The Constraint Is Trust, Not Access: A 13.02% Worst Class Zeroes the Yield

On 27 August 2026, on a panel at the SPE Artificial Lift Conference and Exhibition, Kreg Flowers, a production engineer at Occidental, said the foundational skills of artificial lift have not changed and warned that AI can create false expertise; Maggy Burns, an operations supervisor for produced water at ExxonMobil, stressed input data quality, domain knowledge and human oversight on safety-critical decisions; and Araceli Rivera Mandujano of SLB, the vendor voice on that panel, warned about information overload [1]. Three weeks later, at the SPE Subsea Well Intervention Symposium in Galveston, Texas, Sean McCall, Chief Data Officer at Oceaneering, and Brian Saucier, President of DeepMar Consulting, with SPE's Dana Otillio alongside them, argued that access to modern AI models guarantees no successful deployment, that data quality has to be assessed before AI is deployed against it, that humans must remain in command rather than be observers, and that pilots against historical low-risk projects are the right validation path. Aramco Americas is named in that write-up [2].

Two venues, three weeks apart, one complaint: the constraint is trust, not access. Neither panel contains a number. An unpriced complaint loses to a deck that carries figures, so here is the arithmetic.

What the sourcing actually is

Both legs are pointers. The Journal of Petroleum Technology is the Society of Petroleum Engineers' own journal, but what it published on 8 September 2026 is JPT writing up a panel, not Occidental or ExxonMobil publishing their own account [1]; the date that belongs to the substance is the panel's, 27 August 2026. The second leg is thinner: the two quoted at length are a vendor's chief data officer and a consultancy's president, no operator engineer is quoted, and Aramco Americas appears by reference, which puts an operator's regional arm on the record without making the argument an operator's. JPT published it on 10 September 2026, reporting a symposium held on or before that date [2].

Behind both sits a vendor's document. On 8 September 2026 SLB published a question and answer with Paul Sims, president of its Production Systems division, putting approximately 70% of global production as past peak, approaching 80% by 2030, and conventional reserve replenishment close to 30% [3]. Those are SLB's figures, sourced on its own page to an analysis it does not name, in a position piece that is neither a deployment nor an intent and says nothing at all about AI: what an AI workflow does to that base is our question, not SLB's. It is the size of that base here, never a third voice.

The one archive we measured per class

Ours is the only per-class evidence we recorded ourselves. For an anonymised Texas onshore operator in 2023 we stood up a human-in-the-loop validation queue over a well-log digitisation model, on eight genuine field scans, checked against the operator's own logs by one reviewer [4]. Every scan produced a three-class mask: the background recalled at 0.97, the two curve masks at 0.37 and 0.32 [4]. Eight scans times three classes is 24 regions, and the auto-accept line we shipped put the 8 background regions above the bar and the 16 curve regions below it, the corrected curves landing at a CSV mean absolute error of 0.11 and 0.12 against the operator's own log [4].

Two things transfer: the spread, 0.97 against 0.32 in the same deliverable from the same model, and the share, 8 of 24 regions, which is 33.33%. A figure quoted over a whole archive is an average whose value depends on what fraction of it is that class.

The two accounts

Take an archive in two classes. A fraction $w$ is the class the sensor reads worst, at recall $r$; the remaining $1-w$ is the class it represents well, at the 0.97 we recorded. The average per-class recall, weighted by archive share, is

$$a(w, r) \;=\; (1-w)\,R \;+\; w\,r, \qquad R = 0.97$$

FormulaLaTeX

Copy LaTeX

A vendor account of the workflow is the mature base $C$ times that average, $V(w, r) = C\,a(w, r)$. On its own terms that is not wrong; a miss is an item the workflow did not improve. An engineer signing production surveillance cannot use it, for the reason Flowers named: a recommendation acts on the well whether or not the class behind it was read correctly, so a miss in the class nobody checks is not a foregone gain but an action on a well the model misread, and it costs. Write $k$ for what one escaped miss costs, in units of what one unattended item saves:

$$E(w, r) \;=\; C \left( a(w, r) \;-\; k\,w\,(1-r) \right)$$

FormulaLaTeX

Copy LaTeX

Only the worst class carries the multiple, because the represented class's misses are visible, which is what 0.97 on the background meant [4]. The difference between the accounts is one term, $C\,k\,w\,(1-r)$, linear in the worst-class share. So is the vendor account, but with slope $(R - r)$ rather than $k\,(1-r)$, and that ratio is why the two move at completely different rates. Set the engineer account to zero and solve for the share:

$$w^{*} \;=\; \frac{R}{\left(R - r\right) + k\,\left(1 - r\right)}$$

FormulaLaTeX

Copy LaTeX

The exhibit

The plate carries those two families. Left to right is the worst class's recall, pinned at our recorded 0.32, 0.37 and 0.97 [4]; up the side is the share of global production reached, net, under a rule at 67.90%, which is SLB's approximately 70% past-peak base at 0.97, and 77.60% on the 2030 button [3]. Nothing below that rule is SLB's. Steel is the vendor account, aqua the engineer account, the amber rule is zero, and the escape multiple is a stated assumption, not a measurement [4].

worst class, 0% to 30% of the archive: 15.00%escape multiple, 1 to 20 (assumed, not recorded): 10.00

past-peak ceiling (SLB, not ours)

70% now80% by 2030

At the opening state the vendor account stands at 61.08% of global production on our recorded worst-class recall of 0.32, and the engineer account stands at -10.32%. The other faint curve we recorded, 0.37, reads -4.55%. The class our archive read best, 0.97, is the right hand pin, and it is where the two accounts very nearly meet.

What it shows

It opens at a worst class of 15%, the control's midpoint and an assumption. At our recorded 0.32 the vendor account reads 61.08% of global production and the engineer account -10.32%, a gap of 71.4 points, on an archive where 85% of items are read at 0.97; at 0.37 they read 61.60% and -4.55%. Across the seven shares the vendor family stays inside 13.7 points while the engineer family spans 156.4 points, 11.46 times as wide, four of its seven below zero.

The finding is in the readouts. At an escape multiple of 10 the worst-class share that zeroes the engineer account on our recorded 0.32 is 13.02%. There the vendor still reads 61.98%, which is 91.28% of the 67.90% it prints on an archive with no worst class in it at all, and it does not fall under 90% of stated until 14.92%. The engineer reaches zero 1.9 points of share before the vendor has lost a tenth of what it claims: one party calls a fall from 67.90% to 61.98% a rounding error, the other watches the workflow stop paying for itself.

Press SLB's 2030 ceiling and everything scales, the stated ceiling to 77.60% and the vendor at 0.32 to 69.80% against an engineer at -11.80%, while the crossing does not move at all: a bigger denominator is not a better archive. Drag the escape multiple instead and the share that zeroes our 0.32 runs from 28.78% at 4 to 13.02% at 10 and 6.81% at 20. Our own archive is off the plate: at its recorded 33.33% the engineer account reads -105.93%, zeroed by a multiple of just 3.3235.

What is assumed, and where this is weakest

The escape multiple is the largest assumption and the least defended: we have never priced one escaped miss against one avoided review, and neither panel prices it, which is why it is a slider with its sensitivity printed. Two classes is a simplification, and an optimistic one. And none of this measures anyone's production archive. Ours comes from eight well-log scans for one operator in 2023, raster digitisation and not artificial lift [4], and the only checkable quantity near the argument is out of window: SPE 219553, ExxonMobil authors, presented at the SPE Artificial Lift Conference and Exhibition Americas in The Woodlands in August 2024, reports a closed-loop gas lift workflow on more than 1,300 ExxonMobil Permian wells for an uplift of about 2.2%, with no change to surface or downhole equipment [5]. Two years older than either panel, it is a scale check and nothing more: a figure in the second decimal place.

What a buyer should ask

What is the class your sensor reads worst, and what is its recall? Not the headline accuracy and not the average; our own file had 0.97 and 0.32 in the same deliverable, and it is the 0.32 that decided how the reviewer spent her week [4]. What share of your archive is that class? It is the one quantity here a buyer can measure before signing, and the one that moves the engineer account fastest: at a multiple of 10 the difference between a worst class of 10% and one of 15% is the difference between 15.75% of the base and -10.32%. What does one escaped miss cost against one review avoided? Above 8.5539 at 15%, the workflow is not a saving at our recorded recall. And who signs? Both panels demanded humans in command rather than observers [1][2], and the practical form of that is a routing rule: a queue that auto-accepts the class the model has earned and routes the class it has not [4].

An operator who wants reservoir, wells, lift, chemistry and facilities read as one system, with per-class evidence attached before an AI claim on any of them is priced, is describing what EarthScan 360 exists to do; narrower, we will run a production-surveillance review on your worst class rather than your average. Two panels of the industry's own people named the constraint correctly and in public, and left a deck full of averages standing because nobody put an arithmetic behind the objection.

Key takeaways

1. 1Two SPE panels three weeks apart made the same argument and carried no quantities at all. On 27 August 2026 at the SPE Artificial Lift Conference and Exhibition, Kreg Flowers, a production engineer at Occidental, warned that AI can create false expertise, and Maggy Burns, an operations supervisor for produced water at ExxonMobil, argued for input data quality, domain knowledge and human oversight on safety-critical decisions. Both are reported by JPT, which is SPE's own journal writing up a panel rather than an operator publishing its own account, and the date that belongs to the substance is the panel's.
2. 2The second panel is thinner and it has to be said so. At the SPE Subsea Well Intervention Symposium in Galveston, reported on 10 September 2026, the people quoted are Sean McCall, Chief Data Officer at Oceaneering, and Brian Saucier, President of DeepMar Consulting. No operator engineer is quoted. Aramco Americas is named by reference, which puts an operator's regional arm on the venue's record and does not make the argument an operator's.
3. 3SLB's figures of approximately 70% of global production now past peak, approaching 80% by 2030, come from a position piece published on 8 September 2026 that says nothing about AI and is sourced on its own page to an analysis it does not name. They are used here only as the base the exhibit's y axis is a share of, never as a third voice agreeing with the two panels.
4. 4On eight real well-log scans for one operator we recorded per-class recall of 0.97 on the class the archive represented best and 0.37 and 0.32 on the two it represented worst, in the same deliverable from the same model, with the worst class at 8 of 24 regions, which is 33.33%. A single averaged figure over an archive like that describes no class in it.
5. 5At a worst class of 15% of the archive and an escape multiple of 10, the vendor account reads 61.08% of global production at our recorded worst-class recall of 0.32 and the engineer account reads -10.32%, a gap of 71.4 points. The engineer account is zeroed by a worst class of just 13.02%, and at that point the vendor account still prints 91.28% of its stated ceiling, because the vendor does not fall under 90% until 14.92%.
6. 6Across worst-class shares from 0% to 30% at a recall of 0.32, the seven vendor curves lie inside 13.7 points of the production base while the seven engineer curves span 156.4 points, 11.46 times as wide, and four of the seven finish below zero. Pressing SLB's 2030 ceiling scales the whole plate and moves the crossing by nothing at all: a bigger denominator is not a better archive.
7. 7The escape multiple is this exhibit's assumption and not our measurement. Our record supports only the ordering, that an auto-accepted wrong region is worse than a queued one. At a worst class of 15%, the account stays positive only while an escaped miss costs less than 8.5539 avoided reviews; on our own recorded archive share of 33.33% it takes only 3.3235.

References

[1] Journal of Petroleum Technology. New Tools, Old Lessons: Experts Stress the Importance of Artificial Lift Basics. 8 September 2026, reporting a panel held on 27 August 2026 at the SPE Artificial Lift Conference and Exhibition. Pointer, not an operator-published account: JPT is the Society of Petroleum Engineers' own journal writing up operator employees speaking on a panel. The article carries no quantities. https://jpt.spe.org/new-tools-old-lessons-experts-stress-the-importance-of-artificial-lift-basics

[2] Journal of Petroleum Technology. Industry Executives Offer Practical Advice on How To Use Modern AI Models Wisely. 10 September 2026, reporting the SPE Subsea Well Intervention Symposium held on or before that date in Galveston, Texas. Pointer: the panellists quoted are a vendor Chief Data Officer and a consultant, and no operator engineer is quoted. Aramco Americas is named in the piece by reference. The article carries no quantities. https://jpt.spe.org/industry-executives-offer-practical-advice-on-how-to-use-modern-ai-models-wisely

[3] SLB. Q and A with Paul Sims, President, Production Systems. 8 September 2026. A position piece, neither a deployment nor an intent, making no AI claim of its own. The figures of approximately 70% of global production now past peak, approaching 80% by 2030, and conventional reserve replenishment close to 30% are SLB's, sourced on the page to a recent analysis it does not name. https://www.slb.com/newsroom/features/2026/2026-0908-paul-sims-qa

[4] EarthScan. Eight Real Scans, One Reviewer: Standing Up Human-in-the-Loop Validation. Source file content/case-studies/human-in-the-loop-review-eight-real-scans.mdx. Every figure of ours in this piece is quoted from it: per-region recall of 0.97, 0.37 and 0.32, 24 regions across 8 scans, 8 auto-accepted and 16 queued, and delivered curves at a CSV mean absolute error of 0.11 and 0.12 against the operator's own log. One anonymised Texas onshore operator, raster log digitisation, not artificial lift. https://www.earthscan.io/case-studies/human-in-the-loop-review-eight-real-scans

[5] Society of Petroleum Engineers. SPE 219553, Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence. ExxonMobil authors. SPE Artificial Lift Conference and Exhibition Americas, The Woodlands, Texas, August 2024. Operator-authored and out of window: the work is two years older than the panels above, and it is cited here only as a scale check on what a deployment on a mature estate reports. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379