Most energy companies have spent the last two years asking whether AI belongs in the reserves process. That question is already behind them. Machine learning sits inside the decline-curve software, the seismic interpretation stack, the type-curve libraries, and the analog-well selection tools that feed a modern reserves estimate. It arrived through vendor upgrades and license renewals rather than through a board decision, which is exactly why nobody logged it.
Here is the part that has not caught up. If AI now touches any step of how a US registrant arrives at its reserves numbers, the obligation to describe that process did not change — the process did. The disclosure rule sitting in front of every public operator was written to be technology-neutral, and technology-neutral means it already covers the model you did not know you were running.
Where does AI actually enter a reserves estimate?
Reserves estimation was never a single calculation. It is a chain of judgments: which analogs are comparable, which decline model fits, where the economic limit falls, what counts as reasonable certainty. Each of those links is a place where a statistical model can now propose an answer faster than a reservoir engineer can.
In practice, the entry points cluster in four places. Automated decline-curve fitting, where an algorithm selects the curve family and the fit parameters. Analog-well selection, where a similarity model proposes the peer set. Seismic and petrophysical interpretation, where machine learning classifies facies or picks horizons. And forecasting inputs — price decks, uptime, spacing assumptions — where a model output becomes a manual entry in someone’s spreadsheet with no record of where the number came from.
That last one is the quiet problem. Once a model output is retyped into a spreadsheet cell, it stops looking like a model output and starts looking like an engineer’s judgment. The provenance is gone, and it was gone before anyone considered whether it mattered.
What does the SEC already require you to disclose?
Regulation S-K, Item 1202, is explicit. Under paragraph (a)(7), a registrant must “disclose and describe the internal controls the registrant uses in its reserves estimation effort,” and must disclose “the qualifications of the technical person primarily responsible for overseeing the preparation of the reserves estimates.” Where a third party prepared or audited the estimates, paragraph (a)(8) requires that party’s report to be filed as an exhibit.
Read that with AI in the workflow and three things follow. The internal controls you describe have to be the controls you actually run, not the ones your narrative described several software versions ago. The named technical person stays accountable for estimates that a model now partly produces. And a third party who has quietly folded machine learning into their own workflow is inside a report you file.
No new rulemaking is required for this to bite. The exposure is not that regulators will invent an AI reserves rule; it is that the control description already on file may have drifted away from the process it claims to describe.
Why does the old control narrative stop working?
Traditional reserves controls were built around a human bottleneck. A qualified engineer made a judgment, a supervisor reviewed it, a committee approved it, and the paper trail was the review itself. The control was a person, and the evidence was that person’s signature.
A model breaks that structure in a specific way: it produces an answer without producing a rationale. When an engineer accepts a machine-proposed decline curve, the acceptance is a judgment, but nothing in the workflow captures it as one. The audit trail records the number, not the decision to trust the number. A year later, when someone asks how that estimate was reached, the honest answer is that the software suggested it and it looked reasonable — which is not a control, and does not read like one in a filing.
What should the reserves control narrative say now?
The fix is documentation discipline, not model sophistication. Six questions convert a stale narrative into a current one.
| Control question | The pre-AI answer | What the answer has to cover now |
|---|---|---|
| What tools produce the estimate? | Named software packages | Which of those packages contain machine learning or automated fitting, and which version introduced it |
| Who is the qualified technical person? | Name, credentials, tenure | The same, plus what that person reviews when a model proposes an input |
| How are inputs validated? | Engineering review of assumptions | How a model-derived input is distinguished from a manually derived one in the record |
| How is judgment documented? | Review sign-off | What was accepted, what was overridden, and the stated reason for each override |
| What do third parties use? | Their stated methodology | Whether their methodology now includes machine learning, asked directly and answered in writing |
| Who reports to the board? | Reserves committee summary | The same summary, plus where models influenced the estimate and how that was reviewed |
An operator who can answer all six has a defensible narrative. An operator who cannot has a filing that describes a process the company no longer runs.
Who is actually accountable for a model-assisted number?
Accountability is the hinge, and it is worth stating plainly: the model is not accountable. A named human is, and the rule says so. What changes with AI in the loop is not who signs, but what signing means. It now includes an assertion that the signer understood which parts of the estimate a model influenced and reviewed those parts on their merits.
That is a governance design problem rather than a technical one, and it is the same problem across every AI-influenced decision in an energy company — capital allocation, maintenance scheduling, procurement. The discipline that makes an AI-assisted decision defensible is consistent: know where the model entered, record what a human did with its output, and keep the record at decision time rather than reconstructing it under pressure. That is the organizing idea behind the NIST AI Risk Management Framework, whose Govern and Manage functions are built around documented, repeatable process rather than any particular tool. NIST’s Generative AI Profile extends the same logic to systems whose outputs are hardest to reproduce, which increasingly describes what ships inside commercial technical software. Building that decision record into the reserves workflow is the kind of structure ModalPoint works on with industrial operators.
The question to take into your next reserves review
Ask your reserves lead one thing: for last year’s estimate, can we identify every place a model proposed an input, and show what a qualified person did with it?
If the answer is yes, the disclosure narrative may only need updating to match. If the answer is that nobody has ever asked the question that way, that is the finding — and it is better found in a leadership meeting than in a comment letter, a diligence request, or a deposition.
The reserves number is the most consequential figure a public energy company publishes. It deserves a control story that describes how it is actually produced.
Matthew Bertram is Chief Marketing Officer of OGGN, President of ModalPoint, and CEO of EWR Digital. He works on AI visibility and AI decision governance with industrial and energy organizations.