12.2% Against What? The Missing Baseline Behind an AI Production Uplift

Three national oil companies put artificial intelligence on the record in August 2026, and between them they published one set of performance figures, none of which is a production number and none of which carries a baseline.

ADNOC, with SLB, announced on 4 August that it had deployed its AI-enabled Real-Time Operations Center across a fleet of more than 120 drilling rigs, onshore and offshore, hosted in ADNOC's own cloud environment [1]. That release does carry figures: the platform "reduces engineering effort by 30-40%", helps "reduce incident response times by 4-12 hours", and helps "avoid one to two days of rig downtime" [1]. They are drilling figures, not production figures, and the page states no measurement period, no comparison method and no baseline for any of them. On 11 August Baker Hughes announced a multi-year contract with Kuwait Oil Company for the Ahmadi Innovation Valley project, saying the technologies "will benefit from Baker Hughes' portfolio of digital and artificial intelligence (AI) automation solutions that help operators increase recovery from existing wells, lower operating costs, reduce water production and minimize power consumption" [2]. That is the service company's account of the operator's programme, not the operator's: the only person quoted is Baker Hughes' chairman and chief executive, no Kuwait Oil Company executive is quoted anywhere on the page, and there is no contract value, no well count and no percentage in it [2]. On 26 August CNOOC Limited said in its interim results that it "fully deployed the 'Haineng-Zhiqing' digital platform" and that its "intelligent injection-production interaction scenario for offshore oilfield production was selected as a high-value scenario at the 2026 World Artificial Intelligence Conference" [3]. Neither sentence has a number attached to it.

The only production uplift figures in the window belong to two small operators. Both are worth reading to the letter, because both raise the same question and neither answers it.

The two numbers, read exactly

Robert Thaxton, Vice President of Oil and Gas Business Operations at U.S. Energy Development Corporation, wrote in World Oil that an artificial lift optimisation pilot "was deployed across 12 Permian basin wells" beginning in December 2025, that "the system executed thousands of automated adjustments that previously would have required individual review", and that "production increased 12.2%, compared to the 90-day pre-activation baseline" [4]. This is an article by an operator's executive in a trade magazine, not a press release; the company's newsroom post of 31 August links to it and does not itself carry the figures [5]. No vendor, model or technique is named on the page, and nothing on it says the pilot has scaled past those 12 wells [4].

Presidio Production Company, a small US operator that acquires and optimises producing wells without drilling, said in its Q2 2026 results, filed with the SEC as an exhibit, that it "averaged approximately 22.8 MBoe/d of production for the second quarter", that a new engineering team "is developing Presidio's AI platform, which the Company is deploying first" across its own operations, and that its Asset Intelligence Group "carries a target of three to five percent production growth in 2026" and "has achieved approximately one percent of production uplift to date" [6]. The one percent is credited on the page to the group, not to the platform, and the 22.8 MBoe/d is a rate, not a well count.

That is the whole of the public evidence, and it is enough to ask the question in the title. 12.2% against what?

The answer on the page is: against the average of the 90 days before the system was switched on. That average is not the counterfactual. The counterfactual is what those 12 wells would have produced over the same period if nothing had been done, and producing wells decline. The pre-activation average therefore sits above the counterfactual by an amount set by the decline rate and the length of the window, and neither disclosure states either one. Presidio does not state a baseline window at all.

What we know about this from our own work, and what we do not

Our own reason for caring is a history matching habit rather than a published result. When we condition a reservoir model to production history with an ensemble smoother, the ensemble does two things at once that a single-number uplift claim throws away. It carries the decline the wells were already on, so every forecast is a forecast against a declining counterfactual rather than against a flat one. And the spread between the ensemble members, month to month, is a working estimate of the noise a real signal has to beat. We have no published case study for that ensemble work, and this piece does not pretend otherwise: what follows treats the decline rate and the scatter as the reader's inputs, and the arithmetic connecting them is ours.

What we have published on the machine-learning side of the same problem is an illustrative scenario, a composite case study rather than a client record. In it, an agentic MLOps loop across dozens of models and about 40 producing assets has a monitoring agent watch live production feeds, a reconciliation agent version-stamp the inputs, and a retraining agent rebuild and back-test, with the geoscientist as approver. Retrains go from six weeks to overnight and production-forecast accuracy "lifted 22% across the portfolio" [7]. The number that matters for this piece is not the accuracy lift. It is what the scenario is about: before the loop, drift "went undetected for months", because nobody had a running estimate of what the forecast should have been [7]. A production uplift measured against a trailing average has the same hole in it, and the hole is the same size.

The arithmetic

Take a well set declining exponentially at effective annual rate $d$, so its rate is $q(t) = q_0 e^{-a t}$ with $a = -\ln(1-d)$ and $t$ in years. The system is switched on at $t=0$. The baseline is the mean rate over the window of length $B$ before that moment, and the measurement is the mean rate over a window of the same length after it. With no intervention at all, the ratio of the second mean to the first is

$$K(d,B) \;=\; \frac{\frac{1}{B}\int_{0}^{B} e^{-at}\,dt}{\frac{1}{B}\int_{-B}^{0} e^{-at}\,dt} \;=\; \frac{1-e^{-aB}}{e^{aB}-1} \;=\; e^{-aB}$$

FormulaLaTeX

Copy LaTeX

exactly, because the two integrals differ by one factor of $e^{aB}$. Taking the measurement window equal to the baseline window is a design choice, and it is the choice that makes the correction exactly an exponential rather than a ratio of two integrals. So a measured uplift $U$ against the trailing baseline implies a true uplift against the counterfactual of

$$u_{\text{true}}(U,d,B) \;=\; \frac{1+U}{K} - 1 \;=\; (1+U)\,e^{aB} - 1$$

FormulaLaTeX

Copy LaTeX

which is exactly $U$ when $d=0$, and rises with both the decline rate and the length of the window. There is no fitted parameter in it. It is the arithmetic of comparing a forward window mean with a backward window mean on a curve that is going down.

The second question is whether the design could have resolved the answer. With $n$ wells, $m = B/30.4375$ monthly aggregates in each window, and a per-well relative month-to-month scatter $s$, the relative standard error of the ratio of the two window means is $s\sqrt{2/(nm)}$. At the null the measured ratio is centred on $K$, so that same figure is the standard error carried onto the true-uplift scale, and the smallest true uplift a two-sided 5% test detects with 80% power is

$$\text{MDU}(s,n,B) \;=\; \left(z_{0.975} + z_{0.80}\right) s \sqrt{\frac{2}{nm}} \;=\; 2.8016\, s \sqrt{\frac{2}{nm}}$$

FormulaLaTeX

Copy LaTeX

Two things follow that are worth having in front of you before the exhibit. The decline rate is in the first expression and not in the second, so it changes what the number means without changing what could be measured. And the window $B$ is in both, with opposite signs: it raises $u_{\text{true}}$ through $e^{aB}$ and lowers the floor through $m$.

The bench

The exhibit below computes those two surfaces rather than drawing them. Left to right is the effective annual decline of the well set across the windows being compared, 0 to 70%. Back to front is the baseline window, 30 to 365 days. Height is the true uplift the measured number implies. The amber sheet is the smallest uplift the design could resolve. The white curtain is your window and the pale stem is your decline.

It opens on the one disclosure that states a baseline: 12.2%, 12 wells, 90 days [4], with the decline set to zero, which is what a comparison against a trailing average assumes without saying so. The scatter opens at 8%, and that is an assumption, labelled as one on the plate: we have not published a figure for month-to-month scatter on a producing well set, and the slider is where yours goes. The wells are treated as independent, which is optimistic for twelve wells in one basin, so the floor the sheet draws is the friendliest one available.

measured uplift, 0.5% to 20.0%: 12.2%effective annual decline, 0.0% to 70.0%: 0.0%baseline window, 30 days to 365 days: 90 daysmonth-to-month scatter, 2.0% to 20.0%: 8.0%wells, 1 to 5,000: 12

What the bench shows

At the opening settings the hero reads 12.2%, because at zero decline the correction is exactly one and the measured number is the implied number. The smallest resolvable uplift reads 5.3%, the claim stands at 2.29 times that floor, and three wells would have been enough to resolve it at zero decline. On those numbers the pilot looks unambiguous.

Now move the decline control, which is the one the disclosure sets to zero without saying so. At 30% a year the hero reads 22.5% and the claim stands at 4.23 times the floor. At 70% it reads 51.0%. The amber sheet does not move through any of this, because the expression that draws it contains the scatter, the well count and the window and does not contain the decline. One measured number, one unchanged dataset, and a true uplift that runs from 12.2% to 51.0% depending on a figure nobody published.

Then drag the baseline window from the 90 days the operator states to a year, with the decline left at 30%. The hero goes from 22.5% to 60.2% and the floor drops from 5.3% to 2.6%, so the claim goes from 4.23 times its floor to 22.80 times it. Nothing happened in the field. The only thing that changed is how far back the comparison reaches, and it made the claim nearly three times larger and its error bar half as wide at the same time. That is the finding, and it is why the baseline window is the first thing to ask for: it is the one design choice that flatters a production AI result twice over, and it costs nothing to lengthen.

The floor is where the two small operators separate. Set the measured uplift to Presidio's 1% and the wells readout jumps from three to 340: at 8% scatter and a 90-day window, a 1% portfolio uplift needs several hundred wells before it can be told from noise, and Presidio published a rate, 22.8 MBoe/d, rather than a well count [6]. Push the scatter to 20% with the 12 wells and the 12.2% back in place, and the floor rises to 13.3%, above the measured number: the readout then says the claim clears the floor only above a decline of 3.9% a year, which is a strange sentence until you notice what it means. On noisy wells, the pilot's number is only distinguishable from nothing because the wells were declining. The decline is what makes the result real, and it is the thing the measurement removed.

What the exponential is, and what it is not

One assumption in the model deserves to be met head on, because a reservoir engineer will reach for it first. Decline here is exponential at a constant effective annual rate across the two windows, and unconventional wells are not exponential. The axis is therefore labelled as the effective annual decline of the well set across the windows being compared, not an initial-production decline: a hyperbolic curve at the same instantaneous rate flattens over the window and implies a smaller correction than the exponential does, which is the direction to bear in mind on young wells. The other assumptions are on the plate: the intervention is a constant proportional shift on the counterfactual rate, the two windows are equal in length, and the wells are independent.

None of that changes the shape of what the bench shows. The correction is one at zero decline and grows with the product of the decline and the window, whatever the decline curve's exact form, because it is a statement about comparing a forward mean with a backward mean. The floor falls with the square root of the number of well-months, whatever the wells are producing. Those two facts are the argument. Our numbers only set where you start.

Three questions for anyone publishing a production uplift

What is the baseline window, and why that length? U.S. Energy states 90 days, which is more than most disclosures manage [4], and the bench shows what a year would have done to the same result. Presidio states none [6]. Until the window is on the page, the size of the claim is not determined.

What was the decline rate of that well set over the comparison period? Not a type curve and not a field average: the effective decline of those wells across those two windows. It is the difference between a 12.2% result and a 51% one, and an operator that has the production history has the number.

How many wells, and how noisy are they month to month? Three wells clear 12.2% at 8% scatter; 340 are needed to clear 1% at the same scatter and window. A rate in MBoe/d does not answer this, and neither does a platform name. It is the pair of numbers that decides whether a reported uplift is a measurement or a rounding of the noise, and no release quoted in this piece contains it.

Key takeaways

1. 1ADNOC with SLB published performance figures for an AI-enabled Real-Time Operations Center across more than 120 rigs, with no measurement period or baseline stated and none of them production figures. Baker Hughes described AI automation for Kuwait Oil Company at Ahmadi Innovation Valley with no KOC executive quoted and no number on the page. CNOOC Limited said it fully deployed the Haineng-Zhiqing platform, with no number attached.
2. 2The only in-window production uplift figures are from two small operators: 12.2% against a 90-day pre-activation baseline on 12 Permian wells, from an article by U.S. Energy Development Corporation's VP in World Oil, and about 1% credited to Presidio Production Company's Asset Intelligence Group rather than to its AI platform.
3. 3A pre-activation average is not the counterfactual, because producing wells decline. With the measurement window taken equal to the baseline window, the implied true uplift is (1 + measured) times e to the aB, less one.
4. 4At the opening settings the same measured 12.2% implies 12.2% at zero decline, 22.5% at 30% annual decline and 51.0% at 70%, all on the 90-day window the operator states. The detectability sheet does not move while that happens, because the decline is not in the expression that draws it.
5. 5Lengthening the baseline window moves both surfaces the flattering way at once. At 30% decline, going from 90 days to a year takes the implied uplift from 22.5% to 60.2% and the noise floor from 5.3% to 2.6%, so the claim goes from 4.23 to 22.80 times its floor with nothing changed in the field.
6. 6Three wells resolve a 12.2% uplift at 8% month-to-month scatter on a 90-day window; 340 are needed for a 1% uplift on the same terms. Presidio published a rate of 22.8 MBoe/d rather than a well count, so the reader cannot tell which side of that line it falls on.

References

[1] ADNOC. ADNOC and SLB Deploy AI Platform Across Over 120 Drilling Rigs to Strengthen Upstream Performance. 4 August 2026. https://adnoc.ae/en/news-and-media/press-releases/2026/adnoc-and-slb-deploy-ai-platform-across-over-120-drilling-rigs-to-strengthen-upstream-performance

[2] Baker Hughes. Baker Hughes Awarded Multi-Year Contract by Kuwait Oil Company for Ahmadi Innovation Valley Project. GlobeNewswire, 11 August 2026. https://www.globenewswire.com/news-release/2026/08/11/3342544/0/en/baker-hughes-awarded-multi-year-contract-by-kuwait-oil-company-for-ahmadi-innovation-valley-project.html This is Baker Hughes' release about a Kuwait Oil Company project, not a Kuwait Oil Company release.

[3] CNOOC Limited. CNOOC Limited Focuses on Value Creation, Production and Profit Hit New Highs in H1 2026. 26 August 2026. https://www.cnoocltd.com/english/presscenter/pressreleases/2026/202608/t20260826_122439.html

[4] Robert Thaxton, Vice President of Oil and Gas Business Operations, U.S. Energy Development Corporation. Putting A.I. to work: The disciplined approach to innovation in oil and gas operations. World Oil, August 2026. https://www.worldoil.com/magazine/2026/august/features/putting-a-i-to-work-the-disciplined-approach-to-innovation-in-oil-and-gas-operations/ An article by a company executive in a trade magazine, not a company press release.

[5] U.S. Energy Development Corporation. Putting A.I. to work: The disciplined approach to innovation in oil and gas operations. Company newsroom, 31 August 2026. https://www.usedc.com/putting-a-i-to-work-the-disciplined-approach-to-innovation-in-oil-and-gas-operations/ A short post linking to the World Oil article; it does not itself carry the 12.2 percent or the well count.

[6] Presidio Production Company. Second Quarter 2026 Results. Business Wire release filed with the SEC as Exhibit 99.1, 11 August 2026. https://www.sec.gov/Archives/edgar/data/2083125/000121390026089284/ea030206301ex99-1.htm

[7] EarthScan. Illustrative scenario: Agentic MLOps: From 6-week retrains to overnight. A composite case study, not a client engagement. https://earthscan.io/case-studies/agentic-mlops-six-week-retrains-to-overnight