Phase Two Is Not a Deployment: The Record-Completeness Exponent Behind a 120-Well AI Contract

On 2 September 2026, at Oil and Gas Development Company headquarters in Islamabad, Baker Hughes announced a multi-year contract to raise production from more than 120 wells across OGDC's Tando Alam Oil Complex and the Pirkoh field [1]. The work is split in two. Phase one evaluates the wells and produces tailored redevelopment plans, with Baker Hughes "advising on integrated technology and digital solutions designed to improve production performance and recovery" [1]. Phase two executes them, and phase two carries the sentence this piece is about: "AI-enabled chemical injections designed to improve flow assurance", alongside well workovers and high-graded interventions [1]. Amerino Gatti, an executive vice president at Baker Hughes, and Ahmed Hayat Lak, managing director and chief executive of OGDC, are the named executives, and the trade write-up that carried the item further put it in the same terms [2].

Four things are missing from that page and all four matter. No production uplift figure. No split of the well count between the two fields. The phrases artificial intelligence and machine learning appear zero times, the single AI mention being the chemical injection sentence above. And the operator has not spoken: this is Baker Hughes' release about Baker Hughes' own contract, dateline Houston and London, first-party for the vendor and not for OGDC.

Our own weekly sweep rejected the item on 4 September 2026 on two independent grounds: OGDC is not on the stream's operator list, and nothing AI-related has happened, because the release places the AI in an execution phase that begins only after an evaluation phase, which makes it a plan rather than an action. The re-verification on 5 September re-fetched the page, confirmed the dateline and the wording, and let the rejection stand. The following week's run admitted it anyway. We are keeping the rejection and writing on it rather than around it. What follows is not a report on a deployment. It is an argument about what has to be true inside phase one before phase two means anything, and that argument is ours.

Six documents, and what each one is

Only the OGDC contract attaches an AI claim to a named operator's own programme, and even that is the vendor's word for a phase that has not started. The asymmetry is the finding, so here are the other five with their provenance stated.

The deployment counter-case is operator-authored and not in this window at all. SPE 219553, "Gas Lift Optimization in the Permian Using Machine Learning and Artificial Intelligence", with ExxonMobil authors, was presented at the SPE Artificial Lift Conference and Exhibition Americas in The Woodlands, Texas, in August 2024 [3]. It reports a closed-loop, well-by-well gas lift workflow deployed to more than 1,300 ExxonMobil Permian wells for a production uplift of about 2.2%, with no change to surface or downhole equipment. We reached it through a Journal of Petroleum Technology survey of artificial lift automation published on 4 September 2026 [4], which is a pointer and not an operator statement: JPT is the Society of Petroleum Engineers' own journal writing up what operators presented, and its restatement carries the same result [5]. Whether about 2.2% is distinguishable from noise is a question a sibling piece owns; here the figure serves one purpose, as evidence of what a deployment is able to report.

The denominator comes from SLB. A question and answer with Paul Sims, president of Production Systems, published on 8 September 2026, puts approximately 70% of global production as coming from assets past peak, expects that to approach 80% by 2030, and puts conventional reserve replenishment close to 30% [6]. Those are SLB's figures, sourced on the page to a recent analysis it does not name, and the page is a position piece: neither a deployment nor an intent, making no AI claim of its own. The size of the problem, not a finding.

Two SPE panels sit in the same three weeks and neither carries a single quantity. The first was held at the SPE Artificial Lift Conference and Exhibition on 27 August 2026 and written up by JPT on 8 September [7]: Kreg Flowers, a production engineer at Occidental, said the foundational skills have not changed and warned that AI can create false expertise; Maggy Burns, an operations supervisor for produced water at ExxonMobil, argued for input data quality, domain knowledge and human oversight on safety-critical decisions; Araceli Rivera Mandujano of SLB warned about information overload. Those are operator employees on a panel, which is not an operator-published account, and the date belonging to the substance is the panel's, 27 August, not the article's. The second is the SPE Subsea Well Intervention Symposium in Galveston, Texas, reported by JPT on 10 September [8], where Sean McCall, Chief Data Officer at Oceaneering, and Brian Saucier, President of DeepMar Consulting, argued that data quality has to be assessed and improved before AI is deployed against it, with Dana Otillio of SPE alongside them and Aramco Americas named in the piece. A vendor Chief Data Officer and a consultant: no operator engineer is quoted there at all.

The last item is the control. On 4 September 2026 Baker Hughes announced an offshore production enhancement and stimulation services contract from bp across bp's UK North Sea operations, covering new well development and recovery from mature fields, using the StimFORCE modular package from vessels supported by a UK operating base [9]. No contract value, no well count, no duration, no named asset, and the release does not mention artificial intelligence, machine learning, digital, automation, remote operations or software anywhere on the page. Neither bp nor Baker Hughes has claimed anything about AI there and nothing here should be read as saying they did. It is in the pool because it is the shape of a mature-field contract with the AI wording removed, and the completeness question put to it below is ours, not theirs.

So: two panels of the industry's own people naming data quality as the constraint and unable to price it, one position piece supplying a denominator, one deployment from 2024 reached through a pointer, one contract whose AI phase has not begun, and one with no AI in it at all.

A roster is not a working set

A contract naming 120 wells has counted a request. What an AI-enabled injection programme can be fitted on is the subset of those wells that arrive carrying every record the method consumes, and on our own programmes the two numbers have differed by a lot.

To recommend a dose rate for a scale or corrosion inhibitor and then learn from what happened, a model needs the injection history for that well, meaning what was dosed, when, at what rate and at which point; a fluid analysis, meaning water cut, salinity, scaling indices and souring indicators; an allocated production record at a cadence finer than the response it is trying to detect; and an inspection or sample record saying what actually happened downhole, because that is the label. Four records, and the well is usable only if all four are there for the same well over the same period.

That last clause is where the arithmetic turns. A missing record is not a missing feature the other three can stand in for. A well with no inspection record is not a noisy example, it is an unlabelled one. A well with an injection history but no allocated production has a treatment with no outcome attached. A well whose fluid analysis predates the last workover is describing a different well. Averaging repairs none of that, and imputing the label from the features is the one thing a supervised method must never do. Completeness per well behaves like a conjunction, and conjunctions multiply.

The gas lift counter-case is the same industry, the same estate and the opposite structure. A closed-loop gas lift optimiser consumes the injection rate, which is a valve setting the lift system already commands and logs; the wellhead and casing pressures, which it already measures; and the allocated rate, which exists because the well is already allocated. Control variable and response sit on the same instrument, continuously, because the operation cannot run without them. The count of extra records that workflow had to create was 0, its roster and its working set were therefore the same wells, and that identity, rather than model quality, is what made it a deployment able to report more than 1,300 wells and about 2.2% [3][4].

Flow-assurance chemistry has none of that for free. The dose rate is a pump setting rather than a measured injection into a modelled flow path, and the response is scale, corrosion or emulsion behaviour observed when a technician pulls a sample on a route, which is an irregular manual schedule and not a telemetry stream. The AI-enabled half of a chemistry programme has to create the measurement record that gas lift was handed by its own operation. That is why the AI sits in phase two behind an evaluation phase carrying no figure, and why two SPE panels in the same three weeks can name data quality as the constraint [7][8] without being able to put a number on it.

The arithmetic

Take a roster of $N$ wells, and let the AI step require $k$ distinct records from each. Let $p$ be the per-record delivery rate: the probability that any one required record turns up, for a given well, complete enough to use. Treat the records as independent. Then the wells that arrive complete are

and the rest follows from where $k$ sits in that expression. It is an exponent, and $N$ is a multiplier. Raising the per-record rate from $p_0$ to $p_1$ multiplies the working set by

At 4 required records, going from 60% to 80% per record is a factor of $(4/3)^4 = 256/81 \approx 3.16$. Buying a second field of the same size is a factor of exactly 2, at any $k$ and any $p$. So the two levers can be priced against each other: the per-record delivery that matches multiplying the roster by $m$ is

which at $m=2$, $k=4$ and $p=60\%$ is 71.35%. A rise of 11.35 points in the rate at which one record turns up complete is worth as much as a second 120-well field.

One more expression, because it is how our measured phases reach the same axis. A programme that asked for $a$ wells, received $d$ usable ones and required $k$ records from each implies a per-record rate of

The counts there are measured. The rate is inferred, and it is labelled that way on the instrument.

What our own ledger recorded

We have run this gate and wrote down what came through. Over three phases of borehole-image interpretation for a mid-sized Middle East carbonate operator we kept a two-column ledger, asked against delivered, where delivered meant a well that cleared a defined intake gate rather than one that merely arrived [10].

Phase 1 recorded 5 asked, 0 delivered. All 5 arrived and all 5 were rejected at intake before a single model saw them, because none carried the apparent dip and azimuth picks the detector needed as ground truth [10]. One required column, missing across the whole delivery, and the phase was a zero.

Phase 2 recorded 10 asked, 8 delivered. Two did not clear QC: one shipped a log container the reader could not parse into a clean image, and one was logged with a compact micro-imager where a standard formation micro-imager was needed [10].

Phase 3 recorded 25 asked, 8 delivered, which the file calls the single largest entry in the ledger and the one we reported on hardest [10]. The other 17 were proposed, tracked line by line and never delivered as annotated, ground-truth-bearing wells inside the programme window. Two smaller entries set the scale of a feature-level ask: vug ground truth, total vug percent per 10 cm zone, came from just 2 wells, and 5 horizontal wells arrived late as an unplanned add-on, useful for validating inclination and never part of the original ask [10].

The line the argument turns on is the governance point the file makes itself: read the ledger on delivered wells alone and Phase 2 and Phase 3 land on exactly the same working-set size, 8 wells, despite Phase 3 asking for 2.5 times as many [10]. The ask went from 10 to 25. The working set did not move.

Those counts come with a bound. This is one confidential carbonate borehole-image engagement in Oman, with that programme's own intake and QC gates. It is not a benchmark for how any other operator's data supply will behave, and it was borehole-image interpretation, not a flow-assurance chemistry programme. The law transfers. The numbers do not.

The exhibit

The instrument below is not a chart of any of this. It is the object: a shaft with a stack of perforated planes across it and 120 individual wells released at the top, one solid token per well, each well keeping its own cell on every plane. Each plane is one record the AI step requires, and a plane delivering a per-record rate of $p$ has $p$ of its cells open, to the nearest whole cell, in a different place on every plane. A well that meets a shut cell comes to rest on it as a countable object; a well that clears every plane lands in the working-set tray on the floor, where the survivors can be counted one at a time, and looking straight down the shaft they are the only cells open on every plane. Nothing in it is a height and no quantity is read off a slope. The roster is 120 because the Baker Hughes release names more than 120 wells and 120 is the floor of the vendor's own count: their number, not ours and not OGDC's, and the plate says so.

What the exhibit shows

It opens at 4 required records and 60% per-record delivery. 15 of the 120 wells reach the tray, which is 12.96% of the roster, and the heaps say where the rest went: 48 stop at the first gate, 29 at the second, 18 at the third and 10 at the fourth. That first heap is the one to look at, because 48 wells is more than three times the working set, lost at a single record.

Drag the delivery slider to 80% and 49 wells reach the tray, which is 40.96%. The shares carry the exact factor the arithmetic predicts, 3.16, and the whole-well counts read 15 and 49 because a partial well is not a well. At 60% each additional percentage point of per-record delivery is worth about 1.04 wells; at 80% the same point is worth 2.46, because the slope of an exponent rises with the level.

Now press "double the roster". A second, dimmed cohort of 120 falls into the same shaft, the heaps roughly double, and the tray goes from 15 to 31, because 240 wells at 12.96% is 31.1. That is the linear half of the contract doing everything it can do. Against it the matching-aperture readout prints 71.35%: raising the per-record rate by 11.35 points on the original 120 wells leaves the same 31 in the tray as buying 120 more wells. One of those is a procurement decision with a field attached. The other is an evaluation-phase decision that carries no figure in the release.

The records slider is the exponent on its own. At 60% per record, 1 required record leaves 72 wells of 120 and 6 leave 5. Nothing about the wells changed between those states and nothing about the delivery rate changed either. Only the number of things each well has to carry.

Two toggles set the boundaries. "Instrumented by the operation itself" removes every plane and all 120 wells fall to the tray: roster and working set become the same objects. That is the gas lift geometry, read from SPE 219553 and not our measurement, and it is what an extra-record count of 0 looks like as a picture. "One required column missing estate-wide" shuts the first gate and the entire roster stops on it wherever the other controls sit. That is the shape of our Phase 1: 120 wells piled on one closed gate.

The three ledger buttons drop our measured phases at their real sizes, and they override the sliders because they are counts we recorded rather than a state we model. Phase 1 releases 5 wells against a shut gate: 5 on the plane, 0 in the tray. Phase 2 releases 10 at one gate: 8 in the tray, 2 in the heap, an implied per-record rate of 80.00%. Phase 3 releases 25 behind 3 gates: 8 in the tray, 17 in the heaps, a through-rate of 32% and an implied rate of 68.40%. Press Phase 2, then Phase 3, and watch what does not change. The trays hold the same 8 objects while one shaft released 10 wells and the other 25. The heaps are not the same. The working set is.

Where this model is weakest

Independence is the assumption doing the most work and it is ours, not a measurement. A well missing its injection history is likelier than average to be missing its allocated production too, because both usually come from the same neglected corner of the same system. Positive correlation between the gaps makes the real curve less steep than $p^k$ in the middle of the range, since the missing records concentrate on fewer wells. It does not rescue the tail, and it does not change the ordering of the two levers, because the exponent still lives on the aperture and never on the roster.

The exhibit also opens every plane to the same aperture, and real record sets are not like that: an allocated production record is usually easier to obtain than a downhole inspection, and the binding gate is generally one specific record rather than the average of four. The shut-gate toggle is the honest version of that case and our Phase 1 is its measured instance.

4 is our reading of what a chemistry programme consumes, not a figure from any of the six documents. None of them names the records a model would consume, which is itself the observation, and the slider exists so a reader can substitute their own count. And none of our figures says anything about OGDC's data: nobody outside the contract knows what the Tando Alam Oil Complex or the Pirkoh field hold, because no figure attached to phase one has been published by anyone.

What a buyer should ask before phase two is priced

How many distinct records does the AI step consume per well, and what is the delivery rate of the worst one? Not the average rate, and not the count of wells in scope. The worst single record sets the ceiling, and on our Phase 1 it was the only one that mattered.

What is the working set today, as a count of named wells carrying all of those records for the same period? A programme that can answer has done its evaluation phase properly. A programme that answers with the roster has counted the request.

What does phase one buy, in per-record delivery, and what does it cost? The evaluation phase in a contract of this shape is where the aperture is opened, and it is the half carrying no figure in the release. On a 120-well roster at 4 records, taking per-record delivery from 60% to 71.35% is worth as much as a second field of the same size, and taking it to 84.09% leaves half the roster usable.

Which required records will the operation produce anyway, and which have to be created for the model? That separates a gas lift geometry from a chemistry geometry and it is answerable before anything is built. Records the operation already produces never fall out of the working set. Records that exist only because the AI step needs them have to be scheduled, staffed and gated, and each one is another exponent on the count of wells that arrive complete.

An operator who wants that per-well record set assembled and gated before an AI phase is costed is describing the work EarthScan 360 exists to do, with the intake and QC layer underneath it on the platform overview. Neither is the point here. The point is that a number was published for the roster and none for the gate, and it is the gate that decides how many wells the AI ever sees.

References

[1] Baker Hughes. Baker Hughes Collaborates with OGDC to Optimize Existing Well Production, Supporting Energy Security in Pakistan. 2 September 2026, dateline Houston and London. Vendor first-party: Baker Hughes' own account of its own contract, naming OGDC as the operator and not co-issued with OGDC. The same release is served at a shorter path ending baker-hughes-collaborates-ogdc-optimize-existing-well-production. https://www.bakerhughes.com/company/news/baker-hughes-collaborates-ogdc-optimize-existing-well-production-supporting-energy

[2] Energy Connects. Baker Hughes secures multi-year OGDC deal to boost Pakistan oil and gas output. September 2026. Trade coverage of the release at [1], not an operator statement. https://www.energyconnects.com/news/oil/2026/september/baker-hughes-secures-multi-year-ogdc-deal-to-boost-pakistan-oil-and-gas-output

[3] 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 dated 2024: the work is older than the coverage that surfaced it. https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/550379

[4] Journal of Petroleum Technology. Maximizing Every Barrel: Artificial Lift Embraces Automation. 4 September 2026. Pointer only: SPE's own journal writing up what operators presented, not an operator statement. https://jpt.spe.org/maximizing-every-barrel-artificial-lift-embraces-automation

[5] Journal of Petroleum Technology. Restatement of the SPE 219553 gas lift result. JPT, volume 77, issue 10. https://onepetro.org/JPT/article-abstract/77/10/1/791497

[6] 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% now, 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

[7] 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: JPT reporting operator employees on a panel, not an operator-published account. The article carries no quantities. https://jpt.spe.org/new-tools-old-lessons-experts-stress-the-importance-of-artificial-lift-basics

[8] 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 in Galveston, Texas. Pointer: the panellists quoted are a vendor Chief Data Officer and a consultant, and no operator engineer is quoted. The article carries no quantities. https://jpt.spe.org/industry-executives-offer-practical-advice-on-how-to-use-modern-ai-models-wisely

[9] Baker Hughes. Baker Hughes Awarded Offshore Production Enhancement and Stimulation Services Contract by bp. 4 September 2026, dateline Houston and London. Baker Hughes speaking about bp, not bp speaking. The release does not mention artificial intelligence, machine learning, digital, automation, remote operations or software anywhere; cited here as a labelled no-AI control. https://investors.bakerhughes.com/news/press-releases/news-details/2026/Baker-Hughes-Awarded-Offshore-Production-Enhancement-and-Stimulation-Services-Contract-by-bp/default.aspx

[10] EarthScan. The Attrition Ledger: How We Ran a Subsurface-AI Program on One Number, Asked-Versus-Delivered. Source file content/case-studies/25-wells-asked-for-8-delivered-the-data-scarcity-gap.mdx. Every figure of ours in this piece is quoted from it. https://www.earthscan.io/case-studies/25-wells-asked-for-8-delivered-the-data-scarcity-gap