AI Agents in Industry 4.0: AAS–OPC UA–LLM Architecture as the Foundation of Intelligent Manufacturing Systems in the Context of Industrial Enterprise Implementation

Industry 4.0 and 5.0 technologies have made industrial environments data-rich, yet a persistent cognitive gap remains: operators face substantial difficulty interpreting and acting on this data in unstructured, time-critical situations. This paper presents an architecture that integrates the Asset Administration Shell (AAS), OPC UA, and a Large Language Model (LLM)-based agentic AI within a mandatory Human-in-the-Loop (HITL) framework. The AAS acts as a semantic grounding layer through Retrieval-Augmented Generation (RAG), supplying the LLM agent with ECLASS-referenced technical parameters that reduce the risk of hallucination. OPC UA Methods form a deterministic execution layer that keeps agent actions within PLC-validated safety boundaries. The HITL mechanism enforces a cryptographic approval gate so that no physical machine action can occur without documented human authorization. This requirement was motivated by an industrial survey (n=117), in which 47% of employees stated that human oversight is irreplaceable, combined with enterprise safety and accountability requirements and broader governance considerations for AI-driven actuation in safety-critical cyber-physical systems. Two proof-of-concept case studies evaluate the architecture under controlled laboratory conditions. Proof-of-concept results indicate system processing latencies of 1.7 s (maintenance) and ∼15 s (scheduling), with end-to-end latencies (including mandatory human approval) of 14.9 s and 62 s, respectively, representing estimated improvements of approximately 97% and 96% over expert-estimated manual baselines (∼8 min and 25–40 min). All figures derive from single scripted runs under controlled laboratory conditions and should be read as indicating architectural feasibility at Technology Readiness Level 4, not as statistically validated performance benchmarks: variability bounds and confidence intervals are unavailable, the manual baselines are expert estimates rather than instrumented measurements, and operator deliberation times derive from a single response per scenario. A structured comparison with related work shows that, to the authors’ knowledge, no published approach in the surveyed literature combines AAS semantic grounding, OPC UA deterministic execution, and mandatory cryptographic HITL within a single empirically grounded framework.