Build a Play‑Specific LLM: Practical AI for Field Decisions

Every conference I go to, somebody is on stage talking about AI like it’s going to solve everything. They show the slides. They use the buzzwords. They make it sound like you just flip a switch and the machines start making better decisions than your engineers. And every time, I look around the room and I can see the same expression on the faces of the operators in the audience: that sounds great, but what does it actually do for my wells?

That’s the question nobody on stage wants to answer with specifics. And it’s the question Robert Wichert answered on Wisdom at the Wellhead without any buzzwords at all.

What if AI in oil and gas didn’t start with the technology? What if it started with the data you already have and the decisions your engineers already make?

Start with the Rule That Never Changes

Robert started his answer on AI the way any experienced engineer would. He didn’t start with the model. He started with the data. As he put it: in engineering and every other place you’re using computers, it’s garbage in, garbage out. So first, you need to get some control over what you’re actually feeding the system.

I appreciated that more than he probably knows. I have spent a lot of my career at Total Stream working on exactly this problem. You can have the most powerful analytics tool in the world, but if the data going into it is messy, duplicated, or sitting in five different systems that don’t agree with each other, the output is worthless. Or worse, it’s wrong and you trust it anyway. I’ve said it before: it’s okay to be creative with the data you use to find better solutions, but it’s not okay to be creative with the structure in which that data hits. Because that structure is everything.

Robert gets that. And because Cougar built its data systems clean from the start, connected engineering to accounting to field automation before they ever operated a well, they’re in a position most companies aren’t. They know what’s going into the system because they designed the system to capture it right the first time.

Build Your Own Model. Feed It Your Own Data.

Here’s where Robert’s approach separates from the conference-stage version of AI. Cougar is considering building their own language model. Not buying something off the shelf. Not plugging into a generic platform that was trained on everything on the internet. Building one that focuses on taking the information they’re generating as they get started and feeding it into a model that learns their specific operations.

And the data they can feed it is substantial. Through Total Stream’s link to the US production database, Cougar can pull decline curves and estimated recoveries from any play they’re looking at. They can feed in the types of drilling fluids used on offset wells. The types of completion fluids. The production histories. All of it specific to the play they’re working in.

Robert was clear that initially, this is going to be very play-specific. He’s not trying to build a model that knows everything about oil and gas. He’s building one that knows everything about the specific plays Cougar is operating in. That’s the difference between a tool that sounds impressive and a tool that actually helps your engineer make a better decision on a Tuesday morning.

Running Scenarios, Not Replacing Engineers

Robert described what AI does for Cougar in the most practical terms I’ve heard anyone use. It’s running through different scenarios rapidly and coming up with an optimal solution that then the engineering team can say, yeah, this makes sense.

Read that again. The AI proposes. The engineer decides. That’s exactly how it should work. I don’t know a single petroleum engineer worth his salt who would trust a machine to make the final call on a completion design or a lift change without looking at it himself. And I wouldn’t want to work with one who would. The value isn’t in replacing the engineer’s judgment. It’s in giving the engineer more options to judge, faster, with better data behind each one.

Think about what that looks like in practice. Your engineer is looking at a completion design for a new well in a play where you’ve already got production data from twenty offsets. Instead of pulling up each one individually, comparing decline curves by hand, and trying to figure out which completion style performed best in which part of the formation, the model does that analysis in seconds and presents the engineer with the best options. The engineer still picks. The engineer still applies judgment. But the homework that used to take a week now takes a few minutes.

That’s not science fiction. That’s just good use of the data you’re already collecting.

Why This Only Works if You Did the Hard Part First

Here’s what I think most people miss about AI in oil and gas. Everyone wants to talk about the model. Nobody wants to talk about the years of data work that have to happen before the model is worth anything. You can’t feed a language model data from five different systems that don’t agree with each other and expect it to give you a reliable answer. You can’t train it on production numbers that were re-keyed by three different people into two different spreadsheets and an accounting package that doesn’t match either one.

The reason Robert can even think about building a play-specific AI is because he built the data foundation first. That’s the part that takes discipline. That’s the part that isn’t exciting at conferences. And that’s the part that makes everything else possible.

I’ll be honest, I get a little frustrated when I see companies spending money on AI tools before they’ve solved their basic data problems. It’s like putting a navigation system in a car that doesn’t have an engine. The technology looks great on the dashboard, but you’re not going anywhere. The engine is the data. It has always been the data. AI just makes the engine more powerful, but only if the engine was built right to begin with.

Final Thought

The AI that’s going to matter most in oil and gas won’t be the one that knows the most. It’ll be the one that knows your wells. Your play. Your fluids. Your decline curves. Your completion history. Built narrow, built specific, fed with clean data that you trust because you designed the system that captured it.

That’s not the version of AI you hear about at conferences. But it’s the version that will actually make your engineer’s Tuesday morning better. And in this business, that’s what counts.

Robert Wichert isn’t chasing AI hype. He’s building a play-specific model on top of a data foundation that was designed to be trustworthy from day one. Hear how he’s thinking about it on Wisdom at the Wellhead.