Using an LLM-Powered Framework for Automatic Risk Identification and Mitigation from Past Claims and Supplemental Agreements

Examining historical claims and supplemental agreements (CSAs) can provide critical insights into the underlying factors driving project claims and change orders, thereby strengthening an organization’s overall risk management practices. However, thoroughly understanding these CSA descriptions and developing coherent risk taxonomies is both complex and time-consuming. Moreover, current methods for identifying appropriate mitigation strategies remain slow and inefficient. To address these challenges, this study presents a language model-powered framework, named Change to Mitigate (change2mitigate or C2M), that automates both the classification of risk categories and the generation of tailored mitigation strategies for historical CSAs. Specifically, bidirectional encoder representations from transformers (BERT)opic modeling is employed to cluster CSAs into major risk topics based on semantic content, after which a large language model (LLM) refines and enhances the representation of these risk categories. For mitigation, a multimodal retrieval-augmented generation (RAG)-enabled risk mitigation AI agent (RMAIA) leverages various transportation risk management databases, encompassing textual and visual resources such as best practices (BPs) and lessons learned (LL), to retrieve and synthesize effective response strategies. North Carolina State Department of Transportation bridge replacement projects are employed as pilot cases to test and validate the proposed framework. Our findings revealed that the top root causes for claims are utility relocation delays, closeout conference issues, and plan errors, whereas supplemental agreements (SAs) predominantly resulted from plan/design/method challenges, pavement issues, and underground utility complications. Validation results showed that the RMAIA can deliver timely multimodal mitigation solutions. Other transportation agencies can readily adapt this research to enhance their current risk management practices.