The four types of virtual AI agents FIGURE 1 Maturity level Specialist agents Meta agents Meta agentKnowledge agentAssistant (executing manual tasks)Recommendation (proposing scenarios and actionable insights)Automation (autonomously performing activities) Adviser agent Automation agent Source: Boston Consulting Group (BCG), World Economic Forum. Knowledge agents support workers as intelligent assistants. They analyse and synthesize vast amounts of data to provide real-time operational insights, flag anomalies and create content such as reports and code. By accessing multiple tools and real-time data sources, such as machine logs and sensor data, they add value to functions that require quick insights – for example, in maintenance, quality and logistics. They can also support engineering with machine code generation. Adviser agents go further by generating real-time scenarios to address issues, and recommending actionable insights. They continually refine their recommendations based on real-time feedback, enabling them to learn autonomously and adjust actions such as machine parameter setting, workforce management, production planning and factory layout optimization. They also suggest the best possible scenario based on their optimization objective and received feedback, empowering users to align decisions with business priorities. Automation agents act independently, executing optimal actions without human input. They adapt to new situations through real-time feedback without explicit retraining, allowing them to autonomously optimize machine performance, adjust production parameters, recode instructions or modify production plans. They surpass existing RPA (robotic process automation) by automating not only individual tasks but also entire human activities that require understanding, planning and execution. Meta agents orchestrate specialist agents in the context of multi-agent systems to achieve broader objectives, enabling area- or even factory-wide steering. The long-term vision for meta agents is to consolidate knowledge and automate end-to-end supply chains by integrating diverse specialized agents. Within a factory, these agents could cover an entire production process or group of machines. While specialist agents are already being piloted across industries, meta agents require enterprise- wide AI and further development before real- life implementation. Virtual AI has a significant impact across all manufacturing and supply chain functions, from logistics to production, as well as support functions such as maintenance, quality and engineering. The two use cases described below – production process parameter setting and real-time production planning – illustrate the agents’ capabilities.2.1 Virtual AI – paving the way for autonomous systems Virtual AI agents can manage a wide range of software-based tasks, from routine operations and research to advanced analytics and task automation. In industrial operations, they can enhance responsiveness, improve execution quality, boost productivity and reduce operational mistakes. Unlike traditional machine learning programmes, they can make context-sensitive decisions in real time and adapt through feedback loops. These agents have applications across all operation functions, including production, maintenance, quality, engineering, logistics and planning. The maturity of virtual AI agents can be categorized into three levels: assistant, recommendation and automation. The distinct objectives at each maturity level are pursued by specialist agents: Virtual AI agents have applications across all operation functions, including production, maintenance, quality, engineering, logistics and planning. Frontier Technologies in Industrial Operations 12

Frontier Technologies in Industrial Operations 2025