FlyersSoft's Divakar Kumar proposes integrating AI agents into event-sourced systems to resolve "gray zone" transactions in fraud detection, where rule-based and ML models are uncertain. Kumar recounts a personal frustration: his $3,500 laptop purchase in late June 2026 was repeatedly declined by his bank’s fraud system, and customer service could not explain why. He argues that AI agents, integrated via an orchestrator layer and a shared semantic projection, can resolve these cases without replacing proven systems. Kumar frames the challenge with a three-tier model: rule-based engines (5+ years old) and third-party ML models handle clear "approve" and "block" zones, but a significant fraction of transactions fall into a gray zone. His architecture follows Domain-Driven Design (DDD) principles, with four bounded contexts (Transaction, Account, Device, Payment) each owning data and communicating asynchronously. The transaction context uses event sourcing with Cosmos DB as its event store; a change feed propagates new events to multiple read models. Events from other contexts arrive via a message broker, yielding a projection layer for agent consumption.