Data Infrastructure Modernization as a Foundation for Real-Time AI Decisioning: A Case Study in Relational-to-NoSQL Migration for Large-Scale Retail Platforms

Abstract

Real-time AI/ML-driven decisioning is increasingly central to consumer platforms, from fraud detection to product-eligibility scoring — but these systems depend on a data layer capable of low-latency reads, horizontal scale, and high availability, which legacy relational architectures are often not designed for. This paper presents a case study of migrating a set of Oracle-based enterprise applications, spanning customer profile management and loyalty programs at a large multinational retailer, to a distributed multi-model NoSQL architecture (Azure Cosmos DB) alongside PostgreSQL, carried out alongside decomposing seven-plus monolithic applications into highly available microservices across more than fifteen tier-1 and ten tier-2 production systems. We report the operational outcomes observed — an approximately 50% reduction in infrastructure cost and sustained 99.99% availability (100% during peak holiday traffic) — and discuss the architectural decisions (service decomposition boundaries, phased cutover, rollback strategy) that made them achievable. We then connect these data-layer modernization patterns to the broader industry shift toward embedding AI-driven decisioning directly into consumer-facing transaction flows, illustrating with a public example from a separate large-scale fintech platform where a comparable low-latency data foundation is reported to have supported an AI-based decisioning engine that improved decision speed by approximately 75%. We argue that the data-layer modernization choices described here are a necessary, and frequently underexamined, prerequisite for organizations pursuing real-time AI decisioning at scale.