Earlier in my career, I was part of a team working with a major industrial operator on what was then considered an ambitious concept: a plant with no permanent workforce.

The idea was straightforward. If you wanted to run a facility such as an offshore platform without people on-site, you had to rethink it from the ground up. In practice, that meant innovations like modular equipment that could be lifted out by helicopter and swapped without a maintenance crew, or hardware designed not for repair but for replacement.

See also: Industrial AI has transitioned to the ‘application phase’

And layered across all of it were systems capable of monitoring, diagnosing, and responding to operational events without a human in the loop.

Sound familiar? That was around 2012 and 2013. We did not use the word autonomous then—the industry term at the time was unmanned—but the ambition was identical to what we are discussing today.

This is not to suggest that autonomous operations are old news, but to make a different point: the idea has always been real.

What has changed is the technology available to pursue it and, perhaps more importantly, our understanding of how to pursue it sensibly. Autonomy does not arrive all at once across an entire plant. It emerges first in specific operational layers, particularly optimization and process control.

The difference between hype and reality

Walk into most industrial plants today, and you will still primarily see people in control of operations: experienced operators working with distributed control systems designed, in most cases, to manage versions of the plant that existed years or decades ago.

But increasingly, AI-driven optimization systems are already working behind the scenes. The ambition to modernize is present, often urgently so. The reality is more complex.

That complexity has several sources, and energy systems themselves have changed profoundly. A refinery that once had a clear input-output relationship now faces challenges that would have been unthinkable a generation ago.

In any given moment, an operator may need to weigh whether it is more profitable to produce gasoline or to sell electricity back to the grid, because spot prices make that the better call right now.

See also: Why industrial AI pilots fail: 5 mistakes that kill projects

Renewable sources, storage systems, and grid interactions all add layers of complexity that the control logic of even a decade ago was never designed to handle.

The control logic that was fit for purpose in a simpler environment was not designed for this. Traditional industrial automation systems—DCS, SCADA, PLCs—are built to manage a plant in normal operating conditions.

They handle anticipated exceptions well. What they were not designed for is the kind of continuous, multi-variable optimization that modern operational complexity demands.

At the same time, the workforce is changing. Experienced engineers and operators who carry decades of process knowledge are retiring. The people replacing them expect digital tools, connected systems, and intelligent interfaces, and arrive without the deep institutional knowledge their predecessors accumulated over careers.

That knowledge gap is not a problem that can be solved by better recruitment alone. And running underneath all of this is an unrelenting financial pressure.

Plants need to do more with less. That means operating closer to optimal—which in turn means pushing processes toward boundaries that a cautious human operator, acting alone, would rightly avoid.

See also: With partnerships and software reveals, Siemens leadership plots a course for industrial AI

A human being will always leave a margin of safety. That is not timidity; it is good judgement. But it also means that the plant is never running at its true ceiling, and the opportunity cost of that conservatism over thousands of operational hours is substantial.

From operator support to autonomous optimization

Understanding where most plants are is the starting point for an honest conversation about what comes next. The near-term opportunity is not fully autonomous plants, but a mix of augmented operations and autonomous process optimization.

That does not mean autonomy is purely a future ambition. In many industrial environments today, technologies such as advanced process control and real-time optimization are already making autonomous decisions within tightly defined process boundaries.

These systems are not simply assisting operators; they are continuously optimizing the process itself. What remains non-autonomous is the broader operational layer: coordination across systems, exception handling, commercial trade-offs and strategic decision-making.

The goal is not to replace the operator, but to give the operator tools that are intelligent: Systems that break down the silos between data sources, correlate information in real time, and surface the right insight at the right moment.

Report: Manufacturers split on whether to prioritize industrial AI

An operator needs early warning when an abnormal situation develops. Not an alarm triggered when a threshold has already been breached, but an anomaly signal that alerts them to something deviating from normal before the deviation becomes a problem.

They need context: Has this pattern happened before and how was it handled? And they need the ability to test a response before committing to it, to run the scenario forward in a simulated environment and see what happens.

These are not science fiction capabilities. They are available now; combining them fundamentally changes the operator's role.

ABB's approach brings together these components: anomaly detection that identifies subtle process deviations early; knowledge extraction that retrieves relevant historical data and precedent from across the plant's operational record; and process prediction that allows operators to simulate the outcome of a proposed intervention before it’s applied.

The analogy I use with customers is that of an internet search engine. Before AI-assisted search, you entered a query and received a list of links. The work of correlation, deciding what was relevant, was yours. Now the system synthesizes for you.

It presents a conclusion, not a catalogue. We need to do the same for plant operators. The information they need exists in their systems—the gap is in how it reaches them, and in what state.

Making the numbers work, making the ROI case

That brings us to the challenge which derails more conversations about autonomous operations than any technical limitation: the business case. Any plant manager asked to commit capital to digital transformation will ask, reasonably, what the return looks like.

And that is where many discussions about autonomy falter, because the ROI on augmented operator capability—on making your people smarter, faster, more confident in their decisions—is difficult to quantify. What is the value of a problem that did not happen? What is the financial impact of a decision made better?

My approach is to anchor these investments to technologies where the financial case is already well understood.

See also: Why manufacturing is moving to edge-to-cloud architectures for resiliency

APC and real-time optimization solutions have decades of deployment data behind them and already represent an important form of operational autonomy in many plants today, continuously adjusting processes faster and more consistently than a human operator could achieve manually.

Payback periods are short, often below nine months, frequently around six, and in some cases as little as one month.

When these solutions are combined with augmented operator capabilities, organizations achieve two things simultaneously: a measurable financial return in the near term and more robust infrastructure—integrated data, intelligent systems, and confident operators that make the next step on the autonomy journey possible.

The parallel with sustainability is instructive. Telling a plant manager that autonomous operations will reduce their environmental footprint is true and important, but it rarely closes a business case on its own.

Telling them that the same investment will reduce energy consumption by a specific percentage and deliver a specific saving per year makes it a different conversation. The sustainability benefit is real; the financial benefit makes it actionable. The same logic applies throughout the autonomy journey.

The agentic layer

Looking further along that journey, the concept attracting serious attention, and where ABB is actively developing capability, is agentic AI. The term can sound abstract, but the underlying idea is simple. Most AI systems today are reactive; they analyze, recommend and surface information.

Agentic AI goes further. It acts. It coordinates across systems and workflows, taking on tasks that currently require a human to move between applications, pull together data from different sources and exercise judgement about what to do next.

Stories of AI adoption:

• • Wolfspeed all-in with 22 agents across key company teams
  • Wolfspeed uses ‘three-stage maturity model’ for agent deployment
  • How one semiconductor maker tamed AI in a trio of ways for real performance gains

Think of it as something acting on your behalf, moving between systems, pulling the relevant pieces together and driving toward a conclusion, rather than waiting for a human to do that work of coordination.

ABB is piloting this capability now, and while it remains early-stage, it represents a meaningful next step toward genuinely autonomous operations.

Imagining the plant of 2030

When I think of plants in the near to medium-term, they’re not unmanned. But they will be substantially different from what exists today. Optimization technology will be the norm rather than the exception. AI will be embedded across the operational layer, not as a bolt-on but as an integral part of how the plant runs.

In many cases, routine optimization decisions will happen autonomously in the background, with operators increasingly focused on supervision, exceptions and strategic coordination rather than direct process control.

See also: Why industrial AI requires a data ops foundation to scale

The control room will not look like a wall of alarms and DCS displays; it will be an environment where operators work with correlated, contextualized information and make decisions with intelligence behind them.

And plants will be connected to grids, to markets, and to the broader energy ecosystem, in ways that require exactly the kind of adaptive, intelligent control that the autonomy journey is building towards.

The companies best positioned for that future are not necessarily the ones investing most heavily in autonomous ambitions today.

They are the ones building the foundations: integrating their data, stabilizing their processes, developing their operators, and making incremental investments that deliver measurable returns at each step.