Dr. Ryan Ries here. This is our final installment of the "AI in Your Industry" series, and we saved a big one for last: energy, oil, and gas.
We also got requests to cover agriculture and education. Those will come as blog posts that I'll share next week.
Three quick announcements before we dive in:
- If there are any other industries you’re curious about their AI use cases, let me know and I’ll be sure to write either a blog or another issue of Matrix on it in the future.
- If you run a contact center or customer service operations, check out our upcoming webinar. It’s all about consolidating your tech stack with Amazon Connect.
- Alex G. on my team had a correction to last week’s Matrix image that was important to share…
And with that, let’s talk about energy, oil, and gas.
Traditional AI vs. Agentic AI
I've covered agentic AI in previous newsletters, but it's worth restating in this context because the energy sector makes the concept tangible in a way few other industries can.
I also think a lot of folks are still trying to understand AI vs. agentic AI, and that’s ok. That’s what my team and I are here to help you understand.
Traditional AI in oil and gas has been mainly predictive. Run a model. Get a forecast. Hand the forecast to an engineer. Wait for a decision.
Agentic AI skips the waiting. These systems observe live data, plan a response, execute it across connected systems, and adjust on the fly. Think of it less as a prediction engine and more as an operator that can chain together tasks without pausing for human approval on each step.
For drilling, that means an AI agent reading real time downhole pressure, benchmarking against safety protocols, and autonomously adjusting parameters before a pressure kick becomes a blowout.
For maintenance, this can look like an agent that flags a compressor anomaly, schedules the repair in the ERP, orders the parts, and alerts the nearest technician with instructions. The loop from detection to resolution closes without manual handoffs.
Now that we’ve set some parameters…
Use Cases I’ve Worked On
Oil and gas operations generate TONS of data. Sensors on rigs, pipelines, refineries, and offshore platforms stream temperature, pressure, vibration, and flow readings every second. Seismic surveys produce terabytes of subsurface imagery. Well logs pile up across decades of exploration history.
Most of that data sits in silos, underutilized.
Teams still toggle between disconnected systems, interpret dashboards manually, and make decisions that lag behind real conditions. The human bottleneck persists even as the data firehose intensifies.
That disconnect is precisely where AI, particularly agentic AI, is primed to play a huuuuuge role.
IDP: Buried Infrastructure and Land Records
We're working with a customer on a system to ingest and interpret decades of legacy data to determine where utility lines are buried, who owns the surrounding land, and what work has been performed historically.
This is EXACTLY the kind of unglamorous, high value problem that AI is uniquely suited to solve.
The records exist, scattered across filing cabinets and aging databases in inconsistent formats. AI models that can parse, reconcile, and surface that information on demand turn a liability into an asset.
NLP Chatbot
We partnered with an energy software company that provides subsurface data platforms for oil and gas exploration, geothermal energy, and CCUS (carbon capture, utilization, and storage). Their platform connects subsurface data and helps geoscientists visualize and analyze it, similar to traditional BI tools but specialized for the industry.
The challenge: users needed intimate knowledge of the data structure and platform mechanics to extract value. That's a high bar for adoption.
We built a GenAI powered chat feature that lets users ask questions in plain language. Queries like "What is the highest porosity recorded for Well A01?" or "Which wells have a grain density of 2.65?" return answers directly, no need to understand the underlying data architecture or build custom visualizations. We took this from proof of concept through to an MVP deployed to production for a limited user base.
GenBI
In a second engagement with the same customer, we tackled a different angle.
Their existing dashboard tools weren't extracting the insight buried in the data. We used Amazon QuickSight Q, which leverages GenAI to drive natural language queries, summarize outcomes, and build dynamic dashboards and data stories. Same types of questions ("What is the average permeability between 2100 and 2200m for Well A07?"), but now surfaced through an intuitive interface that dramatically lowers the barrier to insight.
Both projects highlight a pattern I see repeating across energy: the data already exists, often in massive quantities. The bottleneck isn't collection, it's accessibility. Natural language interfaces and GenAI are dissolving that bottleneck fast.
Sometimes the ~sexiest~ use cases aren’t the ones you need first. Or the ones your business needs at all.
Most often, I find that companies see the most ROI on their operations with these very practical AI use cases.
Predictive Maintenance
Speaking of practical…
Predictive maintenance, flagging an asset for repair before it fails, is a mature AI use case at this point. Nothing too new here.
With genAI, when an anomaly surfaces in a wind turbine or a refinery pump, a "Maintenance Copilot" powered by an LLM synthesizes real time sensor data with technical manuals, historical repair logs, and safety protocols to generate a specific repair path and documentation for the technician.
In a previous life, I worked on AR wearables designed to assist with maintenance operations. The concept was the same: get the right information to the technician at the right moment. The difference now is that generative AI can synthesize context from dozens of data sources instantaneously and produce repair guidance that adapts to the specific situation. The shift from manual analysis of maintenance logs to AI generated repair guides reduces Mean Time to Repair and improves technician safety by anchoring every procedure in the most current, contextually relevant information available.
Gas Line Optimization
This is a really interesting one.
We built an optimization system for a gas distribution company that figured out when to start turbines to prepare for demand cycles, making sure there was enough gas in the lines before customers started drawing power. Standard optimization problem on the surface.
BUT THEN: As we modeled the system's behavior, we uncovered operational dynamics the company itself didn't fully understand.
The data told a story about their business that contradicted some of their core assumptions. That discovery drove them to restructure their pricing models to capture more of the market and compete more aggressively.
This is the part of AI consulting that, in my opinion, is the most fun.
The optimization itself delivered value, of course.
But the deeper win was the organizational clarity that emerged from treating the business as a system worth modeling rigorously.
If you’re coming from the newsletter to keep reading, here’s where you left off.
This is a really interesting one.
We built an optimization system for a gas distribution company that figured out when to start turbines to prepare for demand cycles, making sure there was enough gas in the lines before customers started drawing power. Standard optimization problem on the surface.
BUT THEN: As we modeled the system's behavior, we uncovered operational dynamics the company itself didn't fully understand.
The data told a story about their business that contradicted some of their core assumptions. That discovery drove them to restructure their pricing models to capture more of the market and compete more aggressively.
This is the part of AI consulting that, in my opinion, is the most fun.
The optimization itself delivered value, of course.
But the deeper win was the organizational clarity that emerged from treating the business as a system worth modeling rigorously.
Use Cases Worth Watching
Autonomous Drilling and Geosteering
Drilling is the single largest capital expenditure in the upstream sector, and it bleeds money when operations stall from hazards like stuck pipes or lost circulation.
In 2024, SLB and Equinor drilled a 2.6 kilometer section of a well where 99% of the operation ran in autonomous control mode.
The generative component involves real time trajectory optimization. As the drill bit moves through the subsurface, generative models analyze vibration and pressure data to simulate the most efficient path forward, adjusting trajectory in seconds to stay within the reservoir's sweet spot. This process, geosteering, produces more productive wells and faster drilling times.
BP's "Optimization Genie," deployed in the Gulf of Mexico, manages these complex variables at scale. BP has reported that its broader AI for efficiency thesis, which includes tools like the Genie, generated $1.6 billion in savings and made well planning 90% faster.
My team has worked on a project in this space too, helping a client build out optimization models for drilling operations. The core challenge was the same one that makes autonomous drilling such a big deal: too many variables changing too fast for human operators to process optimally.
Smart Grid Management and Synthetic Scenarios
In the power and utilities sector, the primary challenge is managing an increasingly complex and decentralized grid.
Distributed energy resources like residential solar panels, EV charging stations, and battery storage have turned the grid into a dynamic, bidirectional network that traditional SCADA systems struggle to handle.
Generative AI fits this problem because it can produce synthetic datasets that replicate rare and dangerous edge case events.
Now, it’s no secret that our infrastructure is aging. Large utilities have reported that AI driven smart grid optimization can improve overall efficiency by 18%, translating to annual savings between $500 million and $1.2 billion for a large utility through better load balancing, frequency regulation, and congestion management.
DON’T ignore these challenges
I'd be doing you a disservice if I painted this as all upside.
Data QUANTITY is in abundance in this industry.
Data QUALITY remains an obstacle.
Subsurface databases often contain hundreds of tables with ambiguous relationships and non intuitive naming conventions. Text to SQL systems hallucinate when the data architecture is messy, and in this industry, messy is the default.
System integration is another hurdle.
Many operations still run on legacy SCADA systems and older ERP platforms. Plugging agentic AI into that infrastructure requires APIs, middleware, and careful compatibility work.
And, we can’t forget about cybersecurity.
Oil and gas companies are already frequent targets for cyber attacks. Interconnected AI systems widen the attack surface. Insider risk grows too, as more employees interact with AI agents and potentially expose sensitive operational data through prompts.
And governance matters more than most companies want to admit.
An AI agent that autonomously halts drilling operations based on a maintenance prediction that turns out to be wrong generates serious financial consequences.
Human in the loop oversight isn't optional for these types of high stake decisions.
Companies need clear boundaries around where agents can act independently and where they need a human sign off.
My Thoughts
Market projections suggest the energy sector will spend $18.5 billion on AI by 2028, and 70% of energy companies plan to significantly expand their AI and GenAI initiatives.
The pilot phase is over. This is industrial scale deployment.
If you're in the energy, oil, and gas sector and thinking about where to start, here's my short list:
- Pick a data accessibility problem. Natural language query interfaces over existing data stores are the fastest path to demonstrable ROI.
- Start with predictive maintenance. The sensor data is already flowing. Build the agent that closes the loop between detection and resolution.
- Get your data architecture in order. Agentic AI is only as smart as the data it can interpret. Clean naming conventions, documented relationships, and structured views are prerequisites, not nice to haves.
- Define your governance boundaries before you deploy. Know where the agent can act autonomously and where it needs a human checkpoint.
Let me know if you want to chat more about this. I'm always interested in hearing what challenges you're running into with AI in your operations.
Until next time,
Ryan
Now, time for this week's AI-generated image and the prompt I used to create it.
Create an image of a solar powered version of me. I am getting stronger and stronger because I am absorbing all of the solar power. Please use my reference photo attached.
