The number you can't plan around: why upstream AI has to show its uncertainty
By Serhii Honcharuk · Published October 8, 2026
Subsurface decisions are made with incomplete data, and the petroleum industry has had a standard way of saying so for a long time. Under the SPE Petroleum Resources Management System, an estimate is reported as a low, a best and a high case: at least a 90% probability that what is actually recovered will equal or exceed the low case (P90), 50% for the best case (P50) and 10% for the high case (P10). The framework’s own guidance is that uncertainty is communicated best by reporting a range.
Much of the AI now being offered to upstream teams ignores that convention. It returns one answer: a predicted recovery, a picked horizon, a flagged zone. The interface implies a certainty the underlying data cannot support. A planner who reads P90/P50/P10 every day is asked to trust a single number from a system that cannot say how sure it is.
A single number hides what the decision needs
Take two forecasts with the same best estimate of 500 thousand barrels. In one, the realistic spread runs from 450 to 550; in the other, from 200 to 900. (Illustrative figures.) The first supports committing capital. The second may argue for more data, a smaller first phase or a hedge. The best estimate alone cannot tell them apart, and the decision depends on exactly that difference.
People are already overconfident, and software can inherit it
In a well-known 1976 paper in the Journal of Petroleum Technology, E. C. Capen showed that petroleum engineers asked for ranges they were highly confident would contain the right answer set those ranges too narrow. A later SPE study of 56 oil and gas professionals found that giving a best guess first made ranges narrower still. A model trained on interpretations, forecasts and decisions made by overconfident people does not escape that by being software. It has to be tested for it.
What good looks like
We are designing our tools around four habits:
- Speak the audience’s language. Report ranges as P90/P50/P10, not a proprietary “confidence score”.
- Check the ranges against reality. If a tool says 80%, about 80% of held-out wells should land inside its range, and the results of that test should be shown.
- Say when the model is extrapolating. Flag inputs that sit outside the data the model has seen.
- Show what drove the answer, and let the expert overrule it. The engineer decides; the tool keeps the record.
An answer with a range and a reason is something a team can plan around. An answer without them is a guess with good typography.
Sources
- SPE PetroWiki, “Uncertainty range in production forecasting”: https://petrowiki.spe.org/Uncertainty_range_in_production_forecasting
- SPEE Houston Chapter, overview of the SPE/WPC/AAPG/SPEE Petroleum Resources Management System: https://www.spee.org/pdf/2008-11-houston_filler_prms.pdf
- E. C. Capen, “The Difficulty of Assessing Uncertainty,” Journal of Petroleum Technology 28 (1976), pp. 843–850.
- A. Heywood-Smith, M. Welsh and S. Begg, “Cognitive Errors in Estimation: Does Anchoring Cause Overconfidence?”, SPE-116612-MS (2008), https://doi.org/10.2118/116612-MS