FinGuard-AI: A Rules-First, Explainable Framework for Narrative-Based Financial Fraud and Consumer-Harm Triage in U.S. Financial Services

Financial fraud is often visible first in a transaction record, but the complaint narrative usually explains what the ledger cannot: who contacted the consumer, what story was used, how quickly money moved, and what happened when the consumer asked for help. This paper introduces FinGuard-AI, a rules-first risk-triage system that converts those narrative details and a small set of structured case facts into transparent fraud, harm, dispute, and recovery indicators. The system uses explicit weighted phrase rules, not opaque embeddings, to classify narratives into ten substantive fraud, anti-money-laundering, consumer-stress, and disclosure-integrity categories, an "other" category, and an explicit "insufficient evidence" state for narratives that do not support a confident classification. It then produces four bounded 0-100 measures: Fraud Risk, Consumer Harm, Dispute Friction, and Recovery Urgency. Every output can be traced back to matched phrases and visible point adjustments, a nearby-window negation check keeps a denied phrase from scoring as evidence, and a structured-answer conflict check flags contradictions between a consumer's selections and their narrative for manual review. The engine described in an earlier draft of this work as a single local script has since been rebuilt as a tested, publicly deployed platform - a FastAPI backend and a Next.js frontend, covered by a 100-test automated suite and available at a public URL and source repository - so its behavior is open to direct inspection rather than description alone. Following a design-science methodology, the rules engine was exercised on 30 realistic but synthetic cases spanning all requested categories. All cases produced complete classifications, scores, explanations, and recommendations. Mean scores were 74.2, 53.5, 23.6, and 54.7, respectively, and 17 cases met the high-risk queue criterion. Because the cases were constructed for coverage, these results demonstrate functional behavior rather than predictive validity, and no amount of deployment polish changes that. The study contributes an auditable complaint-triage architecture, a multidimensional view of consumer harm, a governance mapping to the NIST AI Risk Management Framework, and a preregistration-ready plan for future evaluation using independently annotated public complaint narratives.