Jim Chappell, Global Head of AI at AVEVA.
Coming from nuclear engineering and the nuclear Navy, I learned early on that in critical systems, one small mistake can become a serious problem, and industrial AI requires the same discipline. After an outage, a power plant operator unintentionally entered the wrong set point for extraction steam on a steam turbine. The plant continued operating within acceptable limits, so its conventional control system didn’t raise any alarms. This error could have gone unnoticed for months. It would have increased the plant’s carbon footprint and cost them millions.
AI often creates the most value when it solves a defined operating problem and delivers a measurable outcome. For the plant operator, predictive analytics detected the change in the operating pattern and pinpointed the problem before months of fuel and water were wasted.
Customer-Specific Intelligence Is Key
Although AI has improved its outputs at a much faster pace than anticipated, some AI models can hallucinate. They may not know how a specific production line behaves under changing supply chain or consumer demand conditions, or why experienced engineers make quick-turn decisions based on years of operational knowledge.
This is an issue for industries operating safety-critical environments that leverage this type of generative AI. For plant operators, responding with “the model told me” isn’t an acceptable explanation for an operating decision. Teams need to know what evidence produced a recommendation, who reviewed it and who remains accountable for acting on it.
The knowledge required to address these limitations already exists within the business. It resides in the standard operating procedures, operational data, lived engineering expertise, established workflows and the experience of the people running these systems every day. Combining that knowledge with LLMs, domain-specific models and trusted, real-time data creates customer-specific intelligence that reflects how an operation actually works.
Back in the 1990s and 2000s, I wrote a lot of C++ code for simulation and expert systems. Some of those systems are still running today because they were built to solve specific operating problems and incorporated the knowledge of the engineers who understood them. The technology has evolved, but the need to combine software solutions with operational context and domain expertise hasn’t.
Context Is The New Differentiator
Consider an unanticipated vibration detected on a turbine blade. Identifying the issue is only the first step in finding a solution. An operator needs to understand its cause, what may happen next, how much time remains before the condition becomes serious and whether the asset can safely operate until its next planned outage. That combination of real-time diagnosis, consequence and timing turns a critical alert into useful operational guidance to inform future responses.
Combined Intelligence
The most useful industrial AI systems often use different capabilities to handle different parts of a problem rather than relying on one model for each task. A language model may provide an intuitive way to interact with the system, while specialized AI, analytics and engineering logic perform the underlying analysis. This desire for contextually aware AI is coming to the forefront of industry, with recent data finding that 49% of industrial manufacturers have “AI use cases already delivering business value,” while “68% expect to be deploying AI at scale” by June 2027.
Consider the challenge of routing coolant piping through an existing 3D plant design. AI has introduced more accessible UX to engineering design, but it needs additional capabilities to carry out the complex task. An engineer can describe the task in natural language, but the system must account for existing I-beams and piping, maintain clearance for workers, control the total length of the pipe and minimize bends that increase head loss. The language model provides the interface, while other analytical and optimization capabilities determine the best route.
See The Issue First
Pressure to demonstrate ROI can lead organizations to begin with the technology rather than the operating need. The better starting point is a practical question: What processes or outcomes should improve if this works, and what tangible value would that improvement create? The answer should define a measurable operational result, connect it to a business value and make clear who is responsible for acting on it.
To start, focus on a narrow use case that operators understand and the organization can test, such as predictive maintenance, anomaly detection or machine vision. For example, nuclear-power plants have been using this type of predictive analytics to warn operators of developing conditions for nearly 20 years while leaving the final decision to people.
Next, before building a model, confirm that the necessary data exists, enough history has been retained and the information is reliable, accessible and trusted. If the data or infrastructure can’t support the objective, change the scope rather than relying on black boxes and magic. Once the system identifies a useful operating method, convert it into a repeatable and auditable process that the organization can test, trust and apply consistently.
At a chemical plant, a deep reinforcement learning system autonomously controlled a complete startup. It adjusted and optimized set points as operating conditions changed, improving efficiency and making the process more consistent. Unlike a foundation model that generates probabilistic outputs, the system was fully deterministic, repeatable and built to perform a specific industrial task.
These same standards should apply before any industrial AI project scales. Teams may not know in advance which operating decision will change or which data will prove to be the most valuable. Discovering those connections is part of AI’s power, but teams should still know who will act on the system’s output and how the result will be measured.
Conclusion
When I co-founded InStep Software, we largely focused on predictive analytics, and since then, the tool has advanced significantly, but the underlying lesson remains the same. Achieving success with any new technology depends on establishing a clear and measurable definition of what “success” means to you from the very beginning.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
