Bank Customer Support Chatbot

How should an AI chatbot answer sensitive banking questions and how do different information-access strategies affect the quality and safety of its responses? This project builds and benchmarks three approaches for a policy-sensitive banking chatbot: a standalone LLM relying on general training knowledge, a retrieval-only system that surfaces relevant regulatory passages directly, and a RAG (Retrieval-Augmented Generation) system that grounds LLM responses in retrieved FDIC and CFPB guidance. Each approach is evaluated against realistic customer support scenarios covering fraud, banking complaints, and error resolution, revealing clear tradeoffs between fluency, accuracy, and regulatory grounding in high-stakes AI applications.