Oil and gas operators will need to up their generative AI game—because the buyer sitting across the M&A table has.

“If you're on the selling side of your asset … you want to make sure you're not leaving money on the table,” Kim Padeletti, head of energy data insights for Amazon Web Services, told NAPE attendees earlier this month.

“If your buyers are doing this, you should also be doing this on an asset that you're selling … and saying, ‘Am I leaving anything on the table? Is there a formation here that might be more valuable than I initially anticipated?’”

For both the buyer and seller, generative AI can also suggest a value that may be lower, but is correct.

“Some of your geoscientists might have had biases about formations—you laugh, but—they thought were going to be really prolific and they were not,” Padeletti said.

While AI can hallucinate, it can also be used as a guardrail against human hallucination. Third-party data may have “said it was a billion-barrel field, but it actually turned out to be a dry hole,” Padeletti said.

“On the buy side, there's an opportunity. But on the sell side, there's an opportunity to make sure you're not leaving money on the table too.”

A couple of operators recently used generative AI when acquiring assets, she added, “but I haven't seen them using generative AI as much as they probably could for that.”

Hallucination in distillation

Regarding hallucination, the proliferation of AI models has created advancements “not just on the compute side, but even just on the hallucination in the AI distillation,” said Andrew Muñoz, COO of energy evaluation software firm 4cast.

Distillation is the part where AI tries to make sense of its knowledge.

Muñoz said, “A lot of the upstream companies that exist today use very user-heavy software [by] geologists and engineers, and this type of software requires training and learning how to navigate it in context.”

An AI agent—everyday examples of these are Siri, Alexa and Google Assistant—can be taught.

“So even though a generative AI model might not be very good at understanding sparse subsurface data—things that we know are problems for machine learning traditionally in our industry—I think these agents are going to make a big impact and are going to kind of change the game there,” Muñoz said.

Pushpesh Sharma, director of product management for software firm AspenTech, said, “One thing I think about is that hallucination is just ‘unbounded creativity’ in a way.

“The model is thinking out of the box and it's giving you something that is not real,” Pushpesh said.

“But there is a good case to actually utilize those kinds of outputs where you are looking into design and any kind of creative input.”

‘Where should I go next?’

AI can be helpful in prospecting, said 4cast geophysicist Muñoz.

“If you're thinking about diving into a new area or a formation you're familiar with … you could prompt it to do a research project for you.”

The Simpson group’s bromide formation in southern Oklahoma could be an example, he said. Say “I want to understand more about the formation—what makes it work and what doesn't—so I can see if I want to invest in this prospect.”

With AI, a user can “sit there and search for 30 minutes and come up with a whole report for you and give you what it finds and how it interprets what's important about that formation.”

He added, “I think the possibilities for lowering the bar of understanding for getting into some of these more complicated prospects is starting to occur.”

Padeletti said, “You could say, ‘Hey, based on all the market intelligence data [and] my legacy assets, where should I go next? What should I buy? What should I divest?’

“And then you can prompt it on ‘Hey, my engineers have this skillset for this type of reservoir or an onshore asset that's adjacent to the Permian. I want to go after … whatever that is.’”

She added, “I would be doing that all day, every day if you're looking for prospects.”

‘Holy cow’

A senior engineer with a U.S. oil and gas producer asked the panelists what operators with small budgets can do. His company tried many AI models and kept coming back to “Holy cow, this is really expensive.

“I just have a feeling that some of the stuff that y'all are talking about might really only apply to people with big budgets,” he said.

“What would you say to that question?”

Muñoz said, “Just wait. You're going to see some free, if not very cheap, ways to run this type of technology.”

For example, he said, since the DeepSeek news on Jan. 27, “the cost of AI has gone down by 80%.” But Muñoz did add a caveat: “I'm not advocating using that model, by the way, because user beware: It's not sourced from the U.S.”

While costs are falling, the best intel models are constantly improving, he added.

“You have to basically look every hour to make sure you're up to date because things are changing so rapidly and the competition is ratcheting up,” Muñoz said.

Padeletti added, “I would say ‘wait,’ but maybe not too long because there's never going to be a perfect time.

“If you want a competitive advantage [and] wait too long, your competitors are moving on and making better decisions quicker.”

If you’re using a model while working on an M&A deal, though, “you don't want to throw that in an open model because some of the open models do use that to train data. … You’d want an encrypted model,” she warned.

“But I do see the cost going down and this becoming democratized in a way that benefits everybody.”

Owning the prompt

A geoscientist asked how to own the query. “How do we keep the copyright?”

Padeletti said, “What we're seeing, especially with customers that are dealing with subsurface data or they're dealing with highly confidential potential deals or M&A deals, is they're using this data in an encrypted tenant."

In this, data is pulled and the results are not pushed out. “It stays in your encrypted tenant.

“So if you're looking at a deal … in the Permian Basin and there's a deal team of 10 people, we would just get them an ‘encrypted instance’ [that] keeps it in your encrypted cloud tenant [and] doesn't ever push that back out.”

A geologist asked how to prevent an AI agent from learning from query.

“If it's an open model that says it trains on your data, yes [it can learn from your prompts],” Padeletti said.

Some models are closed. “They say, ‘We don't train on user data,’” she noted.

But encryption brings the best certainty.

“If you're working on any confidential data, even if you feel your prompt is confidential, like … ‘I'm looking for assets to buy in the Permian Basin.’ You would want to have an encrypted tenant.”

Comments