AI is no longer theoretical in refining; it’s transforming how engineers troubleshoot, optimize and support operations.
Refiners like Marathon Petroleum, Chevron Phillips Chemical (CPChem) and Big West Oil are deploying AI across a growing range of applications, resulting in faster decisions, deeper insights and more effective use of engineering talent.
"It’s a tool to leverage," said Holly Fitch, process engineer with Marathon’s Galveston Bay refinery. "We’ve seen it as basically like a work accelerator, not necessarily that it’s replacing anything, but it is speeding up the process by which we’re doing it."
Fitch said Marathon has rolled out a proprietary large language model — essentially a company-specific version of ChatGPT — that makes it easier to find and verify technical standards. The internal tool has been well received, partly because it surfaces direct references to corporate documents, allowing engineers to fact-check its responses. "It’s not threatening anybody’s role," Fitch noted. "It’s simply enhancing the way we access data."
Marathon, CPChem, Big West leverage AI to boost speed and reliability
Pictured from left are Tyler Harnos with Big West Oil, Holly Fitch with Marathon and Brent Railey with CPChem present during the 2025 AFPM Summit.
Tyler Harnos, a process engineer with Big West Oil’s Salt Lake City refinery, shared how optimization models developed with AI partner Mubit are driving day-to-day improvements. These models can detect process variables every 10 seconds — far faster than the previous approach, which relied on engineers making daily manual adjustments.
"Having their [operators’] knowledge and being part of it has really helped develop these models and push them to a great spot," Harnos said.
Panel moderator Brent Railey, chief data analytics officer with CPChem, agreed that generative AI is having the most traction as a "thought companion" — a kind of assistant that synthesizes information and accelerates decision making, especially for desk workers.
Still, Railey emphasized that not all AI is suited to high-consequence operations. While generative models excel at summarizing documents or answering queries, they can also "hallucinate" answers with unwarranted confidence. "A word sequence predictor is not going to be great for optimization," he warned. "They might help you generate code you can pass to a real optimizer, but they’re not going to do the optimization."
In contrast, more mature forms of predictive AI, such as machine learning models that finetune equipment performance, are already delivering ROI. Harnos described how Big West’s implementation has changed how his team thinks about projects and bottlenecks. "It’s just been helpful. It allows you to maximize your time throughout the day and keep looking forward."
The pace of change is creating its own challenges. With new technologies emerging every few months, panelists acknowledged a tension between innovation and stability. Railey called it the "FOMO vs. tech debt" dilemma — companies want to move fast but can’t afford to chase hype without a clear ROI.
Fitch said the bigger barrier isn’t resistance but perceived complexity. "Nobody’s saying, ‘I won’t use AI.’" But the barrier to entry still seems high, she said, adding that there are lots of programs available that aren’t being used because people feel like it takes a lot of input to get started.
Harnos said early skepticism among operators can turn into buy-in if AI is positioned as a tool that supports their experience. "They see it as a way to reduce the workforce or take their job. What I’ve found helpful is letting them know they’re the reason the models are so good and that changes things," Harnos added.
The panel, held during the AFPM Summit, concluded with a simple message: don’t wait for AI to be perfect. Start small, measure value and scale what works.
