Living Digital Twins Bring AI Closer to Real-Time Industrial Control

| 2 Min Read
AI-enabled digital twins connect live plant data with operational controls, while human oversight keeps emerging factory autonomy within defined limits.

Industrial AI is moving from spotting problems to helping manage the machinery that causes them. Living digital twins are central to that shift: they link a changing digital model of a plant to real operating data and, increasingly, to systems that can act on it. The practical opportunity is faster, more coordinated operations. The constraint is that autonomy still needs clear boundaries and informed human oversight.

From plant model to operating system

A factory’s condition changes continuously. Equipment wears, processes drift, and small irregularities can be early signs of a larger fault. Traditional digital twins have helped engineers simulate conditions and explore possible outcomes, but many remained detached from live operations. They could represent a facility without keeping pace with it or changing what happened on the plant floor.

A living digital twin aims to close that gap. It brings together a high-fidelity model of assets and processes, current operational data, and connections to control systems. That combination can make the model useful during operations, not only during design or planning. As AVEVA chief product officer Rob McGreevy describes it, AI-infused twins can adapt as conditions change, providing a connected view of systems that may previously have operated in isolation.

The underlying data connection is often called a digital thread: a linked record of an asset across its lifecycle, joining engineering and design information with operational data. More affordable sensors now stream readings from internet-connected devices and industrial control equipment. Software platforms can bring those readings into interfaces for operators, while predictive and physics-based models help estimate how machinery or processes may behave.

That connected view can serve different operational priorities. McGreevy says operators might use a twin to examine emissions, check whether an asset is performing healthily, or determine whether production is meeting a target. The value lies partly in bringing those perspectives together: decisions about maintenance, output, and sustainability can draw on a shared picture of plant conditions.

AI can act, but autonomy remains bounded

Agentic AI adds another step by enabling systems to respond to selected insights rather than simply display them. In principle, frequent, small adjustments could help keep operations within desired conditions. That prospect has attracted manufacturer interest. A 2026 Manufacturing Leadership Council survey found that 66% of manufacturers were using or planned to use agentic AI in their operations, while 40% expected some degree of factory autonomy by 2030.

Interest comes with operational caution. In Deloitte’s 2025 smart manufacturing survey, 65% of executives ranked operational risk, including possible business disruption, among their leading concerns. That concern helps explain why the near-term role of AI is likely to stay supervisory. McGreevy compares the expected path with SCADA systems, where the “S” stands for supervisory: a person remains part of the control loop.

“There is a lot more refinement to be done, a lot more proof points, testing, and validation needed,” McGreevy says of the work required before AI systems can operate with greater independence. For now, he expects full autonomy to remain limited to narrow, highly repetitive tasks, with AI more commonly advising or assisting people.

That division of labor changes what operators do. Instead of approving every routine action, workers may spend more time defining acceptable operating limits and checking whether the system’s decisions remain appropriate. The job can also draw on broader expertise: people who once focused on one process line or maintenance task may work across several disciplines. Their experience helps them interpret what AI detects and catch errors or misaligned behavior early.

A digital twin can support that oversight by showing how AI actions relate to live plant conditions. But visibility alone doesn’t set the rules. Organizations need to decide which actions a system may take independently, where human approval is required, and how actions will be monitored and reviewed. The right arrangement can vary: an operator might approve each action in one deployment, while another allows the system to act inside limits set and supervised by people.

Governance is part of the operating design

Company-level safeguards are still developing. In a study by AVEVA and IMD, 43% of organizations said AI had changed decision-making and accountability across ecosystem partners only slightly or not at all. Just 23% reported a significant or fundamental change. Data sharing between partners was also limited: 27% said they shared data substantially or extensively.

Those findings point to a challenge beyond model performance. Industrial AI often depends on information that crosses teams, systems, or company boundaries. When accountability and data sharing haven’t kept pace, it becomes harder to coordinate decisions or establish who is responsible for them. Some organizations are setting internal structures for AI governance; AVEVA, for example, has a responsible AI governance board and published principles for developing and deploying AI. The EU AI Act also sets human-oversight requirements for certain AI safety components used in critical infrastructure.

For plant leaders, governance has to connect to actual operations. Rules need to define how much authority a system has, what conditions trigger escalation, and how people can see what it did. These choices affect whether automation can earn trust in day-to-day work, especially where a faulty action could disrupt production.

Coordination can extend beyond the factory

Living digital twins may also help connect industrial assets across larger networks. McGreevy argues that manufacturing and critical infrastructure have long operated through separate systems. Physical AI, which combines robotics and sensing with AI-based decisions, offers one route toward more coordinated information. Drones inspecting pipelines or power lines, for example, can collect visual data that algorithms analyze for gas leaks and other hazards. The approach could reduce the need for people to enter dangerous environments.

The wider aim is an “integrated value chain”: digital representations that extend across companies and industrial operations. McGreevy says those models could help businesses respond to changes in supply, demand, tariffs, and material availability. Their usefulness would depend on connecting information across the chain, making the data available when conditions change, and maintaining clear responsibility for decisions made from it.

The immediate measure of progress, then, isn’t simply how many tasks an AI system can perform without assistance. It’s whether living twins give operators a timely, connected view of the plant, and whether automation can act within limits people understand and can supervise. That combination is likely to shape how quickly industrial organizations expand autonomy from tightly controlled tasks to broader operations.

This content was produced by Insights, MIT Technology Review’s custom content arm, not its editorial staff. It was researched and written by humans, with any AI tools that may have been used limited to production processes under human oversight.

Source: MIT Technology Review Insights · www.technologyreview.com

Comments

Please sign in to comment.