Industrial AI Autonomy and Human Control: How Living Digital Twins Are Reshaping Factories

Industrial environments are often imagined as cold steel and machinery, but functionally, a factory is more like a living organism: it has a pulse, even a metabolism; it wears and ages over time, and signs of declining health often manifest as barely perceptible irregular fluctuations. Digital twin technology has long sought to mirror this dynamism, but early implementations were always half a beat behind—unable to analyze in real time or intervene proactively. Today, AI-centric "living digital twins" are changing this, offering not only high-fidelity mapping of complex operations but also the ability to intervene in the physical world through operations and control systems.

From Mirroring to Intervention: The Essential Leap of Living Digital Twins

AVEVA Chief Product Officer Rob McGreevy likens the new generation of AI-empowered digital twins to organisms that grow and evolve. These "living" versions are adaptive, evergreen models of plant assets and processes—they track real-time conditions and connect previously isolated systems, providing a visibility and orchestration layer that was missing before, better attuned to the organic rhythms of the facility itself.

Unlike earlier digital twins focused on simulation, observation, and hypothetical scenario planning, living digital twins combine dynamic, high-fidelity digital representations with the ability to intervene in the physical world. Agentic AI further pushes industrial systems beyond diagnosis and recommendation to autonomous action. As the technology evolves toward greater autonomy, it begins to develop a kind of "reflex" for industrial homeostasis: continuous, micro-adjustments that keep the plant running healthily.

But human judgment and operator oversight remain indispensable—they define what "healthy" means and alert when AI systems veer off track. Living digital twins give engineers the means to monitor AI behavior and ensure execution stays on course.

Agentic AI differs from conventional generative AI in its "right to act." Generative models output text or suggestions, while agentic systems can plan steps, invoke tools, execute actions, and adjust strategies based on feedback given a goal. In industrial settings, this means AI doesn't just tell an operator "the valve should be closed 10% more"—it can directly issue commands through the control system. Agentic AI typically comprises four modules: planning, memory, tool invocation, and reflection, with autonomy levels ranging from advisory-only, to execution after human approval, to full autonomy within preset boundaries. Because of its ability to act, agentic AI demands far more governance, audit, and fail-safe mechanisms than purely analytical AI—this is the technical root of the article's emphasis on "humans defining health boundaries."

The Technical Thread That Maintains Synchrony: Digital Thread and Sensors

Living digital twins connect engineering and design information with real-time operational data through a digital thread—an interconnected record spanning the asset lifecycle. Two notable technological shifts have helped stitch this thread together: sensors have become more affordable in recent years, now continuously transmitting telemetry from IoT devices and industrial control equipment; and software platforms bring this information into user-friendly interfaces. Predictive and physics-based models provide AI-enhanced insights, anticipating machine or process behavior.

The operator's perspective on the digital twin also evolves. McGreevy gives an example: "You can look at the digital twin from a sustainability perspective: what are my emissions? Or from a maintenance perspective: is this asset running healthily? Or from a performance metrics perspective: is it producing the quantity of product I need?"

Digital Thread is often confused with digital twin, but their roles differ. A digital twin is a dynamic virtual replica of an asset or process, answering "what is the state now"; a digital thread is an interconnected data链路 spanning design, manufacturing, operations, and maintenance throughout the lifecycle, answering "where did this state come from and what changes has it undergone." The digital thread links CAD drawings, engineering change orders, sensor telemetry, and maintenance records into a traceable record, keeping the twin model synchronized even after asset modifications, avoiding "model drift." Without a digital thread, a twin is just a snapshot at a moment in time; with it, the twin can become the "evergreen model" described in the article and trace anomalies back to design or process root causes.

Setting Boundaries for Semi-Autonomous Systems: Data and Governance Realities

The next stage is for AI to act on selected insights revealed by these perspectives. Manufacturers are positive: according to a 2026 Manufacturing Leadership Council survey, 66% of manufacturers say they are using or plan to use agentic AI, and 40% expect some degree of factory autonomy by 2030. But risk concerns are real: in Deloitte's 2025 Smart Manufacturing survey, 65% of executives cited operational risk (including the possibility of business disruption) as a major concern.

McGreevy believes that moving AI toward autonomy still requires substantial refinement, validation, and testing. "There's more refinement work to do, more corroboration, testing, and validation needed." In the near term, he expects agentic AI to follow the path of industrial technologies like SCADA systems. "The 'S' in SCADA stands for supervisory, and supervisory means there's a human in the middle." In his view, for the foreseeable future AI will mainly be in advisory mode, with full autonomy limited to narrow, highly repetitive tasks.

Human operators' daily work is also changing. Workers who once specialized in a single production line or maintenance task can now take on multidisciplinary responsibilities; they spend less time approving routine operations one by one and more on defining acceptable operating boundaries. Because AI may reinforce patterns in the data it receives, early detection of missteps becomes central, and human knowledge and experience in interpreting what AI sees and does is crucial.

Governance at the company and industry level also needs to keep pace. Organizations must establish clear guardrails for the scope of AI's autonomous decision-making and how its actions are monitored and audited. Yet many enterprises have not yet built comprehensive safeguards: recent AVEVA and IMD research shows that 43% of organizations report AI has only slightly or not at all changed how decisions and accountability work among ecosystem partners, with only 23% reporting significant or fundamental change (and only 27% reporting substantive or extensive data sharing with partners).

Meanwhile, regulators are beginning to define the rules. The EU AI Act classifies certain AI safety components used in critical infrastructure as high-risk and mandates human oversight. Companies are also establishing their own structures; AVEVA has a Responsible AI Governance Council and has published principles for AI development and deployment.

SCADA (Supervisory Control and Data Acquisition) is a classic layer of industrial automation: sensors and PLCs handle acquisition and execution, SCADA handles centralized monitoring and human-machine interfaces, and operators retain supervisory authority in the loop. McGreevy's SCADA analogy for agentic AI is meant to illustrate the industrial path to accepting "supervised autonomy"—first let AI advise at the monitoring layer, then issue commands after human approval, gradually transitioning to limited autonomy. This analogy also implies the technical architecture: agentic AI is likely to be embedded atop existing SCADA/historian systems rather than replacing底层 control, so its autonomy boundaries can be constrained by existing permissions and interlock mechanisms.

The EU AI Act adopts risk-based tiered regulation: unacceptable risk, high risk, limited risk, and minimal risk. AI used for safety components in critical infrastructure (power, water, gas, etc.) is classified as high-risk and must meet obligations for risk management, data governance, technical documentation, logging, human oversight, and accuracy/robustness. For industrial AI, compliance focuses on the demonstrability of "human-in-the-loop"—systems must be designed so humans can effectively supervise, intervene, and override, with audit logs retained. This aligns with the governance approach described in the article—"humans define boundaries, monitor AI behavior"—and means that living digital twins for the EU market need built-in explainability and oversight interfaces at the architecture stage.

Case Studies: From Refineries to Grid Control Rooms

In semi-autonomous systems already in operation, oversight varies by deployment: in some scenarios operators approve every action; in others AI acts autonomously while humans supervise and set the boundaries within which it can operate. McGreevy cites several AVEVA customer examples:

• Saudi Aramco's Yanbu refinery: optimization algorithms adjust production in a closed loop, with operators supervising as needed; at Ras Tanura refinery, algorithms recommend changes and operators decide whether to implement them.
• Nestlé: at the Nesquik factory, AI optimizes and adjusts industrial dryers for powdered beverage ingredients. In a landmark test, the tuning reduced energy consumption by 16% while improving quality and yield.
• SCG Chemicals in Thailand: AI-supported digital twins spanning engineering, operations, and maintenance monitor asset health and flag potential failures. The company reports 99% plant reliability and a ninefold ROI within six months.
• National Renewable Energy Laboratory (NREL) in the U.S.: built a generative AI model for grid control rooms to provide oversight at scale, supporting operators with renewable energy forecasting and grid condition estimation.

The Future Landscape: From Single-Plant Synchrony to Value Chain Collaboration

The synchronization once missing among a single plant's technical systems is equally scarce at larger scales. "For the past 50 years, manufacturing and critical infrastructure have operated as island systems," McGreevy says. But he believes certain emerging technologies are creating opportunities for more holistic coordination.

At the enterprise level, this may involve the integration of physical AI: combining robotics and sensing technologies with AI-driven decision-making. For example, drones flying over pipelines and power lines can feed visual inspection data to algorithms to detect gas leaks and other hazards. This not only creates more accurate, interconnected information flows but also keeps humans out of potentially dangerous environments. "The convergence of robotics and software will mean fewer accidents, less downtime, and more sustainable operations," McGreevy predicts.

The broader version extends beyond any single factory or process to the entire industrial network—what McGreevy calls the "integrated value chain." "Living digital twins make digital representation of the entire value chain possible," he says. "This allows companies to respond quickly to supply chain changes, demand shifts, tariffs, and material availability."

In today's reality of geopolitical turbulence and shifting regulatory demands, such resilience will be "an absolute necessity." McGreevy adds: "Without it, it's very difficult to run a business effectively."

One-line Summary

AI-empowered living digital twins move factories from passive mirroring to closed-loop intervention, but human oversight and governance remain the boundary of autonomy.

Paragraph Summary

The article notes that early digital twins were "half a beat behind" due to their inability to analyze in real time and intervene proactively, while AI-centric living digital twins connect engineering data with real-time telemetry through a digital thread, combining high-fidelity mapping with the ability to intervene in the physical world. Agentic AI drives systems from diagnostic recommendations toward autonomous action, forming a "reflex" that maintains industrial homeostasis, but in the near term it will remain largely advisory, with full autonomy limited to narrow repetitive tasks. The human role shifts to defining "health" boundaries, supervising AI behavior, and interpreting insights. Surveys show 66% of manufacturers are using or planning to use agentic AI, but 65% of executives worry about operational risk; governance frameworks and the EU AI Act are catching up. Cases such as Nestlé's 16% energy savings and SCG Chemicals' ninefold ROI validate the value, and the future direction is physical AI and integrated value chain collaboration.