Living Digital Twins Bring AI Autonomy to Industrial Operations

Oct 28, 2026 424 views

Industrial AI’s next step depends on a practical balance: systems need enough autonomy to respond to changing conditions, while people retain control over the boundaries and the consequences. Living digital twins are emerging as the layer that can connect real-time plant data to those actions, giving operators a view of what AI is doing and a way to judge whether it is working as intended.

From plant model to operating system

A factory can behave less like a fixed collection of machines than a living system. Equipment wears down, processes drift, and early signs of trouble may appear as small irregularities in operating data. A digital model that updates only occasionally can miss those changes or describe them after the physical plant has already moved on.

Earlier digital twins often centered on simulation, monitoring and scenario planning. They could help engineers explore what might happen under different conditions, but they generally weren’t built to analyze live conditions and intervene through operational control systems. A living digital twin adds that connection: it combines a changing digital representation of assets and processes with the ability to inform actions in the physical operation.

AI is making that shift more practical. Rob McGreevy, chief product officer at AVEVA, describes AI-enabled twins as models that can adapt as a plant changes. They connect information from systems that have often operated separately, giving engineers a more current view of how equipment and processes relate. The ambition is not merely to spot a fault. It is to help keep operations within healthy limits through frequent, measured adjustments.

That points toward a form of industrial homeostasis: small interventions made as conditions change, rather than waiting for a major deviation to trigger a response. Agentic AI could eventually carry out some of those interventions itself. But the value of that autonomy depends on the quality of the plant’s live data, the accuracy of its models and the rules that define acceptable action.

A connected thread of operational data

Living twins link engineering and design records with real-time operational information in a digital thread: a connected record of an asset across its lifecycle. More affordable sensors now stream telemetry from internet-connected devices and industrial control equipment. Software platforms bring those feeds together, while predictive and physics-based models can help estimate how a machine or process may behave.

The result gives operators several ways to examine the same operation. They can ask whether emissions are changing, whether an asset is performing within a healthy range, or whether production is meeting its target. McGreevy describes these as different “lenses” on the twin, including sustainability, maintenance and performance. For AI, those lenses also help connect a detected change to the outcome the business wants to protect.

That link matters because industrial decisions rarely optimize a single number. A change that increases output could affect equipment health or emissions. A useful system needs context about the process and its priorities, and operators need a way to see which conditions informed an AI recommendation or action. A living twin can provide a shared operational picture, but it does not remove the need to define what “healthy” means for a particular plant.

Adoption is advancing alongside caution

Manufacturers are interested in letting AI take a more active role. In a 2026 Manufacturing Leadership Council survey of manufacturers (opens in new tab), 66% said they were using or planned to use agentic AI in operations, and 40% expected some degree of factory autonomy by 2030. Those plans come with a clear concern about disruption: in Deloitte’s 2025 smart manufacturing survey (opens in new tab), 65% of executives ranked operational risk, including business disruption, among their top concerns.

Source: MIT Technology Review Insights with data from Manufacturing Leadership Council (opens in new tab), 2026

McGreevy says the technology still needs refinement, testing and validation. He expects many deployments to resemble supervisory control and data acquisition, or SCADA, systems, where people remain involved in the decision loop. AI may advise or handle a narrow, repetitive task, while operators supervise broader activity. That staged approach gives companies a way to build evidence about how a system behaves before granting it more responsibility.

Human roles are changing with the technology. Operators who once focused on one production line or maintenance task may need to work across several disciplines. Routine approvals could take less time, leaving more attention for setting operating limits and recognizing when AI’s interpretation of data doesn’t match practical experience. Because an AI system can reinforce patterns in its inputs, people remain responsible for catching problems early and checking whether the system’s actions fit the real operating context.

Oversight needs to keep pace

Autonomy requires clear rules about which decisions a system can make, how people monitor its behavior and what records are kept for review. Yet governance practices vary. A study from AVEVA and IMD (opens in new tab) found that 43% of organizations said AI had changed decision-making and accountability across ecosystem partners only slightly or not at all; 23% reported significant or fundamental change. The same study found that 27% shared data substantially or extensively with partners.

The oversight model also depends on the application. Some systems ask an operator to approve every action. In others, AI can act within limits set by people, who monitor performance and intervene when needed. The choice affects how quickly a system can respond and how much responsibility operators carry. It also makes the twin’s visibility important: people need to understand the conditions behind an action, not simply see that one occurred.

Rules and company policies are beginning to address these questions. The EU AI Act treats certain AI safety components used in critical infrastructure as high-risk and requires human oversight. AVEVA says it has a responsible AI governance board and published principles for AI development and deployment. These structures are part of a wider effort to define accountability as AI moves closer to operational control.

Connecting plants and industrial networks

The coordination challenge reaches beyond an individual facility. Manufacturing and critical infrastructure have often operated through separate systems, McGreevy says. One emerging route to broader coordination combines robotics, sensors and AI decision-making. Drones inspecting pipelines or power lines, for example, can collect visual information that algorithms assess for gas leaks and other hazards. That can give teams more connected inspection data while reducing the need for people to enter dangerous locations.

McGreevy sees this convergence of robotics and software as a way to reduce incidents and outages and support more sustainable operations. The practical payoff depends on how well inspection data joins with maintenance and control information, and how reliably teams can act on what the system finds.

At a larger scale, living digital twins could represent an integrated value chain rather than a single plant. McGreevy says this could help companies respond to shifts in supply, demand, tariffs and material availability. The idea is to connect operating information across industrial networks so that a change upstream can inform decisions elsewhere.

For now, the most credible path is measured: keep people involved, validate AI behavior and expand its authority where the task and evidence support it. Living digital twins offer a way to connect that oversight to live operations. Their long-term value will depend on whether companies can build shared, reliable data and clear decision boundaries across the systems they want AI to coordinate.

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

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

Pairing industrial AI autonomy and human control