International Review of Business, Trade, and Economics

Governing Enterprise Knowledge with On-Premise Retrieval-Augmented Generation: The AI42 Design Framework

Abstract

Cristian Cosentino, Natale Motta, Carlo Sarugeri, Tommaso Maggisano and Massimiliano Bacchetta

Enterprise adoption of generative artificial intelligence is constrained less by access to capable language models than by the difficulty of connecting those models to fragmented, sensitive, and differently governed organizational knowledge. This paper presents Governed Enterprise Retrieval-Augmented Generation (GE-RAG), a design framework that treats retrieval-augmented generation as a data-governance control surface rather than only a question-answering technique. The framework is instantiated through AI42, an on-premise enterprise AI platform developed by Moveo. Following a design-science approach, the study derives governance requirements from the data-governance, information-retrieval, trustworthy-AI, and technology-adoption literature and traces them to observable architectural mechanisms in the industrial artifact. The analysis identifies six linked capabilities: governed knowledge onboarding, identity- and context- aware retrieval, bounded evidence construction, source-grounded generation, auditable interaction, and feedback-driven improvement. The on-premise boundary strengthens organizational control over storage, processing, model execution, and operational telemetry, but is explicitly treated as an enabling condition rather than proof of compliance or answer correctness. The paper contributes a governance-to-architecture mapping and a preregistration-ready evaluation protocol covering retrieval effectiveness, groundedness, access-control conformance, task-completion time, workload, usability, and user trust. The resulting methodology shows how enterprise RAG can make existing knowledge more accessible while preserving decision rights, provenance, and accountability. Empirical productivity claims are intentionally deferred until the proposed controlled user study is completed with real organizational data.

PDF

Journal key Highlights