Proven oil reserves are increasingly being estimated by AI. Audit rules existed even before that.

Moving from decline curves to models

The traditional method is decay curve analysis. Engineers plot the well drop, fit a curve, and project it into the future. It dates back to the early 1900s and its main benefit is that anyone can see your work. Its weakness is that it cannot handle complex reservoirs well. It cannot account for multiphase flow, interflowing wells, or irregular behavior of shale.

Machine learning has stepped into that gap. Petroleum engineers have spent recent years through the Society of Petroleum Engineers documenting how models built on or in place of decreasing curves improve predictive accuracy, process more data, and discover patterns that standard equations miss. As for accuracy, the case is solid.

The tradeoff is visibility. The decreasing curve shows why. Many machine learning models don’t. The engineers who build them know this, and their own research is now driving towards explainable AI and physically informed models that improve interpretability and limit overfitting. Even if the model is not fully descriptive, it can still be used for internal planning. Once that production is reported as an asset on the books of a publicly traded company, it becomes even more difficult to defend.

What the rules require

Management standards predate this technology by many years. SEC Regulation SX, Rule 4-10(a)(22) — taken almost verbatim from 2026 filings from drillers including MACH Natural Resources and directs readers to that rule for the complete definition — defines proven reserves as quantities that “can be estimated with reasonable certainty to be economically producible” using deterministic or probabilistic methods. Both named methods are forms of human-trackable engineering. This rule does not mention machine learning.

That leaves an opening, and that opening is carrying most of the weight here. The SEC allows for “reliable technology,” which it defines as “a group of one or more technologies (including computational methods) that have been tested in the field to provide reasonably certain results that are consistent and reproducible.” Consistency and reproducibility. Models whose output can change when retrained on new data are uncomfortable with that requirement. A 2008 modernization (Release 33-8995) that rewrote these rules was intended to help businesses adopt better technology. This was not written with an unauditable model in mind.

There are also disclosure rules already in place. Under the SK rules, companies reporting additions to material reserves must provide a “brief summary of the technology” used, but that summary “may be general in nature” and does not need to reveal sensitive details. This carve-out made sense in 2009 to protect trade secrets. It is an open question whether a sentence like “We used a proprietary machine learning model” would give investors enough information to judge the reliability of the reserves.

One number from rough to market capitalization

The chain is short. Engineers feed well data into the model. This model generates production forecasts. This forecast becomes a reserved reserve that feeds into PV-10 and feeds into the company’s valuation. One link in that chain is that an outside company is supposed to test that number and ask if the company is pretty sure.

These outside companies, Ryder Scott, Netherland Sewell, and DeGolyer & McNaughton, are the names listed in the preliminary report at the back of the annual return. Their approval turns internal estimates into numbers that investors can rely on. They will argue that they are not rubber stamps and that their certification is evidence-based and method-agnostic. That is, we test results against well data and our own judgment, rather than which software generated the results. If the model’s predictions match production history and logs, the software behind it does not change its verdict, according to its logic.

This defense is valid, and it deserves to be stated. There is also one practical difference between the model and the curve. You can rerun the engineer decline analysis and arrive at the same answer. Reproducing an opaque model is one thing, explaining why one input is weighted more than another is another. The certification standards and industry SPE-PRMS framework that these companies follow were created for deterministic and probabilistic work. They do not consider the reproducibility or explainability of the model. Auditors are not cutting corners. Their yardsticks were made for different kinds of estimation.

who takes the risk

Exposure is specific. Shareholders own numbers that are partially populated by tools that cannot be investigated. Banks that make reserve-based loans (reserve-based loans that change size every six months based on the borrower’s barrel reserves) are extending credit based on projections that are difficult to confirm. Preliminary auditors carry legal and reputational weight every time they certify an estimate. And the SEC will enforce the 2008 standards against the 2026 tools.

The agency has worked in this area before. Its staff regularly sends out comment letters asking companies to explain the “credible technology” behind the addition of material reserves and justify keeping proven undeveloped reserves on the books beyond the five-year limit that allows undeveloped reserves to be forced off the books absent certain circumstances. If AI begins to compress drilling schedules based on its clock, or if the model’s role becomes large enough to impact reserve estimates, those letters clearly ask: No major lawsuits have yet been landed.

what to see

Three markers indicate whether the gap is closing. Changes to how auditors treat model-derived forecasts will first surface during the spring reserve reporting cycle, when operators file third-party filings. The SEC’s comment letter is second. Note that staff are testing how clearly companies disclose AI in their reserve methodologies. The third is the preliminary audit standard itself. The SPE-PRMS framework and certification rules are where professionals need to incorporate explainability into their work if they decide it is necessary.

Increasingly, the numbers that underpin oil companies’ values ​​are produced by tools the rules did not anticipate and approved by auditors working to standards older than those tools. The gap between the two is something to track.

conclusion

Read the reserve number as an estimate and look at the methodology behind it to find out what that method is. Disclosure of “technology used” in a company’s annual report reveals whether the company relies on “trusted technology” or a proprietary model. The exhibition of third-party reserve reports shows how external companies frame their certainty. For now, the task of asking the questions is left to investors, lenders, and regulators, as standards have not kept up with tools.