Retrieval-Augmented Generation as a Foundation for Trusted Enterprise Artificial Intelligence: Architectural Principles for Accuracy, Explainability, and Operational Adoption

Abstract

A majority of organizations now report regular use of generative artificial intelligence in at least one business function, and enterprise adoption has accelerated accordingly. Trust in large language model outputs has not kept pace. Hallucination, weak provenance, and reasoning chains that resist inspection continue to limit generative systems in regulated and high-stakes workflows, even as pressure to deploy them grows. Retrieval-Augmented Generation (RAG) has become the leading architectural response to this problem: rather than depending solely on parametric memory, RAG grounds model outputs in retrieved, verifiable evidence. This paper treats RAG not as a single technique but as an architectural foundation for enterprise trust, tracing its path from naive retrieve-then-generate pipelines to governed, modular architectures that build in evaluation, attribution, and auditability from the outset. Drawing on the enterprise data architecture literature, the RAG evaluation literature, and recent industry survey evidence, the paper organizes governance-relevant RAG design choices into a maturity framing, compares two leading automated evaluation frameworks, and examines the organizational barriers that determine whether technically sound RAG systems achieve real operational adoption. It contributes a structured account of how retrieval architecture, evaluation practice, and enterprise trust relate to one another, intended for practitioners building production RAG systems and researchers studying enterprise AI governance alike.