Up To Is Not Achieved: The Faint-Class Share Behind a 75% Inspection Saving

ADNOC Gas's second-quarter results of 10 August 2026 say the company "is also scaling artificial intelligence and robotics" across its assets, "from aerial drones and four-legged inspection robots to tank-climbing crawlers", "with the potential to cut inspection costs by up to 75%, complete certain inspections up to 15 times faster and remove personnel from hazardous environments" [1]. Three months earlier its parent ADNOC said a Taurob heavy-duty robot "is now conducting on-site autonomous inspections as the first set of 'eyes' on the ground" at the Taweelah Gas Compression Plant [2]. On 25 August YPF opened its Sala RTIC Security, a command and control room covering more than 520 critical assets with more than 6,500 cameras, 19 operator positions and 11 drones, 5 of them able to launch on their own from fixed stations [3]. In the same fortnight Suncor reported its first Autonomous Operating Zone at Syncrude Mildred Lake and a control refinement called Mud Mode now running on 90% of the Base Plant fleet [4][5].

Four operators, four different verbs, and not one detection rate. None of the four says what share of faults its detectors find, and none breaks that share down by the kind of fault. That is not a complaint about the releases: three are operational announcements and one is a results statement, and none of them claims to be a performance report on a detector. It is the reason the phrase in the title is worth saying. "Up to 75%" is a statement about a threshold, not about a machine. Where you set the confidence bar at which a machine's call is accepted without a person looking is what decides the saving, and the arithmetic behind that bar has a hard edge that no release mentions.

Four announcements, read exactly

What each one does and does not say is the whole of the evidence, so it is worth reading to the letter.

ADNOC Gas's is the only one of the four that attaches a number to artificial intelligence, and it attaches it to a potential. The words are "with the potential to cut inspection costs by up to 75%" and "complete certain inspections up to 15 times faster" [1]. No model is named, no fault class is named, no detection rate is given, and no site is given. The release is a results announcement; the AI and robotics sentence sits inside it as one sentence. Anyone quoting 75% as an achieved saving is quoting something the release does not say.

ADNOC's own robotics release is a deployment and reads as one. The inspector robot "is now conducting on-site autonomous inspections", it carries 3D LiDAR and thermal cameras with 360 degree visibility, and it is there to help engineers "identify potential gas leaks, unusual hotspots and other hazards" [2]. Two things in that release need separating. The heavy-duty operator robot, the one that turns valves and lifts tools, is not deployed: it is being co-developed under the ARGOS Joint Industry Project and "is due to be operational by the end of 2026" [2]. And the partner list in that project, which names Equinor, Petrobras, TotalEnergies, Saft, Taurob and the Net Zero Technology Centre, is ADNOC's account of who is in the project [2]. It is not those operators speaking about their own programmes, and nothing here should be read as their deployment. The same release states that ADNOC's HSE Cockpit.ai "has reduced safety incidents by 30%" [2]. That figure is about safety incidents. It is not a detection rate, it is not an inspection saving, and it belongs to neither of the other two numbers in this piece.

YPF's release is the one that most rewards careful reading. The Spanish text describes a video-surveillance and incident-management centre and says it strengthens "la capacidad de análisis e inteligencia operacional", operational intelligence, and later that the model of security is "basado en tecnología, inteligencia operativa y prevención" [3]. The phrase "inteligencia artificial" does not appear in the release. Several trade outlets covering the opening wrote that the room combines camera feeds with analytics and artificial intelligence. The operator did not write that. What the operator wrote is a count of assets, cameras, positions, drones and autonomous launch stations, and that "el 100% de las operaciones aéreas son gestionadas de forma remota desde la Sala RTIC por pilotos certificados" [3]. Certified pilots, operating remotely. This is autonomy and digitisation, described as such by the company, and it is worth naming as a data point on exactly those terms.

Suncor's two stories carry no AI wording either. The autonomous haulage story says "19 trucks are equipped for autonomous operation at Mildred Lake, with plans to convert an additional 21 by the end of the year" and a planned fleet growth to 70; the summary box above the same story says 14, which is worth noticing if anyone plans to cite a truck count [4]. The Mud Mode story is explicit about its method: "using data already generated by AHS and truck drive systems", the change let trucks hold traction, and version 1.0 "enabled the movement of approximately six million additional tonnes, generating an estimated $25 million in value", while 2.0 "has improved performance by 40% over version 1.0" and reached 90% of the fleet [5]. Those are real operating numbers about a control strategy on a haul fleet. They are not inspection numbers and they are not detection numbers.

So the pool holds one stated potential on inspection cost, one deployed inspection robot with no published detection rate, and two operators reporting autonomy without claiming AI at all. The number that would let anyone check the 75% is absent from all four.

The one review queue we have measured

For an anonymised Texas onshore operator we stood up the first production review frontend over a digitiser that read scanned well logs, and the thing that made it a delivery rather than a demo was a confidence-sorted queue. The model returned a three-class mask per scan: background, curve one, curve two. Per class, recall told the story plainly. The background mask recalled at 0.97. The two curve masks recalled at 0.37 and 0.32 [6].

The queue used those per-class recalls as a confidence proxy against a bar. A class above the bar auto-accepted and never reached the reviewer; a class below it went into the worklist with the overlay ready. Eight scans entered as 24 regions, the 8 background regions cleared on their own, and the 16 curve regions went to a person [6]. That is an auto-accept share of 8 in 24 at a faint-class share of 16 in 24.

Two things in that account matter for the arithmetic below. The first is that the saving came entirely from the class the model was good at, and the reviewer's whole load came from the classes it was not. The second is why that split was safe: a region the reviewer opens is a region whose misses get caught. The 63% and 68% of curve pixels the model did not find were not a delivery risk, because a person looked at every one of those regions. The 3% the background model missed was the risk, because nobody looked.

The arithmetic

Take an inspection population with classes indexed by $c$, each with a recall $R_c$ and a weight $w_c$ that is its share of the items and, on the assumption that faults are equally prevalent across classes, its share of the faults too, and let a share $\varphi$ of the population sit in faint classes. Per-item confidence inside a class has the class recall as its mean, so with a concentration $\kappa$ the share of class $c$ that clears an auto-accept threshold $\tau$ is

$$A_c(\tau) \;=\; 1 - I_{\tau}\!\left(R_c\,\kappa,\; (1-R_c)\,\kappa\right)$$

FormulaLaTeX

Copy LaTeX

An auto-accepted item costs no review, so the saving is the auto-accept share. A fault escapes only when the detector missed it and its item was then auto-accepted, because a queued item reaches a person. So the two quantities are the same weighted sum with one extra factor:

$$S(\tau,\varphi) \;=\; \sum_c w_c\,A_c(\tau) \qquad\qquad M(\tau,\varphi) \;=\; \sum_c w_c\,(1-R_c)\,A_c(\tau)$$

FormulaLaTeX

Copy LaTeX

Two facts fall out of that pair without any further assumption, and they pull in opposite directions.

Where a target saving is available at all, it comes almost entirely from the strong class, because the gate turns the faint classes away at every threshold worth setting. On that part of the space the sums collapse to a single class and

$$M \;=\; (1 - R_s)\,S$$

FormulaLaTeX

Copy LaTeX

which does not contain the faint share. The faint share is not in it because the faint classes are contributing almost nothing to either sum.

The ceiling is the other way round. At the lowest threshold in range the strong class clears the bar essentially in full and the faint classes essentially not at all, so

$$S_{\max}(\varphi) \;=\; S(\tau_{\min},\varphi) \;\approx\; 1 - \varphi$$

FormulaLaTeX

Copy LaTeX

and the recall of the strong class has fallen out. Set that equal to a target and the crossing sits a hair above $1 - S_{\mathrm{target}}$. For a target of 75% that is a hair above one quarter.

The bench

The exhibit below computes that pair of surfaces rather than drawing them. Left to right is the faint-class share of the inspected population, 0 to 80%. Back to front is the auto-accept threshold, 0.500 to 0.990. Height is the share of items auto-accepted, so the review saved. Colour is the second quantity, the share of all faults that escape, from 0 to 3.0%. The white sheet is a 75% saving, which is ADNOC Gas's stated potential [1] and not an achieved figure anywhere. The curtain is the faint share you set, and the labels follow it.

The exhibit opens on the one operating point we have actually run: a faint share of 66.7% and a threshold of 0.713, which is the threshold that reproduces our recorded split of 8 auto-accepted regions in 24 [6]. The class recalls are ours and recorded. The within-class spread is an assumption, and it sits on its own control so a reader who thinks it is wrong can move it and watch what survives.

faint-class share of inspected items, 0.0% to 80.0%: 66.7%auto-accept threshold, 0.500 to 0.990: 0.713within-class confidence spread, 35 to 200: 40missed-fault budget, 0.0% to 3.0%: 3.0%

What the bench shows

At the opening settings the hero reads 33.3% of items auto-accepted, the faults missed at that point read 1.00%, and the ceiling at this faint share reads 35.2%. The readout for the faults missed at a 75% saving reads "off the table", and the last readout says 75% is gone above a faint share of 25.7%. At the mix we measured, a 75% saving is not expensive. It is unavailable, at every threshold the slider reaches.

Now drag the faint share down and watch two readouts behave differently. The ceiling climbs steadily: 70.9% at a faint share of 30%, 75.7% at 25%, 80.6% at 20%, 100.0% at 0%. The faults missed at a 75% saving does not climb at all. It reads 2.25% at a faint share of 25%, and 2.25% at 20%, and 2.25% at 10%, and 2.25% at 0%. The threshold you would set to take that saving moves a long way over the same range, from 0.698 at a faint share of 25% to 0.922 at 20% and 0.958 at 0%, and the miss it costs does not move with it, because $M = (1-R_s)S$ and neither side of that has a faint share in it.

So the shape of the trade is not what a budget conversation usually assumes. A 75% saving does not become gradually more costly in missed faults as more of the population moves into a class the detector sees faintly. It stays at the same 2.25% and then, at a faint share of 25.7%, it stops being purchasable at any threshold and any miss rate. There is no warning slope to watch. The price is flat right up to a cliff, and the only readout that moves before the cliff is the ceiling.

The budget control shows the same thing from the other side. Set the missed-fault budget to 1.0% and the saving available inside it reads 33.3%, and it reads 33.3% at a faint share of 0%, at 25% and at our recorded 66.7%. The exchange rate is the same everywhere: every 0.1% of missed faults you are willing to sign buys about 3.3% of review saved, and that rate is $1/(1-R_s)$, a property of the class you trust rather than of the classes you do not. Then raise the budget and the two constraints swap over. At a budget of 3.0%, the miss we actually shipped on the class we auto-accepted [6], the saving reads 100.0% at a faint share of 0% but only 35.2% at our recorded mix, because there the ceiling has taken over from the budget.

The spread control is where an unconvinced reader should go first. Drag it across its whole range, 35 to 200, and the crossing readout moves from 26.0% to 25.0%. That is the assumption in this model doing almost nothing: the crossing is close to $1 - S_{\mathrm{target}}$ because at a low threshold a class at recall 0.97 is accepted in full and a class at 0.37 or 0.32 is not, and no plausible spread changes that ordering. The finding is not a consequence of the shape we assumed.

One more figure is worth naming, and it is not on the plate: it comes from the same model module, and the unit test for this exhibit checks it. At our recorded mix the population-weighted mean recall is 0.553. A single class at that same mean recall would have a ceiling of 75.3%, against the 35.2% the mix actually reaches. The mean recall says a 75% saving is on the edge of possible. The mix says it is not close. Averaging recall across classes destroys the information the threshold decision needs, which is why any inspection AI programme should be asked for the table and not the headline.

Where the recalls come from, and what they are not

The three recalls on the plate were measured on a semantic segmentation model reading scanned well logs, not on a tank-climbing crawler, an aerial drone or a four-legged robot. A crawler looking for wall loss and a network looking for a faded curve are different problems with different sensors, and the transfer from one to the other is an analogy. The plate says so and so does this paragraph.

What carries across is not the numbers but the structure, and the structure is the part a reader can check without trusting us at all. Any detector applied to a mixed population has a per-class detection rate. Any gate that auto-accepts on confidence auto-accepts the classes it is good at and queues the classes it is not. Any saving that gate produces is therefore bounded by the share of the population in the classes it is good at, and any fault that escapes escaped because it was both missed and auto-accepted. Put your own recalls on the sliders and the ceiling still lands a hair above one less the target, because that is where the two sums put it.

The shape of confidence inside a class is an assumption, and it has a control. The one thing this model would need to be told, and cannot compute, is which classes an operator's inspection population actually falls into and what its detector recalls on each. That is the table none of the four releases publishes.

Three questions for a 75% inspection saving

What is the per-class detection rate, and which class is worst? Not an average and not a headline accuracy: the table, one row per fault type the programme is meant to catch. On our own queue the average across three classes was 0.553 and the number that set the outcome was 0.32 [6], and a single figure would have hidden it.

What share of the inspection population sits in the worst class? That coordinate alone sets the ceiling, and above about a quarter it puts a 75% saving out of reach whatever confidence bar is chosen. The ADNOC Gas release states a potential saving [1]; it does not state the population split that would make it reachable, and neither does any other release quoted here.

What miss rate will the integrity engineer sign, and has anyone converted it into a saving? The conversion is one division, $S = m/(1-R_s)$, and it is worth doing before the saving goes into a plan rather than after. A robot that is already conducting autonomous inspections [2] and a room that already watches 520 assets [3] are both places where that division has a real answer today. The releases show the equipment. The threshold is the decision, and it is still an unpublished one.

Key takeaways

1. 1ADNOC Gas states a POTENTIAL saving of up to 75% on inspection costs and up to 15 times faster inspections, with no model, class or detection rate named. ADNOC separately reports a Taurob inspector robot already conducting autonomous inspections at Taweelah, while the ARGOS operator robot is due to be operational only by the end of 2026.
2. 2ADNOC's release is the source for Equinor, Petrobras and TotalEnergies being in the ARGOS project: that is ADNOC's account, not those operators speaking about their own programmes. ADNOC's separate figure of a 30% reduction in safety incidents through HSE Cockpit.ai is about safety incidents, not detection and not inspection cost.
3. 3YPF's own release says operational intelligence, never artificial intelligence, over more than 520 assets, more than 6,500 cameras and 11 drones with 5 autonomous launch stations, and says all aerial operations are flown remotely by certified pilots. Trade coverage that called it artificial intelligence was not quoting the company. Suncor's two stories carry no AI wording at all.
4. 4On the one review queue we have measured, per-class recall was 0.97 on the background mask and 0.37 and 0.32 on the two curve masks, and 8 of 24 regions auto-accepted while 16 went to a reviewer. The saving came entirely from the class the model was good at.
5. 5Behind a confidence gate the missed faults a 75% saving costs is 2.25% at every faint-class share where that saving is available, because it equals the target times one less the strong class recall and contains no faint share. The saving does not get more expensive as the faint share rises.
6. 6The ceiling is the constraint instead, and it depends on the faint share alone: 100.0% at a faint share of 0%, 80.6% at 20%, 75.7% at 25% and 35.2% at our recorded 66.7%. A 75% saving goes off the table above a faint share of 25.7%, and that crossing moves only from 26.0% to 25.0% across the whole assumed spread range.
7. 7The population-weighted mean recall is a decoy. At our recorded mix it is 0.553, and one class at that mean would reach a ceiling of 75.3% against the 35.2% the real mix reaches. Ask for the per-class table, not the average.

References

[1] ADNOC Gas. ADNOC Gas Delivers Resilient Q2 Net Income, Takes FID on Major Growth Projects. 10 August 2026. https://www.adnocgas.ae/en/news-and-media/press-releases/2026/q2-2026-results

[2] ADNOC. ADNOC Deploys Industry-First Heavy-Duty Robot to Strengthen Safety, Reliability and Performance. 21 May 2026. https://www.adnoc.ae/en/news-and-media/press-releases/2026/adnoc-deploys-industry-first-heavy-duty-robot-to-strengthen-safety-reliability-and-performance

[3] YPF. Comunicado de prensa: YPF inauguró la Sala RTIC Security, un centro de comando y control de última generación. Buenos Aires, 25 August 2026. https://novedades.ypf.com/r/documents.html?p=inauguracion-rtic-security.pdf

[4] Suncor Energy. Back-to-back wins on our autonomous mining journey. 13 August 2026. https://www.suncor.com/en-ca/news-and-stories/our-stories/wins-on-our-autonomous-mining-journey

[5] Suncor Energy. When the going gets muddy, Mud Mode gets going. 20 August 2026. https://www.suncor.com/en-ca/news-and-stories/our-stories/when-the-going-gets-muddy-mud-mode-gets-going

[6] EarthScan. Eight Real Scans, One Reviewer: Standing Up Human-in-the-Loop Validation. https://earthscan.io/case-studies/human-in-the-loop-review-eight-real-scans