Manual assessment of adverse drug reaction (ADR) causality can require reviewing hundreds of medical records per case, creating a workload that can contribute to underreporting and inconsistent assessments. We present an evidence-guided AI agent for automated Naranjo assessment using real-world electronic medical records (EMRs) generated during routine clinical care. A large language model extracts candidate evidence from clinical notes, medication records, and test results. Predefined rules verify its source, relevance, and timing before guiding item-level answers and calculating scores. This separation is designed to reduce hallucinations, or unsupported AI-generated claims, by requiring evidence verification before causality judgments. The agent searches additional records for unresolved questions and returns answers, scores, causality categories, and supporting evidence. In a preliminary evaluation using EMRs from 19 development cases reported at Gachon University Gil Medical Center (2020-2023), item-level agreement under the evaluation criteria was 94.1% across eight Naranjo questions (152 responses), and causality category agreement was 89.5% (17/19). All 309 checked quotations matched source records. By integrating evidence retrieval, verification, and scoring, this framework provides a traceable foundation for automating pharmacovigilance review.