Infrastructure. Data includes creators, hosts, aggregators, and users; Models include creators, hosts, and users; Programs include users, hosts, integrators, and developers; Infrastructure includes users, hosts, integrators, and developers. This taxonomy deliberately includes runtime influences such as retrieval corpora, agents, serving systems, and storage platforms, because all of them can contribute to the final output of the AI system (Sheh et al., 19 Nov 2025).

The literature also identifies a specifically labor-centered AI supply chain. In this formulation, the AI supply chain is the organizational and geographic system through which human data work is procured, coordinated, and delivered. Empirical work across France, Madagascar, and Venezuela distinguishes marketplace platforms, firm-like embedded vendors, and intermediate “deep labour” arrangements, showing that AI preparation, verification, and impersonation tasks are allocated differently depending on confidentiality, quality requirements, and desired flexibility (Tubaro et al., 7 Feb 2025).

2. AI as an operational layer in physical supply chains

In supply-chain management proper, AI is positioned as a decision-support and optimization layer spanning forecasting, inventory, transportation, supplier evaluation, warehousing, product lifecycle management, reverse logistics, and carbon reporting. One U.S.-focused framework targets demand forecasting and inventory optimization, transportation route optimization and fleet management, supplier evaluation and risk management, smart warehousing and energy management, product lifecycle management and reverse logistics, and carbon footprint tracking and reporting. Its policy context is shaped by the Environmental Protection Agency’s finding that transportation contributed about 29% of total U.S. GHG emissions in 2023, by the World Economic Forum’s estimate that supply chains drive over 60% of global emissions, and by incentives under the Inflation Reduction Act of 2022 (Hasan, 2024).

A more implementation-oriented logistics study specifies the data and modeling stack in greater detail. It uses datasets from governmental transport agencies and proprietary logistics databases, with features including shipment details, transport metrics, environmental indicators, and operational factors such as weather conditions and traffic congestion. The modeling layer combines Linear Regression, MLPRegressor, RandomForestRegressor, XGBRegressor, ElasticNet, SVR, K-Means, and DBSCAN, and feeds ML predictions into a vehicle routing problem with time windows. In reported results, XGBRegressor achieved MAE = 0.454 and $R^2 = 0.999$ for carbon emissions modeling; XGBRegressor achieved MAE = 1.919 and $R^2 = 0.474$ for travel route optimization predictions; and MLPRegressor achieved MAE = 0.396 and $R^2 = 0.635$ for demand forecasting. K-Means with $K=3$ produced interpretable route segments, and DBSCAN flagged 48 outlier routes for remediation (Shawon et al., 18 Mar 2025).

Domain-specific AI supply-chain applications extend beyond logistics. In food systems, deep learning has been used from farm to fork: encoder–decoder LSTM with attention for greenhouse plant growth and tomato yield prediction, LSTM for refrigeration de-freezing prediction during peak load shedding, and FCN plus CRNN for optical recognition and verification of food expiry dates. Reported results include RMSE values of 0.0026, 0.0028, and 0.0029 for one-, two-, and three-step Ficus growth forecasting with WT–ED–LSTM–AM; 98.2% detection accuracy for expiry-date localization; and 95.44% recognition accuracy for CRNN-based date recognition (Kollia et al., 2021).

The resilience literature connects these operational capabilities to broader supply-chain properties. In an AIoT-based resilient supply-chain model, four dimensions—Triggers, Vulnerabilities, Capabilities, and Empowerment—are prioritized through fuzzy nonlinear decision analysis. At the dimension level, Empowerment receives weight 0.455, Capabilities 0.253, Vulnerabilities 0.180, and Triggers 0.112. At the component level, Transparency (0.1449), Power of prediction (0.1330), Risk management (0.1107), and Flexibility (0.1089) rank highest, which suggests that visibility and prediction are treated as the most leverageable resilience mechanisms in AI-enabled operations (Aliahmadi et al., 2022).

3. Agentic and generative planning systems

Recent work shifts from point prediction and rule-based automation toward multi-stage agentic planning systems. One line of work, “Generative Probabilistic Planning,” formulates each SKU’s supply chain as a directed graph and combines attention-based graph neural networks, offline reinforcement learning, and probabilistic policy simulation. The system trains 12 policies with different risk parameters on 39,000 weekly training snapshots across 500 high-volume SKUs, then evaluates them with Monte Carlo simulations over a 13-week horizon. Against historical policy, reported steady-state results include lost sales reductions of 75% with excess stock reduction of 20% in one setting, and lost sales reduction of 81% with excess stock reduction of 4% in another (Ahn et al., 2024).

A second line uses LLM-based orchestration for enterprise planning workflows. The Supply Chain Planning Agent deployed at JD.com recasts planning as a five-stage loop of data acquisition, plan formulation, plan execution, diagnosis and early warning, and plan correction. Its tools include retrieval-augmented generation over SOPs, text-to-SQL, atomic operation code generation through Filter, Transform, Groupby, and Sort, and function calling. JD.com’s deployment scale is reported as over 10 million self-operated SKUs, thousands of warehouses and distribution centers, tens of millions of daily orders, and 600 million annual active users. Operational outcomes include a ~40% reduction in weekly data processing and analysis time, a 22% increase in the proportion of plans with accuracy deviation below 5%, and a ~2–3% improvement in stock fulfillment rate (Qi et al., 4 Sep 2025).

Retail replenishment has also been re-architected as a supervised agent network. Flowr decomposes supermarket operations into Demand Forecasting, Inventory Monitoring, Procurement and Ordering, Supplier Coordination, DC Replenishment Planning, and Exception and Alert agents, coordinated by a central reasoning LLM and exposed through a Model Context Protocol interface for human approval. In reported evaluation, the Procurement and Ordering Agent received 4.7/5 for correctness, completeness, and operational usability, while DC Replenishment Planning achieved a 16% average route optimization efficiency gain versus the manual baseline and received 4.6/5 for operational accuracy and interpretability (Bandara et al., 7 Apr 2026).

Agentic systems have also been applied to disruption monitoring across deep-tier networks. A seven-agent framework combines GPT-4o with deterministic tools such as Neo4j graph queries and risk calculators to detect disruption signals from news, traverse supply chains up to Tier-4, evaluate Tier-1 exposure, and recommend mitigations such as alternative sourcing. Across 30 synthesized scenarios, the system reports F1 scores between 0.962 and 0.991, mean end-to-end analysis time of 3.83 minutes, and mean cost of \$0.0836 per disruption (AlMahri et al., 14 Jan 2026).

Agentic coordination is not uniformly stabilizing. In a three-echelon simulated supply chain, a VMI-style “collaborative” LLM agent underperformed non-AI baselines because inventory was hoarded at the Manufacturer while the Retailer starved; early VMI implementations produced service levels often below 5%. A corrected two-layer design—high-level AI policy-setting plus low-level collaborative execution with proactive downstream replenishment—restored stability, and in a transportation disruption experiment all generated strategies achieved 100% service level, with costs ranging from \$15,363.00 to \$19,767.50 depending on the speed–cost trade-off (Dhar, 19 Aug 2025).

4. Visibility, data integrity, and supply-chain risk intelligence

A major theme in the literature is that supply-chain risk cannot be managed without better visibility beyond direct counterparties and better integrity in upstream data pipelines. In supply-network surveillance, one approach augments link prediction with generative AI embeddings over context-rich “quintuplets” such as $(h,\text{supplies},x,\text{to},t)$ and $v \in V$0. Using an automotive knowledge graph with 43,131 companies across 79 countries, 927 products, and 5 certification types, the GenAI-enhanced pipeline reports near-perfect balanced accuracy for many country-level tasks; for example, for $v \in V$1, Canada LogReg/LSTM/CNN/AutoEncoder all achieved $v \in V$2 with all-MiniLM-L12-v2 embeddings (Zheng et al., 2024).

Upstream data integrity is itself treated as part of the AI supply chain. A survey-integrity framework for readiness assessments in supply-chain AI adoption combines rule-based contradiction checks, a basic NLP layer using BERT embeddings and cosine similarity, and supervised classifiers. On 99 responses, with 14 manually labeled as fake, Random Forest achieved Accuracy 0.92, fake-class Precision 1.00, Recall 0.50, and F1 0.67; the reported test-set confusion matrix had 21 true negatives, 2 true positives, 2 false negatives, and 0 false positives. The framework is positioned as a “data integrity layer” upstream of applications such as safety stock optimization and deployment planning (Mani, 14 Jan 2026).

The broader supply-chain risk-assessment literature corroborates the centrality of AI-driven sensing and classification. A systematic review and bibliometric analysis reports strong use of Random Forest, XGBoost, hybrid ensembles, deep sequence models, graph-based models, and federated learning across credit risk, disruption detection, demand and commodity risk, shipment feasibility, and hidden dependency discovery. Quantitative results summarized in that review include CSL-RF with Accuracy ≈ 97.22% and AUC ≈ 0.985, CNN early warning with Accuracy ≈ 94.7%, and Bi-LSTM–CRF improving disruption-event F1 from ≈ 80% to ≈ 85% (Jahin et al., 2023).

Taken together, these results suggest that “AI supply chain” increasingly includes not only downstream optimization, but also the upstream evidentiary substrate—knowledge graphs, surveys, unstructured news, sensor streams, and integrity filters—that conditions whether subsequent decisions are trustworthy.

5. Governance, compliance, and political economy

When AI systems themselves are treated as supply chains, governance questions become structural. One software-centered analysis decomposes the AI software supply chain into four architectural layers—data acquisition, model training, integration and inference, and a cross-cutting substrate—and identifies four integrity gaps: verifiability, versioning, observability, and traceability. Its reference stack of 48 production-grade open-source projects declares 4,664 direct dependencies, resolves to 11,508 transitive packages, and totals roughly 392M lines of code, making explicit the scale of the trusted computing base behind contemporary AI systems (Cesarano et al., 30 Apr 2026).

Compliance work extends this structural view to model provenance and licensing. The AI Supply Chain Galaxy system constructs a directed dependency graph over 908,449 Hugging Face models and applies a rule-based compliance engine to explicit reuse edges such as fine-tuning, adapters, quantization, and merging. The reported ecosystem-wide result is that 55.46% of models exhibit compliance risks or metadata conflicts or omissions. Operation-specific patterns are pronounced: adapter derivations show a 56.67% license omission rate in children, fine-tuning shows an 8.05% license drift rate, and merge edges exhibit an 86.53% overall conflict rate (Han et al., 15 Jun 2026).

Stakeholder analysis introduces a further governance vocabulary centered on harm and remedy. One study organizes AI supply-chain stakeholders into infrastructure providers, data providers, model providers, intermediaries and AI services, user-facing providers, and users or consumers, then proposes four response types to AISC-induced harms: recourse, repair, reparation, and prevention. Its healthcare case study shows that the feasibility of each response varies across vertical integration, horizontal integration, and free-market structures, because traceability, bargaining power, and reversibility are not evenly distributed across the chain (Hopkins et al., 3 Jul 2025).

Economic regulation adds a formal market model. In a one-provider, two-downstream-firm game, downstream application quality is $v \in V$3, where $v \in V$4 is the downstream firm’s preprocessed proprietary data volume. The paper concludes that pro-price-competitive policies increase consumer surplus only when compute or data preprocessing costs are high, compute subsidies are effective only when these costs are low, and pro-quality-competitive policies always improve consumer surplus. It also reports that pro-price-competitive policies and compute subsidies can generate joint gains for the provider, downstream firms, and consumers in some regimes, whereas pro-quality-competitive policies increase provider profits while reducing downstream profits (Qian et al., 13 Mar 2026).

6. Environmental, documentary, and labor dimensions

The environmental dimension of AI supply chains is now measured not only through routing and inventory but also through the information-processing substrate of supply-chain operations. In document-intensive workflows, a three-scenario sustainability assessment compares manual processing, AI-assisted human-in-the-loop processing, and multi-agent agentic AI. At a reference workload of 5,000 documents per day, manual processing consumes 36.3–194.7 kWh/day, emits 10.5–56.1 kg/day of CO2e, and uses 35.1–58.4 L/day of water; AI-assisted HITL consumes 6.1–16.2 kWh/day; and agentic AI consumes 9.8–20.5 kWh/day. Reported reductions versus manual range from approximately 70–90% for energy, 83–92% or 90–97% for CO2e depending on configuration, and 89–98% for water. In the replicable proforma-invoice use case, the framework reports 100% numerical accuracy (Gosmar et al., 10 Nov 2025).

The social dimension is equally salient. Empirical work on outsourced and offshored data work shows that AI supply chains procure labor through marketplace-like platforms, embedded vendors, and “deep labour” intermediary chains, with uneven consequences for remuneration, job security, and working conditions. France appears as proximate but often disembedded, Madagascar as distant yet embedded, and Venezuela as both distant and disembedded. Concrete pay points reported for Venezuelan platform work include captchas at 0.50–1.50 USD per 1,000 items and selfies at 3–5 USD per set; in some projects, Venezuelan workers were temporarily limited to 10% of tasks (Tubaro et al., 7 Feb 2025).

Across the literature, unresolved directions are recurrent. They include dynamic and temporal supply-chain graphs, implicit lineage inference, cross-organizational attestation, uncertainty quantification, multimodal data fusion, reinforcement learning for real-time routing, dynamic Bayesian networks for disruption propagation, and broader incorporation of compute, service, and infrastructure artifacts into supply-chain-aware auditing (Hopkins et al., 28 Apr 2025, Cesarano et al., 30 Apr 2026, AlMahri et al., 14 Jan 2026).

In aggregate, the research portrays AI supply chain as an inherently multi-layered object. It is at once a set of AI-enabled methods for planning and executing material flows, a provenance and dependency problem for AI artifacts and services, and a governance problem spanning law, labor, sustainability, and market structure. The common analytical move across these strands is to replace isolated models with explicit chains—of goods, documents, signals, software, data, or organizational dependencies—and to treat the performance, risk, and legitimacy of AI systems as functions of those chains rather than of any single model alone.