For this episode of the GraphRAG Curator Podcast, I had the chance to hear at length from Frédéric Verhelst—a veteran data transformation executive and hands-on semantic graph data modeler with a PhD in Applied Physics. Frédéric’s career has spanned decades at major energy corporations including TotalEnergies and Shell. Now at VIKING Life-Saving Equipment, Frédéric shares how data management has evolved and improved to become AI ready.

Frédéric’s perspective is both conservative and pragmatic, considering the high stakes world of offshore drilling—where a single stuck drill bit costs $200,000 a day.  Consider all the safety and other risk concerns of offshore oil and gas in the context of agentic AI, and you start to realize how important reliability and accuracy become.

Digital twins are widely used in oil and gas. Frédéric explains the transition the industry has made from simple 3D visualizations to semantic models that can actually reason. He argues that AI agents should be treated like contractors: they need a strict scope of work, clear guardrails, and a solid knowledge graph foundation to ground agent answers and enable more trustworthy action.

Are the same techniques from Oil & Gas used elsewhere?

Frédéric’s answer is a nuanced “Yes, with a shift in focus.” While he has moved from energy to maritime safety equipment, the fundamental data management and modeling techniques consistent:

Technique	Oil & Gas Application	Maritime/Service Application
Semantic Layers	Connecting well logs to regulatory filings.	Connecting customer personas to field service efficiency.
Digital Twins	Visualizing built platforms and real-time sensor data.	Managing functional locations and hyperlinked instrumentation.
Uncertainty Management	Probing subsurface seismic waves for drill or no drill decisions.	Creating guardrails for AI agents to prevent legal or operational surprises.

His core philosophy–thorough, precise modeling– hasn’t changed. In the oil and gas industry, that precision prevents a $100 million dry well; in maritime services, it prevents an AI agent from making costly, unauthorized promises to a client.

Frédéric blogs at https://www.linkedin.com/in/fredericverhelst. You can view the recording and full transcript below.

Edited Interview Transcript

Alan Morrison: Welcome everybody. It’s Alan Morrison with another episode of the GraphRAG Curator Podcast. And today I’ve got a special guest, Frédéric Verhelst of TotalEnergies. He’s somebody who’s got decades of background in digital transformation, including digital twins. He’s knowledgeable about hybrid kinds of AI with some touch points in ontologies, etc. You can see his whole portfolio at fredericverhelst.com and he also blogs on LinkedIn.

00:10:17

Alan Morrison: So, Frédéric, welcome today.

Frédéric Verhelst: Thank you.

Alan Morrison: Tell us about your background and how you got to where you are.

Frédéric Verhelst: Okay, first a small correction. I’m no longer with TotalEnergies. I’m now with VIKING Life-Saving Equipment. If you see on the back here, we are one of the main players in the world on life rafts, lifeboats, and safety equipment in the maritime sector. I’m based in Esbjerg, Denmark. My background—I’m Belgian, and I have an engineering degree first, and then afterwards I took a PhD in applied physics/geophysics, exploration of oil and gas, and that’s how I rolled into the oil and gas industry. I’ve been working with contractors, with small startups working on technologies; worked with TotalEnergies, worked closely with Shell, Chevron, and so on. So I have quite a broad range of experiences and background.

00:11:49

Alan Morrison: So, perhaps we could hear a brief history of where the energy industry has been in terms of its data journey and where you think it’s headed today. You’ve been in transformation initiatives over the years. How are those differing from the kinds of initiatives that we’re seeing today?

Frédéric Verhelst: So, for me, there are a few key points. To start with, beginning in the 2000s, the Society of Petroleum Engineers, an international association for the oil and gas industry, started with forums where the idea was to think six to ten years ahead and to get different viewpoints and perspectives into the mix. In 2003, they had a first one on what I call “Digital Energy,” where it was called “Smart Fields.” It was actually the start of a series of those types of events. In 2008, there was another event in Dubrovnik where we were talking about petabytes of data for asset management.

00:13:23

Frédéric Verhelst: That also gives you an idea of what volumes of data there are in the oil and gas industry. Geophysics, especially, deals with terabytes of data. But on the other side, there is another key point: there was a pretty small oil and gas field offshore Norway with Norsk Hydro at the time, now a part of Equinor, where in 2005/2006 they were seeing that the platform as it was operating would not be viable anymore in 2008. So they needed to do something very drastically. Using the fiber cable that was available, they looked at all the tasks being done offshore and saw if those could be done onshore. By moving tasks from offshore to onshore, they reduced from a normal manning of 41 persons to 21. That shifted the economic cutoff so much that now, 20 years later, the Brage field is still producing and generating money.

00:15:01

Frédéric Verhelst: So that’s for me one of the first examples—perhaps there are others, but from what I know—of where digital technology and data really made a big impact on the viability and profitability of a field. Those learnings were used when Statoil and Norsk Hydro merged into Equinor. They used that as a footprint where they now have 100 people in an integrated operations center supporting all their fields across the world. They have established that it’s really earning them billions per year in US dollars. It’s a very good case.

Alan Morrison: Talk about the challenges that oil and gas companies face in an exploration mode.

00:16:35

Alan Morrison: You’ve got this huge amount of data that is desirable to collect and analyze. Then at the sites, you have limited bandwidth and constraints on resources when monitoring exploration and drilling efforts. Can you paint the picture of the challenges you’ve been involved with trying to address?

Frédéric Verhelst: Actually, the exploration game is very challenging because there is a huge amount of uncertainty. You’re probing the subsurface using seismic waves with a limited bandwidth and all kinds of hypotheses when you want to analyze and interpret those. That leads to decisions of “shall we drill or not?” where an exploration well can typically cost up to $100 million per well. You have to make sure you take all available data, use analogies from other fields, and use old information lying around.

00:18:23

Frédéric Verhelst: All that information needs to be collected and interpreted in a good way. That’s where data management originates in the oil and gas industry—from the subsurface part. They had a real need and knew the consequences of a dry well. Once you go for development, building a platform costs around $10 billion, where uncertainty is still a big part of it. On the other side, for the drilling part, a lot of effort was put in the early 2000s into getting real-time data from drilling rigs to centers onshore where people could analyze and assist.

00:19:50

Frédéric Verhelst: Making sure a drill bit is not stuck in the well is vital because that costs $100,000 to $200,000 per day. Data standards were important to ensure the stream of data was continuous, having real-time data centers just to make sure the data was flowing 24/7.

Alan Morrison: Can you compare and contrast the structured data that seems prominent in the industry with the need for knowledge management and less structured information?

Frédéric Verhelst: Yes. In exploration, there is a lot of unstructured data available—reports and such. I saw a trend ten years ago where everything was dumped in a data lake, which people then called a “data swamp.” Without structure, everything is there, but you need to find the needle in the haystack.

00:21:52

Frédéric Verhelst: In the Danish affiliate where I was heading the Data Office, we had a semantic layer and a virtualization layer connected to all data sources. We were contextualizing them in one model and exposing that as a semantic layer to end users or software. For example, the concept of a “well” was in the production database (structured), in drilling reports (unstructured), in reservoir models (proprietary databases), and in regulatory filings. The semantic layer ensured that if you called for a well, you would get everything you needed.

00:23:19

Frédéric Verhelst: There was one truth out of that. It was also connected to the BI systems running on top.

Alan Morrison: Does graph technology and semantic metadata help with the integration effort? Are you finding use cases for graph databases in addition to standard relational and document databases?

Frédéric Verhelst: I think the best example is a Digital Twin. There are different degrees of Digital Twins. Generation one is mostly visualization with pictures or point clouds (laser scanning) for “as-built” information. Generation two integrates real-time data. Generation three is where you make sure you have a semantic model where you can start reasoning with engines on top of your Digital Twin.

00:25:15

Frédéric Verhelst: These are emerging. With the focus on Generative AI, people will see the benefit more. Our Digital Twin uses the CFIHOS (Capital Facilities Information Handover Specification) standard, which is now being semantized. There is a similar effort for the subsurface called OSDU—a data platform and knowledge capture system for the exploration domain. It will be a base for engines and multiple Digital Twins to use that semantic foundation.

Alan Morrison: Let’s unpack this evolution of Digital Twins. It seems important that if you have simulation and modeling, meaningful information allows AI to do more accurately. Is leadership behind the foundation needed to create these interactive Digital Twins?

Frédéric Verhelst: It’s a good question. I was in an affiliate, not the head office. Some leaders have a vision and understand why things need to happen. However, in a company like TotalEnergies, affiliates in the North Sea have a different maturity than other places. What we think is insufficient might be a very high step upward for other affiliates.

00:28:47

Frédéric Verhelst: If you have a centralized approach, you might get a Digital Twin that needs to cater to all different countries, which is a consideration. I clearly see the generations. Some affiliates were happy just to bring unstructured documents into a visual context. We use “functional locations” as key points—pumps and pieces of kit on the platform—and then hyperlinked P&IDs (piping and instrumentation diagrams).

00:30:18

Frédéric Verhelst: You can go on the drawing, go upstream or downstream, and find the relevant pieces of kit, then visualize data sheets with all parameters. For some, that is a big step. For others, you want that data in a structured database (e.g., Structured Query Language or SQL), then the next step is a graph database to implement those connections explicitly. The next step is a Knowledge Graph with an ontology on top. That is the journey from descriptive to diagnostic, predictive, and prescriptive.

00:31:39

Frédéric Verhelst: There were leaders in the digital program with that vision, but things take time to build momentum. TotalEnergies has a CEO who is an engineer, which helps. He understands what this can do and is a driver for these difficult technologies.

Alan Morrison: It’s important to have active leadership. Could you compare the energy industry to others regarding maturity? You’ve worked on assessments.

Frédéric Verhelst: I see similarities with Pharma. They need to connect data from many sources and have a strong imprint of uncertainty. Decisions have big consequences.

00:34:43

Frédéric Verhelst: From the outside, I have the impression semantic technologies are more embedded in Pharma than in oil and gas. Oil and gas was considered very conservative in 2000, and I don’t think things have changed a lot. If that sense of urgency at the Brage field hadn’t been there, perhaps nothing would have happened.

Alan Morrison: Talk about the industry you’re in now and how that differs.

00:36:14

Frédéric Verhelst: This is a much more low-margin company. One of the pros is that if you come up with a good idea that has a substantial effect, it might be easier to get it through here than in oil and gas. I’m now working on the digitalization of customer and field services. The efficiency of field technicians has a big impact—using them more efficiently and planning better.

Alan Morrison: The modeling challenge is different—understanding humans and their needs on the customer side.Are you using the same techniques you used in oil and gas?

00:37:51

Frédéric Verhelst: Not yet. We are using Dynamics 365 by Microsoft, and they have announced they will make labeled property graphs available in their system.

00:39:36

Frédéric Verhelst: It’s a step forward, even if it’s not a full semantic ontology yet.

Alan Morrison: Are companies really thinking about Knowledge Representation to make AI accurate?

Frédéric Verhelst: You hear managers asking for AI agents because they read about it and think it’s a golden solution.

00:40:59

Frédéric Verhelst: I don’t think all managers understand the consequences. I wrote a post this morning that Agentic AI should be seen as a contractor, not an employee. You need a specific scope of work, guardrails, escalation triggers, and monitoring. You need to formalize those so the machine can stick to it. That’s where the ontology comes in—it needs to be 100% precise so a “key client” is defined clearly, otherwise you get surprises.

Alan Morrison: Unexpected consequences.

Frédéric Verhelst: Yes. Air Canada was one of the first confronted with consequences. Lawyers have had issues with made-up cases. Generative AI looks so damn convincing. It is making things up, but doing it in such a convincing way.

Alan Morrison: What would be the most beneficial thing companies could do starting in 2026?

00:44:23

Frédéric Verhelst: They should start setting up a team that collects knowledge and formalizes that into a knowledge graph/ontology. There is a lot of modeling by data architects that can be reused. There is business process data and descriptions of offshore facilities. You could look at the skills for different roles.

00:45:59

Frédéric Verhelst: Gathering different information makes one plus one much more than two. It makes things like agentic AI possible. It needs to be phased, starting small but having a mandate high enough in the organization to go across departments.

Alan Morrison: Makes complete sense.

Frédéric Verhelst: To you and me, but perhaps for others not.

Alan Morrison: That’s the challenge—building alliances across the organization.

00:47:29

Alan Morrison: Well, Frédéric Verhelst, thank you so much for taking the time today. Very illuminating.

Frédéric Verhelst: Thank you.

Alan Morrison: That was great. I appreciate the insights.

Transcription ended

---