Pairing industrial AI autonomy and human control

Digital twins used to be rearview mirrors: models that simulated a plant and suggested what might happen. The pitch now is a "living" twin that keeps pace with real conditions and can actually act through control systems. That is a meaningful shift, because it moves AI from advising engineers to adjusting refineries, dryers, and grids.

Manufacturers want it. A 2026 Manufacturing Leadership Council survey cited in the piece found 66% using or planning to use agentic AI, and 40% expecting some factory autonomy by 2030. But a Deloitte survey from 2025 found 65% of executives ranked operational risk among their top concerns. Appetite and anxiety are running side by side.

We think this is one of the more grounded corners of the agentic AI story. Plants already run on sensors, control loops, and supervisory systems. Adding a model that makes small, constant adjustments is an evolution of that, not a leap into the unknown. The examples on the table are concrete: a Nestlé dryer tuning test that cut energy use by 16% while improving quality and yield, and SCG Chemicals reporting 99% plant reliability with a digital twin watching asset health.

What I like most is the honesty about the middle ground. AVEVA's product chief leans on the SCADA analogy: the S stands for supervisory, which means a human stays in the loop. Aramco shows both modes at once, with one refinery running closed-loop optimization under oversight and another where algorithms only recommend and operators decide. That is the right shape for the next few years. Autonomy where the task is narrow and repetitive; advice everywhere else.

Now the caution. This piece comes with a vendor's voice and vendor case studies, and the numbers are reported by the companies themselves. We would want independent audits before treating a ninefold return in six months as a norm. The bigger worry is governance lag. An AVEVA and IMD study found only 23% of organizations saw a significant change in decision-making and accountability across partners, while 43% saw little or none. Plants are wiring in agents faster than they are rewriting who answers when one gets it wrong.

If this scales without that rewrite, the failure mode is not a dramatic robot mishap. It is quiet drift: a model reinforcing patterns in its own data, operators approving less and noticing less, and a small error compounding across a connected value chain that used to be siloed. Integration that makes plants resilient also makes failures travel. We are for living twins. We are also for boring, explicit rules on what they may touch.

A digital twin is a software model of a physical asset or process. The newer versions connect design and engineering records to live operational data through a so-called digital thread, enabled by cheaper sensors and better software platforms. In Europe, the AI Act already treats certain AI safety components in critical infrastructure as high-risk, with human-oversight requirements attached.