AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Pairing industrial AI autonomy and human control

Collected Oct 28, 2026

Industrial software vendor AVEVA is positioning what it calls "living" digital twins as the next stage of factory and plant automation, according to a sponsored MIT Technology Review piece built around an interview with Rob McGreevy, the company's chief product officer. The core claim is that digital twins are shifting from passive simulation toward models that stay synchronized with live plant conditions and can intervene in the physical world through operations and control systems.

The distinction matters technically. Earlier digital twins were used mainly for observation and what-if scenario planning, and they typically lagged physical reality by some margin, so they could not perform real-time analysis or act on what they found. A living twin, as AVEVA describes it, is an adaptive, continuously updated model of assets and processes that connects systems which previously ran in isolation. That connection supplies visibility and orchestration across a facility rather than a static replica of it.

Underneath sits a "digital thread" — a connected record of an asset that persists across its lifecycle. Two developments make this practical. Sensors, whose costs have fallen in recent years, stream telemetry from IoT devices and industrial control equipment, and software platforms aggregate that telemetry into interfaces operators can read. Predictive and physics-based models then layer AI-derived insight on top, estimating how a machine or process is likely to behave. McGreevy frames the value in terms of interchangeable lenses: emissions and sustainability, asset health and maintenance, or output against production targets.

The more consequential shift is toward agentic AI — systems that go past diagnostics and recommendations to act on their own. McGreevy ties this to a notion of industrial homeostasis, in which small continual adjustments keep a plant healthy. Adoption intent is high. In a 2026 Manufacturing Leadership Council survey cited in the piece, 66% of manufacturers said they are using or plan to use agentic AI in operations, and 40% expect some degree of factory autonomy by 2030. Risk appetite is lower: in Deloitte's 2025 smart manufacturing survey, 65% of executives ranked operational risk, including business disruption, among their top concerns.

McGreevy's own expectation is incremental. He says more refinement, proof points, testing and validation are needed, and he compares the likely near-term path to SCADA, the supervisory control and data acquisition systems that have long underpinned industrial control. The supervisory in that name implies a human in the middle, and he expects AI to stay largely advisory for the foreseeable future, with full autonomy confined to narrow, highly repetitive tasks. This is a useful calibration for anyone planning automation roadmaps: the near-term product reality is suggestion-with-approval, not hands-off control.

Human oversight remains load-bearing in this model. Engineers define what healthy operation looks like in the first place and flag when AI behavior drifts out of alignment. Operator jobs change shape: people who specialized in one process line or maintenance task take on multiple disciplines, spend less time approving routine actions one at a time, and spend more time setting acceptable operating limits. Because AI can amplify patterns present in its training data, catching errors early becomes a core responsibility rather than an exception. In practice, this suggests autonomy shifts work from execution to boundary-setting and exception handling — a skill profile closer to control engineering than to manual operation.

Governance has not kept pace. In a study by AVEVA and IMD, 43% of organizations said AI has changed decision-making and accountability across ecosystem partners only slightly or not at all, while 23% reported a significant or fundamental change; only 27% said they share data substantially or extensively with partners. Regulation is beginning to bite: the EU AI Act classifies certain AI safety components used in critical infrastructure as high-risk and requires human oversight. AVEVA says it maintains a responsible AI governance board and published principles for AI development and deployment.

The piece offers several deployment examples. At Saudi Aramco's Yanbu refinery, optimization algorithms adjust production in a closed loop with operators overseeing as needed; at Ras Tanura, the same class of algorithms recommends changes that operators choose to implement or not — a neat illustration of the difference between closed-loop and advisory modes at the same customer. Nestlé uses AI optimization on industrial dryers at Nesquik plants, where a landmark test cut energy use by 16% while improving quality and yield. SCG Chemicals in Thailand runs an AI-supported digital twin spanning engineering, operations and maintenance that monitors asset health and flags potential failures; the company reports 99% plant reliability and a ninefold return on investment within six months. The US National Laboratory of the Rockies has built a generative AI model for grid control rooms that supports operators with renewable energy forecasting and estimates of grid conditions.

Looking further out, McGreevy points to physical AI — robotics and sensing paired with AI decision-making — such as drones inspecting pipelines and power lines and feeding visual data to algorithms that spot gas leaks and other hazards. He argues this also removes humans from dangerous environments, and predicts fewer incidents and outages from the convergence of robotics and software. He extends the idea to an "integrated value chain," arguing that living twins make digital representations of whole supply chains possible, letting companies respond faster to demand shifts, tariffs and material availability. He characterizes manufacturing and critical infrastructure as having operated as siloed systems for 50 years, and calls resilience an absolute necessity amid geopolitical and regulatory turbulence.

A reasonable reading is that the technology story here is less about any single model than about integration: telemetry, lifecycle data and control systems joined well enough that an agent's action can be trusted and audited. The hard parts are unglamorous — sensor coverage, data quality, latency, and clear records of who authorized which action. The survey figures point to a gap between ambition and preparedness that will likely define adoption speed as much as model capability does.

Why it matters: Product and platform teams building for industrial customers should expect demand for audit trails, permissioning and human-in-the-loop approval flows alongside raw model performance, since the stated near-term pattern is advisory AI under supervision. The cited reliability and efficiency results at SCG Chemicals and Nestlé give buyers concrete benchmarks to test vendor claims against. The EU AI Act's high-risk classification for certain safety components in critical infrastructure means oversight features may become a compliance requirement, not a nice-to-have — an inference from the regulatory detail rather than a stated conclusion.

Read at MIT Technology Review · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt