arXiv:2610.26889v1 Announce Type: new Abstract: Context. Despite advances in software testing and quality assurance, defects continue to recur in industrial software systems. Defect data are routinely collected in issue-tracking systems, but their use is largely confined to detection, reporting, and short-term resolution, leaving the knowledge embedded in historical defects and their fixes under-exploited for prevention and long-term improvement. Objective. This paper investigates how historical defect and defect-fix data can be analyzed, with AI assistance under human oversight, to identify the human-error origins of recurring defects and to feed that understanding back into prevention and QA improvement. Method. We developed an AI-assisted, human-in-the-loop analytics approach that classifies defects by failure type and by human-error origin across the SDLC, determines how each could have been prevented, and synthesizes engineer-focused capability-building insights. It was deployed and evaluated in a large software company through an industrial case study, using practitioner surveys and interviews, classification-reliability checks, and an effort-normalized analysis of defect density. Results. Practitioners and independent review judged the classifications accurate; the outputs were rated useful and actionable; and defect density decreased after the introduction of the feedback, a tentative association we interpret as consistent with, but not proof of, a positive effect, given the study's uncontrolled pre/post design. Conclusion. AI-assisted, human-governed analysis of defect-fix data can help shift organizations from defect detection toward prevention, continuous learning and improvement.