AI-Enhanced Risk Assessment for Improved Supply Chain Resilience

Sammanfattning

Introduction: This thesis explores the integration of Artificial Intelligence, specifically Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG), to quantify qualitative disruption signals in global supply chain risk management. As modern high-tech supply chains face increasingly unpredictable geopolitical, macroeconomic, and multi-tier shocks, translating heterogeneous "soft signals" and unstructured text into actionable data is crucial for transitioning from static, intuition-based risk matrices to dynamic, data-driven inventory buffering. Research Question: The primary research question is: How can qualitative expert judgments and heterogeneous disruption signals be quantified and estimate a Component Risk Score based on machine learning algorithms in a globally expanded complex supply chain? Method: To address this, a multi-modal AI architecture - the Component Risk Estimator was developed and evaluated using historical operational data and market intelligence across 26 critical electromechanical supply chain categories. The methodology utilized a sequential, time-indexed RAG pipeline to extract and fuse features from unstructured market reports and deterministic operational data. A Random Forest Regressor was then employed to compute a continuous Component Risk Estimate (CRE), which was comparatively evaluated against the historical heuristic baselines of human Global Material Managers (GMMs). Results: The results indicate that advanced LLM architectures can successfully quantify and replicate human heuristic judgment with high mathematical precision. The primary model (GPT-5.4) achieved an 89.7\% Within-Margin Accuracy and 100.0\% Ordinal Adjacency, ensuring complete operational safety with zero catastrophic multi-tier misclassifications. Conversely, a comparative analysis highlighted that smaller, lightweight models (GPT-4o and GPT-4o-mini) suffer from instruction-following decay and a severe "Alarmist Bias," defaulting to extreme risk over-estimations when exposed to macroeconomic noise. Discussion: These findings demonstrate that advanced, well-prompted AI frameworks can safely standardize subjective risk assessments across a global enterprise, allowing supply chain professionals to cognitively offload manual data processing and focus on strategic resilience. While the study's scope is currently delimited to electromechanical components and relies heavily on high-quality external data retrieval, it establishes a functional quantitative state-vector for risk. Future research should explore integrating this continuous risk output directly into Deep Reinforcement Learning (DRL) agents to achieve fully autonomous inventory safety-stock execution.