A Hybrid Agentic AI Framework for Intelligent Supply Chain Analytics

1 INTRODUCTION

• Supply-chain analytics spans database querying, KPI computation, performance analysis, forecasting, and optimization, drawing on expertise fragmented across several organizational roles.Company-specific database schemas and industry-specific KPIs add to the range of required expertise.Source: §1 · p. 1
• Planners need to combine ad hoc analysis with standardized workflows, but existing systems often trade flexible exploration against reliable, auditable execution.The paper identifies a unified natural-language system with domain-safe execution as largely absent in operational supply-chain settings.Source: §1 · pp. 1–2
• Prompt-centric encoding of analytical logic and routing is presented as a way to support rapid iteration and lower adoption barriers as supply-chain capabilities evolve.The authors contrast agent instructions with traditional systems that hard-code logic in application code.Source: §1 · p. 2
• The framework combines domain-specific deterministic workflows with open-ended agentic analysis, supporting structured supply-chain processes and flexible exploration in one system.It is designed to bridge natural-language interaction with deterministic decision workflows.Source: §1 · p. 2
• A coordinator delegates distinct analytical responsibilities to specialized agents, a design presented as improving scalability, auditability, and context usage over single-agent baselines.The introduction contrasts this modular design with monolithic agents that face context pressure and limited interpretability.Source: §1 · p. 2
• The paper evaluates single-agent and multi-agent systems and demonstrates the architecture through a supply-chain use case in a proprietary industry simulator.The simulator acts as a supply-chain digital twin.Source: §1 · p. 2

2.1 LLMs with Tool Use and Agentic Systems

• Tool-augmented language models support database querying and multi-step workflows, while monolithic agents can face scalability and context limitations.Prior work also includes centralized multi-agent coordination and applications in automation, discovery, and forecasting.Source: §2.1 · p. 2

2.2 Intelligent Supply Chain Management

• Supply-chain analytics draws on classical inventory, service-level, and forecasting methods alongside newer machine-learning approaches and digital-supply-chain analytics.The cited examples include stochastic inventory models, ARIMA, and data-driven forecasting and optimization.Source: §2.2 · p. 2

2.3 AI and LLMs for Supply Chain Analytics

• Existing AI and LLM supply-chain work addresses specific domains or high-level paradigms, whereas this paper proposes their integration in a controlled multi-agent framework.The proposed combination includes natural-language interaction, schema-grounded retrieval, deterministic workflows, and exploratory reasoning.Source: §2.3 · p. 3

3 Agentic System Architecture

• The system design targets usability, extensibility, and scalability as supply-chain data, tools, functions, and organizational workloads change.Usability is framed around planner access without technical expertise; extensibility around adding capabilities; and scalability around managing larger workloads and costs.Source: §3 · p. 3

3.1 Hub-and-Spoke Multi-Agent Architecture

• A coordinator interprets natural-language requests, decomposes tasks, and delegates to open-ended or workflow agents; specialist communication passes through the coordinator and explicit data artifacts.The architecture uses artifact references and metadata to support reuse of intermediate results and an audit trail.Source: §3.1 · pp. 3–4, Figure 1 · p. 3
• Workflow agents implement predefined, auditable procedures, and the coordinator routes requests between these procedures and flexible hypothesis-driven analysis.The design places deterministic, domain-safe procedures and exploratory agents in one framework.Source: §3.1 · p. 4
• Open-ended specialists support retrieval, visualization, and hypothesis generation, while KPI, sub-optimality-detection, and forecasting agents provide supply-chain-specific capabilities.The listed workflows include skyline analysis of DIO and fill-rate trade-offs and selection of a forecasting method by error.Source: §3.1 · p. 4

4.1 Experimental Testbed

• The testbed models one supplier, one warehouse, ten retail stores, and five item types, with randomly generated demand and simulator-produced data stored in a relational database.The database includes operational attributes such as inventory, forecasts, prices, lead times, and deliveries.Source: §4.1 · p. 5

4.2 Experiments and Tasks

• The evaluation uses 100 manually reviewed queries expanded from 20 domain-expert queries, comparing reasoning and non-reasoning single- and multi-agent configurations over five runs.Queries cover analysis, data retrieval, visualization, and metric calculation; the single-agent prompt consolidates specialist prompts.Source: §4.2 · p. 5

4.3 Evaluation Protocol

• An LLM judge scores intent coverage, completeness, and relevance on a three-level scale, and a domain expert assessed 20 queries to validate automated scoring.The expert fully agreed in 17 cases, partially agreed in two, and disagreed in one.Source: §4.3 · pp. 5–6

5.1 Experimental Results

• Reasoning models score above 90%, while non-reasoning models score approximately 65%.The reported scores are means averaged across five runs.Source: §5.1 · p. 6, Table 2 · p. 6
• Reasoning models show a slight performance decrease with multiple agents, whereas non-reasoning models improve slightly with lower variance.The authors hypothesize that global prompt visibility benefits stronger reasoning models, while explicit task division can help non-reasoning models.Source: §5.1 · p. 6
• The single-agent architecture consumes roughly four times more input tokens, while multi-agent communication can increase response latency.The single-agent prompt includes all system roles and workflows; the multi-agent design activates only task-relevant prompts.Source: §5.1 · p. 6

5.2 Analysis of failures

• The largest reasoning versus non-reasoning performance gaps occur in data retrieval and visualization; structured metric calculation shows a smaller gap.Reasoning models concentrate successful responses, while non-reasoning models show a more even distribution between successful and partially correct responses.Source: §5.2 · pp. 6–7, Table 4 · p. 7, Table 3 · p. 7
• Non-reasoning models often fail in data retrieval because of execution errors such as date formatting, missing tables, or misconfigured data connections.Because visualization and downstream analysis depend on retrieval, these errors can propagate into compounded failures.Source: §5.2 · p. 7, Table 5 · p. 7

5.3 Case study: Hybrid Workflows

• The sub-optimality workflow screens stores against a fulfillment target, uses skyline analysis on selected KPIs, then applies comparative or retrospective diagnosis.It returns suboptimal stores with symptoms and likely root causes.Source: §5.3 · p. 8, Figure 2 · p. 8
• The coordinator retrieves data from planner queries and triggers deterministic analysis, producing diagnostic evidence intended to support traceability and auditability.The paper says embedding the process in the agentic architecture reduces reliance on data analysts and accelerates data-driven decision-making.Source: §5.3 · p. 8

• The cited passage lists store_01, store_02, store_03, store_04, store_05, store_06, store_07, store_08, store_09, and store_10, without identifying them as meeting a target fill rate.Source: §1 · p. 9

• Store_01 and store_08 each have fill rate 1.00 and DIO 0.Source: §3 · p. 9

• Stores 03, 05, 07, 09, and 10 have fill rate 0.99 and DIO 0; store_02 has 0.98 and DIO 0; stores 04 and 06 have 0.99 and DIO 3.Source: §4 · p. 9

• The section references a saved figure file named skyline.png.Source: §5 · p. 9

6. Single-Store Diagnostics (Dominated Stores)

• MAPE ranges from 0.05–0.07; the primary symptom is lower Fill Rate or higher DIO, with no definitive root cause identified within predefined indicators.Source: §6 · p. 9

7. Multi-Store Comparative Analysis

• Multi-store contributing factors include poor forecast accuracy (forecast accuracy ratio > 1.1), excessive safety stock (ratio up to 6.0), and demand variability differences.Source: §7 · p. 9
• Forecast re-optimization is recommended for stores with poor forecast accuracy or excessive safety stock, including store_02, store_05, store_06, store_09, and store_10.Source: §7 · p. 9
• The forecast optimization agent evaluates Gaussian Process regression, Exponential Smoothing, and ARIMA before selecting, fitting, and deploying a model.Source: §7 · p. 9
• The optimized forecast tracks demand more closely, with sustained positive inventory and increased shipments mitigating baseline stock-outs.The comparison uses a supply-chain digital twin.Source: §7 · p. 9

• The architecture is competitive with a single-agent baseline while using roughly 4X fewer tokens.Source: §6 · p. 9
• The multi-agent system balances flexibility with deterministic control for natural-language supply-chain analytics.Source: §6 · p. 9
• Future work includes additional workflows and tools, improved multi-agent performance through prompt strategies, and long-context management.Source: §6 · p. 9
• Figure 3 presents the interface and what-if simulation results comparing baseline and optimized forecasts for sub-optimality detection and demand re-forecasting.Source: Figure 3 · p. 10