Data is not the problem

For most operators, the issue isn’t a lack of data. It’s fragmentation. Information is often spread across different systems, from control platforms and maintenance tools to spreadsheets, or teams may be working in isolation.

In some cases, the most important knowledge still sits with individuals rather than in the data itself. Engineers rarely have a complete, real-time view of what’s happening across an asset, which is why having more data doesn’t necessarily lead to better decisions.

The trust problem

Even when the data challenge is addressed, another issue is making sure that engineers can understand what an AI system is recommending and, more importantly, why.

This is where many current AI approaches fall short. Large language models (LLMs) are good at generating responses, but the responses are generally not based on real physics. LLMs work by recognizing patterns in data, which means they can sound convincing even when they’re wrong.

This creates a fundamental problem in oil and gas engineering environments, where decisions will have a direct safety implication. It’s not enough for an AI recommendation to just sound reasonable. It has to be verifiable.

Many people are already skeptical of AI tools, thanks to experiences with clunky earlier generations that failed to deliver on their promises. So it’s important that the output is not only right, but also explainable.

Moving beyond the “black box”

The recent shift towards “physics-informed” AI models is an important development. Rather than relying purely on statistical and predictive patterns, physics-based models rely on real physical and chemical principles that govern the world around us.

Conventional AI models predict what’s likely to happen based on historical data, but real engineering requires an understanding of constraints, the relationships between different variables in a system, and things like cause and effect. Physics-informed AI brings these approaches together and allows models to generate outputs that are grounded in reality.

Not only do these types of models consider real-world conditions, but their outputs can be tested against real simulations, validated against known behavior, and traced back to their underlying assumptions.

This gives engineers the confidence to assess a recommendation with credibility before acting on it. It also means that, instead of acting as a standalone black box decision-maker, the AI tool becomes part of an engineering workflow that supports human judgement rather than replacing it.

From theory to application

Unsurprisingly, partnerships between engineering specialists and AI developers are becoming increasingly important. AI technology has advanced rapidly, but the real challenge in oil and gas is applying it in a way that engineers can trust.

KBR has been working with their partner Applied Computing, combining AI models with decades of engineering expertise to develop a physics-informed AI. Rather than relying on pattern recognition or prediction, the approach grounds recommendations in engineering principles that can be validated against known operating conditions.

One of the first applications of this partnership is KBR INSITE, an AI-enabled platform used in production environments. The system combines operational plant data with an understanding of process chemistry and engineering constraints to identify anomalies and highlight opportunities for optimization.

Keeping humans in the loop

Even with more advanced AI models, human oversight is still essential. AI can highlight patterns and test different scenarios, but it can’t take responsibility for decisions.

It’s possible to look at this in three ways. At the most basic level, AI can be used to monitor conditions and flag anomalies. Next, it can be used in a more advisory role, to suggest actions that engineers can evaluate. Finally, but only in limited cases, AI may be used to automatically optimize specific processes within tightly defined boundaries.

Even then, experienced engineers are still needed to interpret the results, assess the risk and make the final decision.

A shift in skills and mindset

As AI becomes more embedded across industries, those industries are seeing a change in required skills. Engineers need to understand how these models work, where they fall short and how to validate the output.

This requires a wider cultural shift. Organizations will need to upskill their workforce, and that won’t happen overnight. Some will move faster than others. But adoption will ultimately come down to trust in high-risk environments.

AI has clear potential in oil and gas, particularly in areas such as asset integrity, optimization and operational planning. But its impact will not come from automation alone. It will come from improving the way decisions are made — through better visibility, validation and trust.