Oil and gas has spent decades paying for better pictures of rock nobody can see, and it kept score.
By Matthew Bertram · Owner & President of ModalPoint, CEO of EWR Digital · October 2026
Westwood Global Energy's 2018 benchmark of the appraisal behind 70 producing fields found that 43 came in more than 10% off their sanction plan, and that 84% of the mismatches were rooted in the reservoir. Set beside IPA's finding that 75% of upstream projects missed their planned NPV, Amoco's exploration success rate with and without 3-D seismic, and the winner's curse that Capen, Clapp and Campbell described in 1971, the industry's own record shows what a better subsurface model is worth. The evidence on AI is thinner. OpenAI's GDPval measures the quality of what models produce, and I found no published study that follows AI-assisted decisions through to their outcomes.
In 2016, at the SPE Annual Technical Conference and Exhibition in Dubai, Neeraj Nandurdikar of Independent Project Analysis presented IPA's review of a decade of upstream project performance at the dinner panel the Project, Facilities, and Construction discipline holds every year. As the Journal of Petroleum Technology reported it, 75% of projects failed to achieve the NPV that was planned.
Three in four. These were not wildcats. A project only has a planned NPV after someone has approved it. The article names one contributing factor, and it is the thin early work of appraising the reservoir and designing the development before anyone commits the money, the stage the industry calls front-end loading.
The strongest case against spending more on appraisal sits inside the best study of appraisal I know. In February 2018, Westwood Global Energy published a benchmark of the appraisal programmes behind 70 producing fields. The write-up, by project geoscientist Pierrick Rouillard, compared what each field was expected to deliver at sanction with what it delivered once it came onstream. Of the 70, 43 landed more than 10% above or below the sanction plan on reserves or production.
Then comes the uncomfortable number. Fields appraised with more than five reservoir penetrations still deviated from plan 52% of the time. A CFO who signed for the sixth appraisal well, and the seventh, is entitled to ask what the money bought.
Westwood's own answer is careful: "The number of reservoir penetrations is important, but it is by no means the only contributing factor." Where the wells went and what they measured count too.
Here is what the extra wells did buy. Among fields with fewer than five reservoir penetrations, 70% did not perform as expected. Among fields with more than five, the figure was 52%.
Better data did not make anyone right. It made them less wrong, which sounds like a consolation prize until you look at the size of the misses.
Half the fields in the study had their proved plus probable reserves revised after production started. Where the revision went down, the average cut was 60%. Picture a development sanctioned on 100 million barrels that turns out, after the platform is built and paid for, to be a 40 million barrel field. Where the revision went up, the average increase was 68%, and that is not free either, because facilities sized for the field you expected put a ceiling on the field you got.
Of the fields that did not match expectations, 84% had problems rooted in the reservoir. The leading causes were wrong estimates of the hydrocarbons in place (49%), pressure support that fell short because the connection between producers and the water injectors or the water leg had been misdiagnosed (42%), and reservoir quality that differed from the forecast (35%).
Read that list the way a reservoir engineer would. Every item on it is a modelling error. The oil in place was whatever it was. The connectivity was whatever it was. What failed was the picture of the reservoir that the sanction decision rested on, and Westwood says so plainly: "Most of the subsurface risks identified could have been mitigated before the start of a project with more effective appraisal."
The cleanest comparison I have found sits in a fact sheet from a 1999 Department of Energy report on exploration technology. Amoco's exploration wells drilled on 3-D seismic between 1990 and 1995 succeeded 48% of the time. Its wells drilled without 3-D seismic succeeded 13% of the time.
Seismic does not change the rock. It changes the picture of the rock that a drilling decision is made against. In Amoco's portfolio, a better picture went with a hit rate of about one well in two instead of about one in eight.
This was not a controlled trial, and I would not present it as one. Companies shoot 3-D over prospects they already like, so part of that gap is selection. Hand half of it to selection and the 3-D wells still succeeded at more than twice the rate of the others.
The same case study has a detail I find more useful. Conventional surveying had turned up eight prospects, each carrying an economic success probability somewhere between 22 and 53 percent. The 3-D data cut the eight to two and correctly predicted that those two had a potential success rate of 60 percent. Six wells not drilled is what a better model looks like in a budget.
Notice what the Westwood numbers do not say. They do not say that well-appraised fields hit their plans. Most did not. They say well-appraised fields went off plan less often than thinly appraised ones. That is a relative claim, and relative is the only kind that matters when two companies want the same acreage.
The industry has a name for what happens when you forget that. In 1971, three Arco employees, Ed Capen, Robert Clapp and William Campbell, published "Competitive Bidding in High-Risk Situations" in the Journal of Petroleum Technology. They had been asked to keep the company from repeating a lease sale in which it won everything it bid on and then spent a couple of years in budget trouble. Their finding was that winning bids had been too high to leave a decent profit. The paper coined the term for it: the winner's curse.
The mechanism is short. When a dozen companies estimate the same block and bid on their estimates, the highest estimate wins, and among a dozen honest estimates the highest one is the one most likely to be too high. Shading every bid down is one response. A better model is the other, because the less error there is in your estimate, the more often a high number means a good block rather than a bad guess.
That is what I mean when I say whoever has better data and better models wins. Not that your model is right. That it is less wrong than the one across the table, and that you know roughly how wrong it is.
The pattern shows up outside the oil patch. In 2011, Erik Brynjolfsson, Lorin Hitt and Heekyung Kim surveyed 179 large public companies and found that firms which had adopted data-driven decision-making had output and productivity 5 to 6% higher than their other investments and IT use would predict. They tested whether the arrow ran backwards, with successful firms simply buying more data, and concluded that the effect does not appear to come from reverse causality.
Arie de Geus, then head of planning at Royal Dutch/Shell, made the competitive point in a 1988 Harvard Business Review article. He described planning as the process by which management teams "change their shared mental models of their company, their markets and their competitors," and warned that the process ran too slow "for a world in which the capacity to learn faster than competitors may be the only sustainable competitive advantage."
A reservoir model is a shared mental model with a file format. It is the thing a sanction committee argues with.
When I went looking for the same evidence on the AI side, I expected one of the frontier labs to have published it. The closest thing is OpenAI's GDPval, released in October 2025. It builds tasks from the real work of professionals averaging 14 years of experience, across 44 occupations, and reports that the best frontier models "are approaching industry experts in deliverable quality."
Deliverable quality is a fair thing to measure. It is not a decision measure. GDPval asks whether the model's memo is as good as an expert's memo. It does not ask whether the company that acted on the memo came out ahead. I have not found a published study from OpenAI, Google or Anthropic that follows AI-assisted decisions through to their outcomes the way Westwood followed 70 fields from sanction to production.
The subsurface is ahead here, and not because its models are better. It is ahead because it built the lookback. A sanction case records what the model said when the money was committed. Production records what happened. Westwood's study is those two records laid side by side, seventy times.
Most AI-influenced decisions have no sanction case. The prompt, the model version, the answer and the person who acted on it are rarely written down in one place, and a decision with no record of what the model said can never be scored. That record is what the DIG standard is built around, and what decision integrity captures at the moment a decision is made.
Ten years from now someone will try to run Westwood's study on the AI decisions being made this year. Whether they can depends on whether anyone wrote down what the model said.
Every figure above comes from the sources below. If your company has run a sanction-to-production lookback on its own fields, I would like to compare notes on what share landed off plan and which part of the model was wrong. Message me on LinkedIn.
How to cite: Bertram, Matthew. "What Seventy Oil Fields Know About Your AI Model." MatthewBertram.com Research Series, 2026. https://matthewbertram.com/research/better-models-better-decisions