A Taxonomy and Reference Architecture for AI-Enabled Industrial Knowledge Management

Abstract

Much of the knowledge that keeps an industrial operation running is never written down: it sits with individual employees, scattered across incompatible systems, and cannot be found in time—a risk that becomes acute whenever people change roles or retire. Artificial intelligence (AI) is widely proposed as a remedy, but existing approaches concentrate on documentary, white-collar work; embodied, blue-collar work is comparatively underserved, particularly by large language models with no native grounding in physical activity. Existing work also treats AI as a single undifferentiated capability, leaving practitioners without a principled basis for choosing among technologies or integrating their outputs. This paper proposes a taxonomy of five AI paradigms (perceptive, dialogic, interpretive, structural and contextual), each defined by its contribution to one of three knowledge processes (capture, structuring, transfer) and by the type of work it serves, with a boundary marked where collective tacit knowledge resists codification. A separate orchestration layer, realised by autonomous agents and enabling technologies (knowledge graphs, retrieval-augmented generation, augmented reality), connects the paradigms into a pipeline. Applied diagnostically to three tools in one manufacturer’s training programme, an assembly-guidance system, a structured interview system, and a RAG-based conversational assistant, the taxonomy shows all three occupy the capture or transfer columns while structuring goes unserved, leaving each tool’s knowledge inaccessible to the others. A pump-assembly scenario shows how an agent-orchestrated pipeline over a shared knowledge graph, with human validation, could unify their outputs. The tools’ reported gains, a 29 per cent onboarding-time reduction and an over 90 per cent retrieval-time reduction for 3,000+ daily users, are taken at face value; whether integration compounds them, and for whom, is a working hypothesis, not a demonstrated result. The paper concludes with a staged evaluation strategy measuring cross-tool retrieval coverage and validation throughput to isolate the structural layer’s impact.