Design and Evaluation of a Retrieval-Augmented Conversational Framework for Employee Benefits Automation

Modern enterprises face substantial challenges in managing employee benefits, particularly during life events such as marriage, childbirth, or relocation. While conversational agents have been applied to general HR support, there remains a gap in solutions that effectively integrate structured event detection with retrieval-augmented policy guidance. This paper addresses this gap by presenting a scalable, retrieval-augmented conversational agent architecture designed to proactively detect employee life events and deliver contextual, policy-grounded guidance. The proposed framework integrates structured data ingestion from Workday, unstructured policy retrieval via a Retrieval-Augmented Generation (RAG) pipeline using Amazon OpenSearch and large language models, and event-driven orchestration through AWS Step Functions and Lambda. The primary research motivation is to evaluate whether combining real-time event monitoring with policy retrieval improves benefits decision-making accuracy and operational efficiency compared to conventional HR workflows. To assess the approach, we conducted a pilot deployment within a benefits administration environment, measuring metrics including time to resolution, reduction in HR ticket volume, and retrieval accuracy of policy content. Results indicate that the system reduced benefits-related tickets by over 40% and improved enrollment completion rates during the annual enrollment window. The architecture demonstrates potential for generalization across additional HR processes, offering a reproducible pattern for combining event-driven data pipelines and conversational AI to improve employee experience. Future work will explore multilingual capabilities, expanded life event coverage, and robustness against retrieval inaccuracies.