AI PB: A Grounded Generative Agent for Personalized Investment Insights

Compliance-First Architecture is Feasible at Scale: AI PB demonstrates that generative AI can operate in highly regulated financial environments by implementing deterministic routing between internal and external language models based on data sensitivity levels, rather than relying on a single model for all tasks. This component-based approach allowed the system to meet Korean financial regulations while maintaining production deployment.

Grounding and Retrieval Reduce Hallucination Risk: The system combines domain-specific embedding models with traditional search (OpenSearch) to ground investment recommendations in actual data rather than generating unsupported claims. This hybrid retrieval pipeline was critical for ensuring the AI provided factual, verifiable insights in a domain where inaccurate information carries real financial consequences.

Layered Safety Mechanisms Enable Trustworthiness: Rather than relying on a single safety measure, AI PB employs multiple safeguards including rule-based heuristics, behavioral modeling, and contextual bandits for recommendations. The paper's emphasis on "layered safety" suggests that responsible AI in finance requires redundant verification systems rather than trusting any single component.

Proactive Personalization Requires Careful Design: Unlike passive chatbots, AI PB generates unsolicited, user-specific investment insights. This proactive approach demands particular attention to personalization accuracy and bias—the system uses sequential behavioral modeling to understand individual users, raising questions about how such systems should handle diverse financial situations and risk tolerances.

On-Premises Deployment Supports Accountability: By operating fully on-premises rather than relying on cloud-based third-party APIs, the organization maintains direct control over data, model behavior, and audit trails—a practical approach to responsible AI governance in sensitive financial contexts.