Oil companies are shifting reserve estimates from human engineering to machine-learning models.
getty
Permian Resources’ latest quarterly report, filed this spring, names “artificial intelligence and its application in our industry” as a risk to the business. The same filing repeats a line that has run in oil-company reports for years: reserve engineering is “a process of estimating underground accumulations of oil and natural gas that cannot be measured in an exact way.” A new tool flagged as a hazard, a few pages from an admission that the number it helps produce was never exact.
To be clear about what that filing does and does not say: it’s a standard risk-factor disclosure, the kind that now shows up in the company’s press releases too. It is not Permian saying it books reserves with AI. Its year-end 2025 reserves were prepared by Netherland Sewell & Associates, an outside engineering firm. The story here is broader than one company. Across the industry, the math under the reserve number is changing, and the rules that turn that number into a stock price have not.
Proved reserves are what an oil company is mostly worth on paper. They drive PV-10 (the standardized present value of the oil in the ground), the impairment tests that can force a writedown, and a five-year clock on “proved undeveloped” reserves — wells a company has committed to drill. Change the reserve figure and the valuation moves with it. The forecast feeding that figure used to come from a chart an engineer drew by hand. More and more, it comes from a model.
The shift from decline curves to models
The traditional method is decline-curve analysis. An engineer plots a well’s falling output, fits a curve, and projects it forward. It dates to the early 1900s, and its main advantage is that anyone can check the work. Its weakness is that it handles complex reservoirs poorly — it can’t account for multiphase flow, wells draining into each other, or the irregular behavior of shale.
Machine learning has stepped into that gap. Petroleum engineers have spent recent years documenting, through the Society of Petroleum Engineers, how models built on top of or in place of the decline curve improve forecast accuracy, processing more data and finding patterns the standard equations miss. On accuracy, the case is solid.
The trade-off is visibility. A decline curve shows its reasoning. Many machine-learning models do not. The engineers building them know it, which is why their own research now pushes toward explainable AI and physics-informed models that improve interpretability and limit overfitting. A model you can’t fully explain is workable for internal planning. It’s a harder thing to defend once its output becomes a reported asset on a public company’s books.
What the rules require
The governing standard predates this technology by years. SEC Regulation S-X, Rule 4-10(a)(22) — quoted nearly verbatim in 2026 filings from drillers including MACH Natural Resources, which directs readers to that rule for the full definition — defines a proved reserve as a quantity that “can be estimated with reasonable certainty to be economically producible,” using deterministic or probabilistic methods. Both named methods are forms of engineering a person can trace. The rule doesn’t mention machine learning.
It does leave an opening, and that opening carries most of the weight here. The SEC permits “reliable technology,” which it defines as “a grouping of one or more technologies (including computational methods) that has been field tested to provide reasonably certain results with consistency and repeatability.” Consistency and repeatability. A model whose output can move when it’s retrained on new data sits uneasily against that requirement. The 2008 modernization that rewrote these rules — Release 33-8995 — was meant to let companies adopt better technology. It was not written with an un-auditable model in mind.
There’s also a disclosure rule already in force. Under Regulation S-K, a company reporting material reserve additions has to provide “a concise summary of the technology or technologies” it used, though that summary “may be general in nature” and need not reveal proprietary detail. That carve-out made sense for protecting trade secrets in 2009. Whether a line like “we used a proprietary machine-learning model” tells an investor enough to judge a reserve’s reliability is an open question.
One number, from the rock to the market cap
The chain is short. An engineer feeds well data into a model. The model produces a production forecast. That forecast becomes a booked reserve, which feeds the PV-10, which feeds the company’s valuation. At one link in that chain, an outside firm is supposed to test the number and ask whether the company is reasonably certain.
Those outside firms — Ryder Scott, Netherland Sewell, DeGolyer & MacNaughton — are the names in the reserve-report exhibits at the back of an annual filing. Their sign-off is what turns an internal estimate into a number investors can rely on. They aren’t rubber stamps, and their certifications are evidence-based and, they’d argue, method-agnostic: they test the result against the well data and their own judgment, not against which software produced it. If a model’s forecast matches the production history and the logs, by that logic the software behind it doesn’t change the verdict.
That defense holds up, and it deserves to be stated. It also runs into the one real difference between a model and a curve. You can re-run an engineer’s decline analysis and land on the same answer. Reproducing an opaque model — and explaining why it weighted one input over another — is a different job. The certification standards these firms follow, and the industry’s SPE-PRMS framework, were written for deterministic and probabilistic work. They don’t contemplate the reproducibility or explainability of a model. The auditors aren’t cutting corners. Their measuring stick was built for a different kind of estimate.
Who carries the risk
The exposure is concrete. Shareholders own a number partly set by a tool they can’t examine. Banks that lend against reserves — reserve-based lending, resized every six months on a borrower’s booked barrels — are extending credit on forecasts that are harder to check. The reserve auditors carry legal and reputational weight each time they certify an estimate. And the SEC enforces a 2008 standard against 2026 tools.
The agency has acted in this area before. Its staff regularly send comment letters asking companies to explain the “reliable technologies” behind material reserve additions and to justify keeping proved-undeveloped reserves on the books past the five-year limit, beyond which undeveloped reserves can be forced off the books absent specific circumstances. If AI starts compressing drilling schedules against that clock, or a model’s role grows large enough to matter to a reserve estimate, those letters have an obvious next question to ask. No marquee case has landed yet.
What to watch
Three markers will show whether the gap narrows. The spring reserve-report cycle, when operators file their third-party exhibits, is where any change in how auditors treat model-derived forecasts would surface first. SEC comment letters are the second — watch for staff testing how clearly companies disclose AI in their reserve methodology. The third is the reserve-auditing standards themselves: the SPE-PRMS framework and the certification rules are where the profession would have to write explainability into the job, if it decides the moment calls for it.
The number that anchors an oil company’s value is increasingly produced by a tool the rules never anticipated and signed off by auditors working to a standard older than the tool. The gap between the two is the thing to track.
The Bottom Line
Read the reserve number as an estimate with a method behind it, and find out what that method is. The “technologies used” disclosure in a company’s annual report tells you whether it leans on “reliable technology” or proprietary models; the third-party reserve-report exhibits show how the outside firm frames its certainty. The standard hasn’t caught up to the tooling, so for now the work of asking falls to the investor, the lender, and the regulator.
This article is for informational purposes only and is not investment advice.