💻 computer science

LLM-Augmented Decision Intelligence for Real-Time Supply Chain Risk Assessment: A Parameterized Simulation and FMEA-Based Framework

This paper presents a novel LLM-augmented decision intelligence framework that integrates FMEA-based risk quantification with large language model contextual reasoning to achieve statistically significant, high-recall real-time supply chain risk assessment, as validated by both parameterized simulations and extensive empirical testing on the DataCo Smart Supply Chain dataset.

Original authors: Nikhil Reddy Donapati

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nikhil Reddy Donapati

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a global supply chain as a massive, high-speed train system. Thousands of trains (ships, trucks, planes) are carrying goods from factories to stores. Usually, the system runs smoothly, but sometimes, things go wrong: a storm hits, a factory breaks down, or a shipment gets lost.

For a long time, companies tried to predict these disasters using old-fashioned rulebooks. These rulebooks were like rigid traffic lights: "If the train is 1 hour late, turn yellow. If it's 2 hours late, turn red." The problem is that the real world is messy. A 1-hour delay might be fine if you have extra cargo, but it could be a catastrophe if you have no backup. Old rulebooks couldn't understand the context or the story behind the numbers.

On the other hand, we now have AI "super-readers" (Large Language Models, or LLMs). These are like brilliant detectives who can read thousands of news reports, emails, and weather logs to understand the story of a delay. But, these detectives are great at storytelling and bad at math. They might say, "This looks risky!" but they can't give you a precise, auditable number to prove it to a boss or a regulator.

This paper proposes a "Dream Team" solution: It combines the math of the old rulebooks with the storytelling of the AI detective.

Here is how the authors built this system, explained simply:

1. The Six-Layer "Kitchen"

Think of the system as a high-tech kitchen preparing a meal (a risk assessment) for a busy restaurant (the supply chain).

Layer 1 (The Delivery): Ingredients (data) like order times, shipping modes, and inventory levels arrive.
Layer 2 (The Prep): The ingredients are chopped and measured (normalized) so they are all in the same format.
Layer 3 (The Math Chef - FMEA): This is the old-school expert. It uses a strict formula (called FMEA) to calculate a "Risk Score." It asks: How bad would this be? How likely is it? Can we spot it early? It gives a hard number (like a credit score) that is impossible to argue with.
Layer 4 (The Storyteller - LLM): This is the AI detective. It takes the hard numbers from Layer 3 and looks at the context. It asks: Wait, is this delay because of a known strike? Is there a fraud alert? It adds a human-like explanation to the score.
Layer 5 (The Head Chef): This layer combines the Math Chef's score and the Storyteller's advice to decide: Is this risk Critical, High, Medium, or Low?
Layer 6 (The Waiter): It sends the final verdict and the explanation to the restaurant manager (the company's software) instantly.

2. The Goal: Don't Miss the Fire

In a supply chain, the worst mistake is missing a disaster (a "False Negative"). If you miss a fire, the whole warehouse burns down. If you raise a false alarm (a "False Positive"), you just waste a few minutes checking.

The authors designed this system to be extremely sensitive. They would rather scream "Fire!" 10 times when there is only smoke, than miss a real fire once.

The Result: On a massive real-world dataset (180,000+ records), their system caught 96.8% of the actual risks.
The Trade-off: Because they were so careful to catch everything, their overall "accuracy" score looked lower than some other systems. But the authors argue this is like a smoke detector that beeps loudly for every little puff of steam—it's annoying, but it saves the building.

3. The "Stress Test" (The Simulation)

To prove their system works, the authors didn't just look at past data; they built a video game simulation.

They created 120 different "what-if" scenarios, mixing delays, inventory shortages, and fraud.
They found a specific "tipping point": If you have less than 10% extra inventory AND a delay of more than 5 days, the risk of running out of stock explodes like a balloon popping. Old linear models missed this explosion; their system caught it perfectly.

4. Why It Matters for Business

The authors claim this isn't just a theory. They built it to plug directly into the software big companies already use (like SAP, Oracle, and Salesforce).

It's Auditable: Because it uses the "Math Chef" (FMEA), a manager can look at the score and say, "Yes, this is a 600 risk score because of X, Y, and Z."
It's Explainable: Because of the "Storyteller" (LLM), the system can write a plain-English note explaining why it flagged the risk, so humans understand the decision.

Summary

The paper argues that to manage modern supply chains, you can't just use rigid math (which misses context) or just use smart AI (which lacks hard numbers). You need both.

By combining a strict risk calculator with a smart AI reader, they created a system that is 96.8% good at spotting real dangers, can explain its reasoning in plain English, and fits into the software companies already own. It's like giving a supply chain manager a super-smart assistant who never misses a clue and can always back up their hunches with hard math.

Technical Summary: LLM-Augmented Decision Intelligence for Real-Time Supply Chain Risk Assessment

1. Problem Statement

Modern supply chain systems operate in volatile environments where conventional risk management paradigms—relying on static rule-based models and historical trend analysis—are structurally insufficient. These traditional approaches fail to capture compound, multi-dimensional disruptions that propagate non-linearly across interconnected supply nodes. While Large Language Models (LLMs) offer strong capabilities for reasoning over unstructured enterprise data, their standalone deployment lacks the deterministic, quantitative grounding required for auditable risk prioritization in regulated environments. Conversely, traditional Failure Mode and Effects Analysis (FMEA) provides rigorous quantitative risk quantification via Risk Priority Numbers (RPN) but is often manually intensive and insufficiently adaptive to dynamic, real-time signals.

The core problem addressed is the lack of a unified framework that combines the deterministic auditability of FMEA with the contextual adaptability of LLMs to enable real-time, high-stakes supply chain risk assessment.

2. Methodology

The paper proposes a six-layer modular decision intelligence architecture designed for enterprise integration. The framework operates as follows:

Layer 1 (Data Ingestion): Accepts heterogeneous streams (ERP, CRM, logistics feeds) and validates input schemas.
Layer 2 (Feature Engineering): Computes and normalizes six primary risk-relevant features: delay duration ( $d$ ), demand deviation ( $\Delta D$ ), inventory buffer level ( $I$ ), supplier reliability ( $R$ ), transport risk class ($TR$), and a fraud indicator.
Layer 3 (FMEA Risk Quantification): Calculates Severity ( $S$ ), Occurrence ( $O$ ), and Detection ( $D$ ) scores on a [1,10] scale to produce a deterministic Risk Priority Number (RPN) and a Stockout Probability ( $P_s$ ). The $P_s$ model is grounded in the classical newsvendor framework, capturing non-linear interactions between delay, demand deviation, and inventory.
Layer 4 (LLM Contextual Reasoning): Submits structured JSON payloads (scenario variables + FMEA scores) to an enterprise-grade LLM (GPT-4 class). The LLM performs contextual anomaly flagging, adjusts detection scores for integrity-based failures (e.g., fraud), and generates natural language risk explanations. The system uses a low temperature ( $T=0.2$ ) to ensure output determinism.
Layer 5 (Decision Intelligence Synthesis): Aggregates FMEA scores and LLM outputs via a weighted fusion mechanism. The system is explicitly calibrated to optimize Recall over Precision, based on Proposition 1, which posits that in supply chain contexts, the cost of a false negative (undetected disruption) significantly exceeds the cost of a false positive.
Layer 6 (Enterprise Output Interface): Exposes results via a REST API compatible with SAP, Oracle, and Salesforce platforms.

Experimental Design:
The framework was validated using two complementary datasets:

DataCo Smart Supply Chain Dataset: A real-world benchmark of 180,519 transaction records. The target is binary delivery risk (Late Delivery/Shipping Canceled), representing 59.1% of the data.
Parameterized Simulation: A synthetic dataset of 120 discrete scenarios designed to stress-test compound risk interactions and non-linear thresholds under controlled conditions.

3. Key Contributions

Hybrid Architecture: A novel integration of formal FMEA-based RPN scoring with LLM contextual reasoning within a unified, six-layer enterprise pipeline.
Synthetic Benchmark: A theoretically grounded 120-scenario benchmark with a deterministic stockout probability model derived from the newsvendor framework, enabling reproducible stress-testing of edge cases.
Empirical Validation: Rigorous evaluation on the DataCo dataset demonstrating statistically significant superiority over baselines, including 5-fold stratified cross-validation and ROC-AUC analysis.
Operational Threshold Identification: The identification of a critical non-linear risk threshold where concurrent conditions of Inventory Buffer < 10%** and **Delay Duration > 5 days lead to a stockout probability approaching 1.0 and non-linear cost escalation.

4. Key Results

Real-World Validation (DataCo Dataset)

Recall Performance: The proposed framework achieved 96.8% recall, significantly outperforming the Rule-Based System (70.4%) and the Decision Tree (58.3%). This high recall is critical given the asymmetric cost structure where missed disruptions are far more expensive than false alarms.
Statistical Significance: McNemar's test confirmed the superiority of the LLM-FMEA framework over the baseline with $\chi^2 = 156.4$ ( $p < 0.001$ ).
Stability: 5-fold cross-validation showed highly stable performance (Recall: $96.9\% \pm 0.4\%$ ).
Accuracy Context: While the framework's accuracy (59.3%) is lower than the Decision Tree (71.1%), the authors argue this is a principled trade-off. The system is calibrated to maximize recall in a highly imbalanced dataset (59.1% positive class), aligning with the theoretical optimal operating point defined in Proposition 1.

Controlled Simulation Results

Performance: The framework achieved 92.6% accuracy and 93.8% recall on the 120-scenario benchmark.
FMEA Validity: A strong Pearson correlation ( $r = 0.89, p < 0.001$ ) was found between the computed RPN and Cost Impact, validating FMEA as a financially grounded risk proxy.
Ablation Study: The full hybrid system outperformed "FMEA Only" (+15.7 F1 points) and "LLM Only" (+14.5 F1 points). The LLM component provided the most significant value in detecting integrity-based failures (fraud) and stabilizing classifications in borderline RPN scenarios where deterministic scoring was uncertain.

5. Significance and Claims

The paper claims that integrating deterministic FMEA quantification with LLM-based contextual reasoning yields a scalable, auditable, and operationally robust foundation for next-generation supply chain decision intelligence.

Bridging the Gap: The framework addresses the specific gap in existing literature where LLMs are used as standalone assistants rather than being integrated with formal quantitative risk frameworks.
Enterprise Readiness: Unlike proof-of-concept research, the architecture is explicitly designed for deployment via REST APIs into major ERP/CRM ecosystems (SAP, Oracle, Salesforce), with latency budgets (under 4 seconds) and cost models ( $0.01–$ 0.05 per transaction) considered feasible for enterprise adoption.
Actionable Insights: The identification of the $I < 10\% / d > 5$ days non-linear threshold provides a concrete, data-driven guideline for enterprises to adjust safety-stock policies for high-risk segments.

The authors maintain a modest stance regarding generalizability, acknowledging that while the framework performs well on the DataCo dataset, validation across diverse sectors (e.g., pharmaceuticals, defense) with different disruption dynamics is required for broader claims. The work is positioned as a validated prototype moving toward production-grade infrastructure, with future work planned for reinforcement learning optimization and multi-agent architectures.