Industrial AI can make a wrong call in more ways than producing an unhelpful answer. When software is connected to equipment, power systems or hazardous sites, a mistake can reach beyond the screen. AVEVA’s Arti Garg says the challenge is to gain the benefits of more capable AI while keeping people responsible for decisions that affect safety and operations.
Industrial AI is moving closer to the machinery
AI has been part of industrial software for decades, often in specialized roles such as predictive analytics. What is changing, Garg says, is the spread of foundation models, physical AI and agentic AI. These systems can support more complex tasks, from software-based automation to robots that gather information in physical environments.
That shift is drawing more interest from industrial companies. Garg cited one study suggesting AI adoption in the sector had risen by nearly 78% over two years. The source did not identify the study, so the figure is best read as context for the momentum she describes, not as a standalone measure of the whole industry.
The difference between industrial AI and a tool that only handles digital tasks is not just a matter of setting. Industrial systems can affect equipment and infrastructure used to supply power or extract resources, often in places where people work around hazards. Newer AI models can also be harder to predict and explain than more narrowly designed systems, and their behavior may change as they are tuned or learn from new data.
That makes greater capability a mixed bargain. A system may be able to do more, but the consequences of letting it act without clear limits may also be greater.
Better decisions start with connected data
Before AI can help diagnose a plant problem, an organization needs to bring together information that may live in different systems. That can mean equipment telemetry, maintenance and service logs, engineering records, and original design documents. In many industrial operations, connecting those records is a substantial part of the work.
Garg says newer tools, including graph databases and AI systems that can match information across data sets, may help operators find relevant context faster. Instead of manually searching several sources to understand a pump or mixer, an operator could use a tablet-based system to bring together its live readings, maintenance history and documentation while diagnosing a problem.
Robots could extend that approach beyond the operator’s desk. A robot moving through a hazardous area could collect information for an operator without requiring a person to enter the space. The AI may help gather or interpret evidence, while a human remains available to assess what it means and decide what to do.
This is a less theatrical part of industrial AI than a robot acting on its own, but it is also a practical prerequisite. If the relevant information is incomplete or scattered, a more capable model does not magically make the underlying records coherent.
Autonomy needs operating limits
AVEVA describes its responsible AI approach in terms of security, efficiency, human safety and oversight. Garg’s central point is that AI should support people rather than replace them in critical decision loops. In practice, that means deciding which tasks a system may handle independently and where a person must remain accountable.
One example is adjusting equipment set points. AVEVA has worked with customers on systems that use operational data and simulations to recommend changes based on conditions in a plant. Garg says some pilots have also tested moving from a recommendation to automatic adjustment.
That step changes the nature of the decision. A suggestion can be reviewed before it affects the equipment. An automatic adjustment acts first, which is why operators may set limits on how far the system can move a set point or restrict automation to specific operating areas. Those boundaries give the system room to work without handing it unrestricted control.
There is no single set of guardrails that fits every application, Garg says. The limits depend on the system and the risks involved, and organizations will need to learn from deployment. The point is not to pretend that a human supervisor can anticipate every edge case. It is to define what the AI is allowed to change, make its role clear, and keep a person responsible for understanding the risk.
The distinction matters because industrial automation is not simply a question of whether a model can complete a task. It is also a question of what happens when it is wrong, what happens when conditions change, and who has authority to intervene.
AI may help carry expertise forward
The industrial workforce is changing too. Garg said that nearly half of the sector’s workers are expected to retire within the next five years, a shift that could take substantial experience out of operations. AI might help preserve and share some of that knowledge, including by applying lessons learned at one site to another.
That possibility is not the same as turning every experienced worker’s judgment into a reliable model. It does suggest a role for AI in making operational knowledge easier to find and use, particularly for newer employees who need to understand how equipment and processes fit together.
Garg also expects AI-assisted coding to let people with deep industrial knowledge build applications even if they do not have a software-development background. She predicted that this could become more common within 18 months. The appeal is straightforward: people closest to operational problems may be able to create tools for needs that software teams did not know to ask about.
But new tools do not remove the need to change how work is organized. Garg argues that organizations often struggle less with the existence of AI than with adapting business processes to its strengths and limits. A system that makes recommendations, for example, needs a workflow in which someone can review them, understand their basis and act on them appropriately.
Sustainability runs in both directions
AI could help manage power grids as renewable generation expands. More distributed and intermittent sources, including rooftop solar, make it harder to track supply, demand and power quality across a system. AVEVA has been working with Idaho National Laboratory on an AI grid-resilience project focused on helping operators understand grid conditions, identify anomalous behavior and respond earlier.
There is a second question, though: the environmental cost of running AI itself. Garg chairs an IEEE working group developing a standard methodology to measure that impact. The group’s P7100 work covers electricity and energy use, resources, water and carbon. Garg says the effort is intended to establish a common method that could be used for reporting and potentially oversight or regulation.
Without an agreed method, individual claims about the footprint of a particular AI query can be difficult to compare. The measurement question is not an aside to deployment. It affects whether organizations can judge the cost of a system against the work it performs.
Garg also argues for choosing a model suited to its task. Larger models require more computing, which is likely to draw on more resources. A purpose-built model may use less, and, if its design fits the application, may also behave more predictably. That does not make model selection simple, but it gives organizations a practical place to start: use only the capability the task calls for.
A customer example, with caveats
AVEVA describes SCG Chemicals, a petrochemical company in Thailand, as a customer working to build an AI-driven reliability platform. The work began with bringing operational and engineering data together, then adding AVEVA Predictive Analytics, which uses proprietary models to detect anomalies and identify possible operational risks.
The stated goal is to turn unplanned downtime into planned downtime and improve plant reliability. Garg said SCG Chemicals is targeting about 99% plant reliability and that early pilots produced an ROI of nearly 9x.
Those figures are notable, but they are not a universal forecast for industrial AI. The reliability figure is a target, and the ROI figure refers to early pilot results as described by AVEVA. The example also reinforces a less glamorous point: the AI layer rests on work to organize the data beneath it.
The test is how the system changes the work
Robots and drones could take on inspection or data collection in hazardous and remote locations. AI tools could help operators diagnose equipment problems, support grid management, or turn domain knowledge into new software. Those are plausible uses described by Garg, not a guarantee that automation will improve every operation.
For organizations, the practical work is to match the system to a task, set limits that reflect the physical risks, and give workers a clear role in reviewing or supervising its actions. They also need to decide how to measure the system’s resource use and whether existing processes can make use of what it provides.
Industrial AI is not just software with a harder hat. It is software whose decisions can meet the physical world, where the right answer includes knowing when not to act.