iTranslated by AI
re:Invent 2025: Driving Energy Innovation with AWS AI at Baker Hughes
Introduction
By transcribing various international presentations into Japanese articles, we aim to make hidden, high-quality information more accessible. This time, we are covering the following presentation for this project!
For transcription articles from re:Invent 2025, please check the information summarized in this Spreadsheet.
📖 re:Invent 2025: AWS re:Invent 2025 - How Baker Hughes is Driving Energy Innovation with AWS AI (AIM347)
In this video, Cheo Alvarez from Baker Hughes explains the utilization of AI technology in the energy sector through a partnership with AWS. With energy demand expected to grow by 165%, and the oil and gas industry generating 15 petabytes of data from drilling rigs, the company is aiming to realize a digital oilfield utilizing agentic AI. Through a digital platform called Leucipa, they are achieving data contextualization, workflow automation, and the integration of physics-based models and machine learning, deploying specialized agents such as reservoir monitoring agents and pump optimization agents. Key factors emphasized include verification processes with a human in the loop, ensuring explainability, and adaptability to the different tools and processes of each customer. By leveraging technologies such as AWS Energy Agents, they are promoting safe and effective AI adoption in heavy industry.
Note: This article was automatically generated while maintaining the content of the original presentation as much as possible. Please be aware that there may be typos or incorrect information.
Main Content
Challenges Facing the Energy Industry and Long-term Efforts Toward Digital Oilfields
Well, everyone, thank you very much for attending our first session today. My name is Cheo Alvarez. I am with Baker Hughes, and today I will be talking about how we are advancing energy with AI technology through our partnership with AWS.
I have a few points I'd like to discuss. First, I want to set the stage regarding the challenges we are facing in the energy sector. I believe this is primarily driven by the demand for LLMs, agentic AI, machine learning, and so on. Next, I will talk about how we are addressing this, specifically the use and application of agentic AI, and the lessons we have learned along the way. We have been working on this for a long time. We have been bringing digital projects to market for years and years. Of course, agentic AI is very new, but we are weaving these into every layer of our stack to truly accelerate our ability—and our customers' ability—to deploy these types of solutions at scale. Then, a bit about the internal mechanics, which I can talk about in more detail after the session. And finally, I will discuss our future outlook and where we are heading.
Now, for years, the energy industry has pursued the same vision of a "digital oilfield." However, global energy demand has only accelerated in recent years. This is a statistic that surprises me every time I see it: the projection of 165% growth in energy demand is truly astonishing, primarily driven by hyperscaling needs and the energy demands of hyperscalers. In the oil and gas industry, we have been on a journey to capture data and truly understand the value of that data at every stage of operations. From the reservoir, to drilling and completing the well, how we measure that well, its continuous operation, and then downstream. As you can see, we are generating vast amounts of information. 15 petabytes of data are generated from drilling rigs. There are 1,800 drilling rigs operating around the world at any given time, and again, this is only part of the initial process of drilling a well.
In other words, data is being created everywhere. Extracting a signal from that noise has always been, and will continue to be, a challenge; but the demand—and the opportunity here—is immense. As those who have followed the energy industry know, when I was a child, there was always talk that the world was running out of oil. Actually, technically, that's not true. The world is running out of "cheap" oil. And the way we have addressed that challenge over and over again was by introducing new physical processes like horizontal drilling, hydraulic fracturing, and deepwater exploration and production. We will continue to face it with technology. However, now, more than ever, that technology will inherently be more digital. That is why I am giving this talk today.
This is a quote from McKinsey, and I think everyone here will resonate with it: how transformative agentic AI will be for every industry on the planet. This will transform the operations of every company, and in my opinion, make work much more interesting. Because I can delegate my repetitive tasks to agents, allowing me to focus on more creative, higher-value work. Where we see this impacting—and this is a very high-level discussion—is that by 2028, 33% of enterprise software will include some form of agentic AI, which again drives energy demand. By 2028, 15% of operational decisions will be made by agentic AI. Over one billion agents will be created and deployed at varying levels of quality. And what Baker Hughes sees here is that the quality of our agents will be a true competitive advantage for Baker Hughes and our ability to support energy customers in the future. Because, "garbage in, garbage out." We have a lot of deep expertise that we need to apply to and bring into our agents.
So, let's talk about where we have been, particularly in the technology sector I am involved in. It is called Leucipa. Let me talk a bit about Leucipa. It is about the digital technology program I am participating in, because we have been on this journey for years.
Our current Agentic AI journey is the same one we have been on since the first day of the digital oilfield. For us, the first step in the process was enabling access to data. Garbage in, garbage out. It is something I have heard from everyone I have spoken to here; getting access to data and contextualizing that data. Then we add workflows and automate those workflows. That is, we work with our customers to understand their processes, the data selection and technology choices they have made, the modeling tools they use, and how to apply those to their business processes. I am trying to give a very high-level explanation without using too much industry jargon, but I think this is very similar for all industries.
We want to introduce automation, but again, to ensure that we can eventually deliver some form of results while respecting the customer's technology choices and business processes. We have been doing this since the early 2000s. Initially with physics-based classical techniques. Then we layered on machine learning and AI-type models. And now, we are adding Agentic AI and weaving it throughout this entire stack.
Implementing Agentic AI in the Leucipa Platform and its Architecture
Real-world application of AI, Agentic AI, or in other words, Agentic Enterprise Software. We are taking an approach very similar to yours. We are introducing orchestration, centralized orchestration. I think what truly differentiates us is the quality of the individual agents. Orchestration agents are largely generic, but what sets us apart is the architecture, how we approach them, and the quality of the individual agents themselves. Basically, we have taken Leucipa functions, wrapped them as individual agents, and contextualized them.
In the oil and gas industry, there are reservoirs two miles underground that spread out for miles in every direction. There is no way to actually measure what is happening underground. Therefore, for us, everything is an approximation of the truth. We have the wellbore vicinity that we can measure. There is some mathematical method to approximate what happens in wells far away from the reservoir. We need agents that have some awareness of prediction accuracy. We are always measuring the same values but approaching them in different ways. We need contextualized agents that understand uncertainty and the quality of the models used to arrive at those approximations. This is the area where we are really spending a lot of time and focus.
We have foundational capabilities that we call Leucipa. These are our physics-based models and the classical correlations we apply. The oil industry has truly evolved in line with the adoption of computing in general. It started with very simple plotting functions; as CPUs became more widespread, we gained the ability to iteratively calculate and simulate reservoir simulations, nodal analysis, and various other things. Machine learning, the large-scale adoption of machine learning throughout our industry, and now, bringing all of this together and packaging it as Agentic AI, as you can see here.
This is presented as a chatbot because it is the most user-friendly interface we can present to end users. However, what is happening behind the scenes is that "Lucy"—Leucipa, or Lucy—invokes all the various agents of Leucipa that we provide and contextualizes them, once again improving the signal-to-noise ratio, presenting the right context, calling the right agent at the right time, and surfacing the right recommendations to the end user. As you can see here, everything is stacked on top of one another.
In other words, it is about Leucipa's ability to connect to data, contextualize that data, and cooperate with the customer's stack, regardless of the form or tools they have chosen, whether it be on-premises, SaaS, or leaning heavily into cloud technologies. We truly need to be prepared to meet customers where they are. That means on-premises or in the cloud, and all of this is intended to avoid meaningless recommendations that only add more noise instead of increasing the signals we should be extracting.
I think this is the point where we start to differentiate. It is in our architectural approach.
And it is a future-oriented architecture rooted in our history. The gravity of information is shown on the left as the energy operators—our customers—and on the right as Baker Hughes and our digital platform, Leucipa. Energy operators have a wealth of information from all the different tools and applications they use. They have business processes to execute and specific operations to which they apply them. All the fields operated by customers are generally different, in the sense that they are highly heterogeneous.
When we go to a customer and show them a demo of our software, they say, "I see, it's test validation." You might think this is a very common process standardized across the industry. But when we get into the details of how the customer does it, they say, "That's nice, but that's not how we do it here." What we immediately learned from that was that we needed flexibility in every aspect of this. What this architecture provides us is precisely that flexibility. The flexibility to very quickly adapt our agents to the data sources chosen by the customer, the flexibility for the tools they choose, and by setting up some agent protocols in between, we allow each side to surface what the other is providing.
And we invoke orchestration agents on both sides, not just to deploy agents and tools, but to incorporate a human-in-the-loop so that we can verify the responses when those agents and tools interact. We truly view agents as partners, as teammates. Not as a replacement for human capital, but truly as an extension of human capital.
Let's zoom in and take a concrete look at what one of these use cases actually looks like. A customer may select a specific type of reservoir monitoring agent. It might come from a physics-based tool or a machine learning-based tool, but that reservoir monitoring agent might come from simple monitoring of tags coming out of the wellhead. That reservoir monitoring agent needs to be tuned to specific conditions. In this case, we are discussing a rising oil-water contact. This is because if water enters the well too much, it can permanently damage the well, resulting in producing far more water than desired over the lifetime of that well, effectively causing the loss of that well.
This reservoir monitoring agent is operating on the customer side or the Baker Hughes side, but it is continuously observing in the background. When it detects an event beginning to occur that is known to be influenceable by another agent, it calls on a Baker Hughes specialist agent. Baker Hughes specializes in the manufacture and design of pumps, especially electric submersible pumps. These are pumps located two miles underground, at the bottom of the well. We use various simulators to predict the optimal pump frequency or speed, or liquid production rate, to produce through that pump without damaging the reservoir. Usually, this results in the deceleration or slowing down of the pump.
The Baker Hughes agent model makes a recommendation on how much it needs to be reduced. And here is where it becomes very important for our industry. When applying AI, especially agents, to heavy industry, we must heavily rely on verified recommendations with a human in the loop, by experts who understand the physical principles behind what those agents are recommending. An agent might make a recommendation, but the quality of that recommendation is entirely determined by the models it invokes and the quality of the data input into it. Our humans are in the loop to verify the health checks before sending them back to the customer. This is a service we provide.
We send that recommendation back to the customer. And when we pass it over, the customer puts their own humans in the loop—meaning production engineers verify it. Baker Hughes cannot act on its own based on that recommendation in every single case. Some customers want us to provide that service, while others want to keep their own humans in the loop. Generally, it is verified by the customer's production engineer, and that person approves the recommendation from our experts. And then they push that recommendation to an edge device.
Nowadays, since we have instrumentation installed at every wellhead, we can push those recommendations to edge devices without actually needing to dispatch someone to the field to make physical changes. And the cycle just repeats. That is, the reservoir monitoring agent returns to observation to verify if the pump change actually produced the expected impact. And the process continues. Because we have so many different agents and calculators, and ways to approximate truth and quantify uncertainty, this is truly a generalized process. This is one specific process, but it is very scalable and can be generalized to much of what we do.
Lessons Learned from AI Adoption and Future Outlook
How do we break down complex systems? We break them down into small parts. We have many agents solving each of these small parts. And orchestration—meaning optimization of optimization or higher-level orchestration agents—is really just making everything work better together. So, what we have learned through this—and it is very much the same lesson from when we started, back when AI, ML, and agent tools were not yet widely adopted—is that data quality is everything. Even if you do this with physics models, calculations, or algebraic correlations, data quality is the foundation you stand on.
User experience, and the explainability of the recommendations and opportunities we surface as we increasingly introduce things perceived as black boxes, is absolutely critical. We have users who are highly technical experts who have spent years understanding first-principle physics, such as how this pump functions, how this reservoir behaves, and how to solve these types of solutions numerically and analytically. Therefore, we need to be able to explain how we arrived at a specific recommendation at a very, very low level. This was something said from the very first day we tested the concept of agents with people. For me, explainability is everything.
About adaptability. We use the term "heterogeneity" to describe reservoirs, but organizations are just as heterogeneous. Every customer chooses different tools, different data stacks, different approaches, and different business processes. Every field is also different. We need to be prepared to meet customers wherever they are in their digital transformation journey. And we are working with global customers, not just in North American onshore, but around the world, and each of you is at a different stage of the digitization process.
And finally, governance and cost. How to manage them. In HS&E—or heavy industry—if incorrect recommendations are made based on bad data or outdated models, it can lead to catastrophic results. Setting up guardrails based on Baker Hughes' expertise is one of the areas where we truly believe we can provide the most value to our customers.
What makes this work? Many of the familiar AWS technologies we know and love, which were announced today and heard about in the presentation. There is an open-source project I want to highlight, called "Energy Agents," which was released with AWS sponsorship. You can check it out on GitHub. This shows many of the same technologies we are applying internally. We are building on top of these. Of course, data management and the guardrails I mentioned earlier.
Regarding the future outlook, what we are aiming for is rooted in past experience: connecting to data, contextualizing data, working with customers to document workflows, automating manual processes supported by these agents, and, of course, truly driving some impactful results using them. Until now, we have been doing this manually, working with implementation engineers and developers to create these types of tools. Today, I am specifically talking about the Leucipa application for customers, and this is helping. We see that "agentic" is truly helping us adopt the technology we have been developing for years at scale. However, given the level of tuning and customization required to truly introduce this into assets, it has been much more difficult to deploy at scale.
Therefore, we believe that weaving "agentic" into how we onboard customers, then adjusting workflows to adapt to specific customer needs, and creating and handing them over to the customer so that they themselves can use them, will truly be a game changer. And we believe it is what will enable us to digitize this operation at scale. We think this is something truly made possible by "agentic," and perhaps something that can only be realized by "agentic" for the first time in history.
With that, we have just reached our time, perfectly on schedule. Thank you very much. I am happy to answer any questions if you have them, so I will be around here somewhere, thank you very much.
Note: This article is automatically generated using Amazon Bedrock, while maintaining the information of the original video as much as possible.














Discussion