Agentic Systems for Failure Attribution and Root Cause Analysis

Abstract

Automotive warranty cases are valuable for quality follow-up, but they are not written as root-cause reports. A record can mix customer symptoms, workshop actions, replaced assemblies, parts context, and partial causal clues. The recorded repair scope can describe what was done in service while still hiding the component that most likely initiated the failure. Human reviewers can recover that meaning, but the work is slow and depends on engineering experience. This thesis aims to develop a planner-guided agentic system for failure attribution and root cause analysis in warranty cases. Instead of treating repair text as a single classification input, the system separates text normalization, signal extraction, hypothesis formation, optional critique, dictionary-constrained mapping, historicalcase retrieval, and final judgement. The intermediate outputs are stored with the final decision, so a reviewer can see the fields, comparison cases, and constraints behind the selected label. The evaluation combines manual review analysis, a direct single-pass baseline, mechanism variants, field-level checks, retrieval checks, knowledge-use analysis, and finallabel scoring. Manual review finds 333 CHG-related records inside a 628-record engine-exchange candidate set, showing how component-level failure patterns can sit inside a broad repair outcome. On the 357-record positive benchmark, targetlabel recovery increases from 29.4% with the direct single-pass baseline to 84.3% with the planner-guided workflow, while other non-target outputs fall from 67.2% to 7.6%. On the mixed 628-record basis, the workflow reaches 85.4% accuracy and 87.3% F1 for the target label. ANN retrieval preserves the exact top result in 99.4% of the current retrieval index, and knowledge entries are used in 490 of 628 completed records. The results indicate that an agentic workflow can improve failed-component recovery and make the result easier to inspect than a single-pass request. It does not replace engineering judgement or independently prove physical root cause. Its role is to organize noise from workshops into a standardized label, structured support, and a shorter path for expert review.