A panel of industry professionals shared their perspectives and real-world experiences using artificial intelligence (AI), offering insights into the learning curve required to use the technology effectively. The candid discussion featured leaders from SPE, Oceaneering, and a private offshore engineering consultancy during the recent SPE Subsea Well Intervention Symposium in Galveston, Texas.
Among the key takeaways was that while AI is now a widely accessible technology available to organizations of virtually any size, determining where and how to deploy it remains a company-specific challenge.
Panelists also stressed that AI is not a risk-free technology. However, many of the associated risks can be mitigated through education, experience, and a disciplined approach to its use.
Below are some of the highlights from that discussion.
SPE’s AI Journey
Dana Otillio, the vice president of communications and member experience for SPE, shared details about the not-for-profit’s journey into AI since the advent of large language models (LLM) and generative AI. The big question facing SPE a few years back when these innovations debuted, she said, was: “How do you make nearly a century of petroleum engineering knowledge useful in the digital age and in the AI age specifically?”
SPE entered the AI era with nearly a century of industry knowledge housed in OnePetro, a repository shared with two dozen technical societies, with about half the content coming from SPE.
In late 2023, Aramco Americas partnered with SPE and software firm i2k Connect to develop an AI tool that would allow engineers to query the content using natural-language prompts. The challenge, Otillio said, was enabling AI access while protecting the intellectual property of both SPE and contributing authors.
By 2025, the tool, called EnRG-LLM, was ready for testing by SPE members and was soon made available for licensing to oil and gas companies. However, SPE learned that many companies had already either subscribed to foundational AI models or developed their own internal tools. As a result, many companies were more interested in accessing SPE's content in the OnePetro data set than in licensing a standalone AI service.
"While the EnRG-LLM remains available for licensing, we started realizing as quickly as it was launched that some companies were already moving in a different direction," said Otillio. "Even with the tremendous support of Aramco, which got us to the point of having a fine-tuned model, we learned that AI itself wasn't necessarily the differentiator. We thought of it like the movie, Field of Dreams—if you build it, they will come. And that's not really what happened."
Otillio said what ultimately proved to be the key differentiator was the knowledge contained within OnePetro, which "ties in with SPE's mission and values. AI doesn't really change that. Trusted, authoritative content is the value SPE creates."
SPE also realized that although the model had been gifted by Aramco, maintaining and updating an LLM of this nature would require significant ongoing investment. SPE is now working on a secondary solution with Silverchair, the vendor that maintains OnePetro.
"The advantage was that Silverchair had already ingested all of our content and metadata, which meant everything was close to launch," she said. "The development of this, we've been able to do it within 2 months' time."
As a result, SPE has made it possible for companies to acquire a license that securely connects to SPE's portion of the OnePetro corpus through an MCP or API and allows foundation models to interact with the underlying content. Otillio said the new initiative will roll out in the coming weeks with the proceeds funding a member-facing AI assistant to be launched by year-end.
Otillio also stressed to attendees that, like many companies in the oil and gas sector, SPE has had to climb the learning curve associated with AI and its costs, which she said "do not scale like traditional software" because of the data intensity involved.
One question can trigger significant data retrieval behind the scenes, while others may require far less resources. And as a data provider, it is difficult to predict those workloads and the costs associated with them.
"You have to budget for variability and based on the companies that we've spoken to, many of which you work for, you need to pilot whenever possible," she said, adding that pilot projects and proof-of-concept programs can help organizations better understand operating costs as usage scales.
Another piece of advice she offered was for companies to assess the quality of their data, which in many cases may be unstructured or lack sufficient context. She noted that SPE is fortunate in this regard because of its 2-decade digitization efforts. "Having high-quality metadata will absolutely help your AI tool retrieve the right information more accurately," she said.
Finally, Otillio said companies need to take governance seriously by establishing clear boundaries around what AI systems can access and how proprietary information is protected. "My guess is that those will be long-term conversations and that you may have to adjust your strategy over time."
‘Humans in Command’
These days nearly everyone is aware of AI and its growing capabilities, but as the panelists argued, it is not necessarily a plug-and-play technology. Implementation and adoption within large organizations still takes time and careful planning.
That was a central theme of Otillio's remarks, one that was echoed and expanded upon by Sean McCall, chief data officer and senior director of data strategy and architecture at subsea robotics specialist Oceaneering.
McCall told attendees that as the industry moves deeper into the AI era, he has never seen greater interest among business leaders in discussing data quality.
"I think it's because the data is closer to us, the data is felt, and you've got more access to seeing the connection between high-quality data and an output that's useful to you," he said.
However, the corporate IT executive stressed that the human element of AI remains less discussed but is every bit as important as the technology itself. One reason, he said, is that today's AI models are ultimately probabilistic systems.
“What that means is that it is a pattern-recognition engine, and it is going to produce different results each time, depending on who's asking, depending on what context was provided,” he said, adding that this is a departure from the deterministic software applications that corporations have relied on for decades. “[AI] doesn't understand, it doesn't possess human capabilities, and it can't judge in the way people can judge. What it can do is compare instructions that it's been given and produce probabilistic outputs from that information.”
This matters, he said, because AI is so reliant on its training data that humans remain critical to the optimal use of these new tools. That may not be the case for straightforward, rules-based tasks, such as selling a predetermined number of shares when a stock price threshold is reached. But in situations requiring interpretation or careful discernment, humans must remain in the loop, or as McCall put it, "humans in command."
He argued that only a limited number of activities today are suitable for full AI autonomy with humans serving merely as observers.
“The bottom line is that humans are accountable for what AI does—AI is not accountable for what AI does,” McCall said, adding that this is a major distinction. This point may be especially relevant in oil and gas, where safety is a top priority and project costs are often measured in the tens or hundreds of millions of dollars.
McCall encouraged everyone in the room to experiment with AI tools and evaluate how they fit into existing workflows, but he warned against deploying the technology simply to check a box.
At Oceaneering, he said, "we focus on an intentional AI culture," one that emphasizes providing the appropriate context for specific tasks so that AI tools perform as intended. He encouraged organizations to examine their processes in greater detail, ask better questions, and clearly define what successful outcomes look like before relying on AI-generated results.
One of McCall's final pieces of advice was somewhat unconventional. He said Oceaneering has abandoned its internal AI training program, instead relying on vendors to provide instruction or directing employees to resources on YouTube.
He acknowledged that it "sounds weird to be standing on stage in a professional setting saying that we count on YouTube, but it's the reality. There are tons of content creators that are trying to make a living by keeping up with the tools and that are right on top of the new releases," he said.
Putting AI to Work
In addition to the discussions about some of the pitfalls or cautionary notes of modern AI tools, speakers also highlighted practical opportunities to apply the technology to everyday workflows.
Brian Saucier, president of offshore engineering firm DeepMar Consulting, addressed the topic while bringing the discussion back to subsea intervention.
Saucier noted that one of the symposium's recurring themes, and a core objective of well intervention, is maximizing the value of existing assets, or as he put it, "squeeze more juice out of the orange." He believes AI can play an important role in achieving that goal, provided users understand what to look for.
"My message to you," he said, "is to look for something that you've done that's been challenging and spin it into the AI prompt to get that personal validation" in terms of what the AI might be capable of delivering.
Like the other speakers, Saucier cautioned against overreliance on AI and emphasized the need for balance and guardrails as organizations experiment with the technology.
"I try to make the distinction that AI is a partner, and not to think of it as a replacement," he said.
Saucier also encouraged professionals to understand the difference between generative AI, which can support a wide range of analytical and content-related tasks, and agentic AI, which can perform actions and workflows with appropriate guidance.
One opportunity he highlighted is the preparation of daily intervention reports, a repetitive task that could be well suited for AI. These reports rely on a steady flow of information, including operational updates, weather forecasts, and metocean conditions, all of which could be analyzed to help assess risks associated with specific activities.
He also pointed to logistics planning and more advanced engineering analysis. By allowing AI systems to ingest engineering drawings and technical documents, he said they could help evaluate equipment capabilities and operational limits.
“We have vessel motion, we have structural limits, we have control systems and their response time—there's a lot going on,” he said, adding that these are “multidimensional parameters that have to be considered in the well-specific operating guidelines. That's a perfect example of where an agentic AI solution could be harnessed.”
Saucier then referenced a project he undertook shortly after the debut of generative AI models about 3 years ago, when the technology was considerably less mature than it is today. Using data from a decade-old offshore project, he asked the model to perform a fatigue-life assessment on a set of five wells.
“And lo and behold, it gave me a data validation of this model that our fatigue life that was very close to what we actually had run,” he said.
He shared the example to demonstrate how engineers can test and validate AI tools with relatively low risk before applying them to operational work. Saucier also offered a vision of a future in which fully vetted AI agents operate on engineers' desktops, continuously updated and monitored by dedicated personnel.
Before that future can be realized, he said individuals and organizations must first determine where AI can deliver meaningful value.
"I think generative AI is here. It's valuable, it's working, but we need to hone our skills in command prompt engineering to make sure that we give it all the details," he said. "I know that's easy to say, because sometimes it's hard to look yourself in the mirror, or for your organization to, and really ask, ‘What is value to us? How much risk are we prepared to take, and where are we on that journey?’"
