Skip to content
ScoAI: OR-Native Decision Intelligence for Verifiable Supply Chain Decisions
Industry Insights 20 min read

ScoAI: OR-Native Decision Intelligence for Verifiable Supply Chain Decisions

The article analyzes the ScoAI framework, which integrates supply chain context, operations research, and AI to turn model predictions into verifiable, executable decisions through OR-native interfaces, data-mechanism fusion, decision error decomposition, and agent architectures with multi-layer validation and execution tracking.

JD Tech Talk
JD Tech Talk
JD Tech Talk
Core view: The design and evaluation of reusable learning components should center on the decisions they support, not merely on their outputs.

Introduction

Supply chain inventory management requires decisions on replenishment quantities, inventory allocation, and execution timing under uncertainty in both demand and supply. Foundation models and agents can estimate demand, parse supplier information, and generate candidate plans, but model outputs are not equivalent to business decisions. Only when embedded in explicit business objectives, warehouse capacity and case-pack constraints, transport capacity, and execution processes can they form implementable, verifiable operational plans.

The paper ScoAI: Toward OR-Native Decision Intelligence addresses the gap between model outputs and actual decisions [1]. It proposes ScoAI = SC + OR + AI : supply chain business defines the decision context, resources, and action consequences; operations research translates objectives, constraints, and cross-stage coordination into computable problems; AI provides reusable estimates and judgments across tasks. The three are not independent capabilities but together form a computational loop that runs, validates, and updates around business decisions. As a Perspective paper, it focuses on a methodological question: when foundation models and agents are applied to inventory, allocation, and planning, how to move from generating superficially plausible answers to supporting verifiable, executable, and traceable decisions.

Paper: ScoAI: Toward OR-Native Decision Intelligence

Subtitle: A Perspective on Foundation Models and Agents for Supply Chains

Link: https://www.preprints.org/manuscript/202610.0845

Background

Consider a batch of promotional items at risk of delayed arrival. Even if the model predicts a low delay probability, planners must choose among waiting, expediting, or cross-warehouse transfer. The probability forecast is only one input; stockout loss, expediting cost, other warehouses' available inventory, and transport capacity all affect the final plan. When potential loss is large enough, preemptive intervention may be rational despite low probability. Such problems pervade inventory management: more accurate demand forecasts do not necessarily yield better replenishment quantities; a lower-cost plan for a single warehouse may not reduce total fulfillment cost; approved plan quantities may not be fully executed due to inventory occupation, partial shipments, or manual adjustments. Meanwhile, observed sales during stockouts are usually below true demand, and unmet demand cannot be recovered from sales records alone.

From these phenomena, the paper identifies three key issues when models enter the decision flow: how model judgments translate into concrete actions, how local choices reflect full-chain value, and how plan results connect with actual execution. ScoAI's theoretical analysis, system interfaces, and evaluation framework all stem from these issues.

Screenshot 2026-10-13 00.52.49.png
Screenshot 2026-10-13 00.52.49.png

Core Idea

The paper's main contribution is not a new network architecture but elevating the decision structure to be part of the model interface. In common pipelines, models output predictions while objectives, constraints, and execution rules are handled externally by business systems. ScoAI argues that these elements and their relationships must be explicit from the start of modeling, so they can be independently verified and persist through training, solving, and runtime updates.

This requirement is summarized as OR-native (operations-research-native) . OR-native does not merely mean plugging in a solver; it requires that business objectives, available actions, uncertain factors, information available at decision time, and execution feedback all enter the model or agent interface in explicit form. Then model outputs can directly serve downstream computation, and when business rules change, the system can identify exactly which components need adjustment.

This separation is critical for inventory systems. Inventory conservation is a deterministic business relationship that need not be learned; warehouse capacity, case-pack rules, and service-level requirements are current operational rules that must be explicitly expressed and versioned; demand, lead times, substitution relationships, and supplier responses are uncertain and must be estimated from data. The paper terms the combination of learning components with explicit operational mechanisms data–mechanism fusion .

Data–mechanism fusion is not simply appending an optimizer after a prediction model. The structure of business loss determines whether the prediction stage should focus on mean, quantile, or tail risk; infeasibilities identified by the solver can correct modeling assumptions or candidate plans; actual shipment and receipt records should update the system state for the next decision round. Different feedback acts on different objects and cannot be lumped into generic model error.

2.png
2.png

The paper further decomposes decision error into three categories: high-quality plans missing from the candidate set, the model failing to correctly compare candidates' business value, and insufficient search within limited compute budget. Each error type demands a different improvement path. Improving solver capability only enhances search quality; it cannot add missing candidates or fix erroneous value estimates.

Cross-stage decisions require propagating downstream impacts upstream. Assortment decisions alter subsequent replenishment needs; replenishment plans consume transport and receiving capacity; thus a locally low-cost choice may increase total chain cost. The paper introduces completion value to measure a local plan's potential to form a high-quality complete plan, and uses resource prices to pass the opportunity cost of shared capacity to each stage. For businesses with case-pack requirements, integer variables, and shared resources, the final plan must still pass a joint feasibility check.

System Design

The proposed agent architecture does not require the foundation model to handle the entire decision flow. Models are suited for understanding text, estimating demand and lead times, and proposing candidate plans; deterministic date calculations, inventory conservation, case-pack constraints, warehouse capacity limits, and shared-resource coordination remain the responsibility of programs and solvers; permission checks, resource reservation, and business submission stay with the business system.

Structured judgment models can be integrated via well-bounded interfaces: they read authorized state, judge well-defined problems, and submit option probabilities or rankings to the planner. Model confidence alone does not constitute execution permission. The final plan must still pass validation across five layers: semantic, informational, value, computational, and operational.

3.png
3.png

Execution records are an indispensable part of this architecture. The system should separately store the plan proposed by the model or planner, the approval result, and the actual executed action, and update business state with the actual action rather than the original plan. For inventory management, this affects the accuracy of on-hand and in-transit quantities for the next period and determines which subsequent observations can support analysis and judgment.

4.png
4.png

Supply Chain Scenarios

The paper draws on existing JD research to show that the above interfaces already have technical foundations in various supply chain scenarios. In inventory control, prior work used dynamic-programming-derived decisions as supervision signals to train replenishment policies and adapted them based on actual operating costs; regional and forward warehouse assortment allocation incorporated demand forecasting, shared inventory, warehouse capacity, and fulfillment cost into a single optimization problem; generative planning connected business requests, data, and analytical tools; ORLA involved solvers in generating and selecting optimization models; SCOPE coordinated assortment, sourcing, replenishment cycles, and fulfillment paths across the full chain.

These studies demonstrate multiple ways to connect learning and optimization: OR solutions as training supervision, operating costs as learning signals, solver feedback to correct model outputs, and downstream results to evaluate upstream choices. Note that each work targets different scenarios with different experimental designs; their results cannot be aggregated into an overall ScoAI business benefit. The paper cites them to illustrate how different interface types are realized in concrete tasks.

This methodology also aligns with two recent projects from our team. One recasts the destroy step of large neighborhood search as a learnable structured problem [2]; the other explores how language models can provide prior knowledge and policy structure at the design stage of lost-sales inventory control, while an OR controller handles actual execution [3]. The tasks and technical methods differ, but both embody the principle that AI should provide reusable capabilities in well-defined computational roles, not replace existing constraint expression, optimization solving, and execution control mechanisms.

A constructive example in the paper further shows that business rule changes cannot be handled solely by the model's semantic understanding. When the store shipment limit is adjusted from 4 to 2, the old plan may still receive a high model score but no longer belongs to the new feasible region. Demand estimates and inventory conservation relations remain unchanged; what changed is the allowed action space. The correct response is to update the constraint and re-solve, not to prioritize retraining the prediction model.

5.png
5.png

Case-pack constraints illustrate the same issue. Rounding a per-unit shipment plan directly may simultaneously leave store demand unmet and violate total inventory constraints. The more reliable approach is to write the case-pack rule directly into the joint decision model and recompute. Such basic requirements in warehousing and distribution indicate that explicit decision structures cannot be replaced by free-text generation.

Business Implications

ScoAI's significance lies not in replacing existing business systems with general models, but in clarifying the applicable scope of AI capabilities within the decision chain. Demand and lead times can be estimated from data; inventory conservation and case-pack constraints must be explicitly enforced; supplier text can be parsed by models; whether to expedite must still be decided by weighing cost and resource conditions; candidate plans can be generated by models, but before submission they must pass validation, resource reservation, and permission review. For existing inventory systems, this path is more implementable than directly replacing mature algorithms.

The paper also provides a targeted diagnostic framework. When policy performance degrades, one can separately check whether the candidate set contains high-quality plans, whether candidate value ranking is accurate, and whether the search process was sufficient. Warehouse capacity changes require constraint updates; lead-time distribution changes require response model updates; discrepancies between actual and planned shipment quantities require execution record correction. Many issues vaguely attributed to "model drift" may actually occur at different stages of the decision pipeline. Accordingly, data recording must cover the complete decision process. Saving only inputs and final results is insufficient for accurate attribution; one must also link the rule version used at decision time, candidate plans, system recommendations, approval results, actual execution, and subsequent observations to distinguish the impacts of model, solver, human intervention, and execution deviation. System evaluation should not be limited to prediction accuracy and model invocation cost, but should also assess decision quality, hard constraint satisfaction, task coverage, and end-to-end total cost.

General-purpose models are not suitable for all supply chain tasks. For replenishment scenarios with stable rules and clear problem boundaries, mature specialized OR processes often remain more reliable and economical. Reusable models show advantage only when the same understanding, estimation, or modeling capability can serve multiple related decisions; their value must still be judged by decision effectiveness, task coverage, and overall cost.

Conclusion

ScoAI focuses not on whether AI can enter supply chain, but on what should be learned from data, what relationships must be computed explicitly, and what results need verification against execution records after AI participates in supply chain decisions. The basic division of labor is: supply chain business defines the problem, resources, and action consequences; operations research expresses objectives, constraints, and coordination; AI provides reusable estimates and judgments. What truly improves subsequent decisions is not only the model's initial output but also the actual actions taken and the business feedback they generate.

Paper and Related Work

ScoAI: Toward OR-Native Decision Intelligence – Authors: Yang Liu , Yulin Huang, Xue Yu, Haowei Zhou, Jiaxin Wang, Qiqi Yan, Tianqi Zhou, Yi Yao, Zhongxuan Han, Shaomin Li, Jianshen Zhang, Zhe Chen, Yongzhi Qi*, Wei Qi, Zuo-Jun (Max) Shen.

Recasting the Destroy Step of Large Neighborhood Search as Dense Segmentation – Authors: Yang Liu , Yulin Huang, Jianshen Zhang, Yongzhi Qi*.

SabreAgent: Language Models at Design Time for Lost-Sales Inventory Control – Authors: Yang Liu , Yulin Huang, Xue Yu, Jiong Dong, Jianshen Zhang, Yongzhi Qi*.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Operations Researchagent architecturefoundation modelsData-Mechanism FusionScoAIOR-Native Decision IntelligenceSupply Chain Decision Making
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.