# Oil & gas operator AI failures: evidence screen

**Research date:** 4 October 2026. **Window:** 4 Oct 2024–4 Oct 2026; older examples only as historical context. **Conclusion:** This search did not establish a strong, named case in-window of an oil/gas operator abandoning, rolling back, or materially missing expectations for a specifically identified AI project, with a documented stop decision and reason. This is a negative search result, not evidence that no such cases exist. No operator case is counted merely because ROI is undisclosed or a pilot lacks a public scale-up announcement.

## Findings

### No in-window case met the evidence bar

Searches of operator disclosures, press releases, regulator/government records, technical literature, and targeted web queries for named AI projects combined with terms such as *abandoned, cancelled, paused, stopped, rollback, failure,* and *missed expectations* surfaced announced pilots and deployments, but no adequately documented termination or material miss. Examples that surfaced as live/positive projects (and thus are not failures) include:

- **Serica Energy / Rahd AI well decommissioning pilot** (2024 ESG report): six legacy wells; report describes a pilot examining digitized records and AI modelling as a future planning baseline. It does not report cancellation, failure, costs, or results. [Serica Energy ESG Report 2024](https://www.serica-energy.com/downloads/ESG/ESG-Report-2024.pdf). Confidence: high for what the report says; no inference of success/failure.
- **ADNOC / AIQ ENERGYai**: announced completion of a 90-day proof-of-concept in Jan. 2025 and positive results. This is not evidence of failure; public announcement does not establish later deployment economics. [ADNOC release](https://adnoc.ae/en/news-and-media/press-releases/2024/adnoc-and-aiq-successfully-complete-trial-phase-of-agentic-ai-solution). Confidence: high.
- **Saudi Aramco / Yokogawa Fadhili Gas Plant AI**: Aramco says rollout occurred Oct. 2024–Apr. 2025, with evaluations still ongoing and early measured usage reductions. Not a failure. [Aramco AI and Big Data](https://www.aramco.com/en/what-we-do/energy-innovation/digitalization/ai-and-big-data). Confidence: high.

These examples demonstrate why “no public ROI” or a pilot-stage disclosure cannot properly be coded as failure.

### Anonymous operator anecdote: excluded

An AppIntel post claims an unnamed, now-defunct operator terminated an in-house AI regulatory-application system because it did not save time or improve results, and says the work was spun off. It gives no operator name, dates, project cost, underlying records, or verifiable corroboration. This is a potentially relevant lead, not a citable named case; excluded from findings. [AppIntel, “In-house AI attempts fail-80%”](https://www.appintel.info/2026/in-house-ai-attempts-fail/). Confidence: low; claims unverified.

### Historical context outside window: Google stopped taking new upstream AI custom-development work (2020)

Google Cloud said it would no longer build custom AI/ML algorithms to facilitate upstream extraction; contemporaneous reporting noted the existing Total (now TotalEnergies) agreement remained in place at that time. This was a **supplier policy decision**, not evidence that Total or another operator cancelled a failed project, and it predates the research window. It should not be presented as an operator AI failure. [Axios report, 19 May 2020](https://www.axios.com/2020/05/19/google-oil-microsoft-amazon-artificial-intelligence). Confidence: medium-high for the policy announcement; no operator failure established.

## Search scope and limits

Conducted 11 targeted searches covering operator AI projects “discontinued/abandoned/cancelled/paused,” named-company terms, regulator/government records, and operator disclosures. Followed up on leads from company reports/releases and screened out anonymized vendor anecdotes and generic claims that AI pilots fail. Primary-source review included the Serica and Aramco operator pages and ADNOC release; SEC-hosted Shell and Equinor annual-report searches did not surface a named AI project termination. Search indexing is incomplete; private project decisions, internal postmortems, and non-public budget/ROI data are not observable here. No costs, stop rationale, or quantified misses could be substantiated for any qualifying named operator case.

## Focused technical-literature follow-up (8 searches)

**Searches run:** (1) `site:onepetro.org Devon Energy artificial intelligence machine learning pilot false alarms optimization`; (2) same for EQT; (3) same for ConocoPhillips, with model generalization/deployment; (4) same for ExxonMobil; (5) `site:onepetro.org/SPE Shell machine learning false alarms field deployment`; (6) same for Equinor, generalization/pilot operationalization; (7) same for ConocoPhillips, limitations; (8) same for Devon Energy, workflow deployment. Searched OnePetro/SPE-indexed technical publications, prioritizing the requested operators. Results were mostly positive deployment accounts or modeling papers, not postmortems. No qualifying discontinued optimization or failed named operator trial surfaced.

### Useful concrete limitations, but not abandonment evidence

- **ExxonMobil/XTO, Permian gas-lift optimization (SPE-219553-MS, Aug. 2024).** The paper describes the opposite of a failed pilot: closed-loop optimization deployed on >1,300 wells, with reported ~2% average oil uplift. Its technical boundary conditions are useful: production data undergoes rigorous QC; wells lacking downhole pressure gauges rely on an ML recommendation model; uplift is tracked through a model-based approach grounded in productivity and steady-state well tests. A JPT article on the same work says deployment proceeded in two stages, with the first three months requiring manual approval of setpoints, then describes autonomous setpoint implementation. This shows a staged trust/automation gate, not poor adoption or project failure. [OnePetro abstract](https://onepetro.org/SPEALCE/proceedings-abstract/24ALCE/24ALCE/D031S008R001/550379); [JPT account](https://onepetro.org/JPT/article-abstract/76/10/30/569510/Unsung-Hero-Artificial-Lift-Boosts-Production?redirectedFrom=PDF). Confidence: high for reported scope/design; outcomes are operator-authored/publication-reported, not independently audited. The conference date is Aug. 2024, just outside the 2-year window by publication/event date.

- **Shell India, ML prediction of multiphase flow (SPE-214348-MS, June 2023).** This is a research/model-development paper, not evidence of a cancelled Shell operations deployment. Models were trained on representative lab data / synthetic datasets generated from Shell Flow Explorer and OLGA; the paper describes transfer learning to experimental data as a next step and says integration with live data could be future work. Thus operational generalization and field adoption were not established in that publication. Excluded as a failure: planned next steps and lab-to-field gaps do not show the model failed or was abandoned. [OnePetro abstract](https://onepetro.org/SPEEURO/proceedings-abstract/23EURO/23EURO/D021S002R002/519987). Confidence: high for the paper's stated research scope; no evidence found on subsequent deployment status.

- **Equinor-affiliated field-development DRL paper (SPE-203951-PA, 2022).** The paper explicitly notes conventional optimization solutions can be scenario-specific and fail to generalize to unseen earth models or price assumptions; its own DRL is trained across a large simulated range and reports preliminary application to a real field. It is a useful statement of the generalization problem and a positive research result, but predates the main window and does not document a failed Equinor production system. [OnePetro paper summary](https://onepetro.org/SJ/article-abstract/27/01/226/465130/Deep-Reinforcement-Learning-for-Generalizable?redirectedFrom=fulltext). Confidence: high for the paper's claims; no failure inference.

### Operator leads screened out

- The Devon-targeted OnePetro search returned a 2024 Devon paper on stacked-pay drainage and development optimization using seismic inversion, completions diagnostics, and geochemistry—not an AI failure or AI model trial. [URTeC abstract](https://onepetro.org/URTECONF/proceedings-abstract/24URTC/24URTC/D021S029R001/551024) (search result surfaced under Devon query; no AI failure content established).
- Targeted EQT and ConocoPhillips searches did not return a directly attributable SPE/OnePetro postmortem or discontinued AI optimization trial. Results were generic or unrelated; no attributable named case added.
- Additional OnePetro/SPE searches returned successful/current AI studies, including a 2026 74-well Bakken rod-lift pilot and a 2024 offshore condensate RVP model with a described data-conditioning issue (disturbance periods should be filtered; its selected deployment was a real-time visualization tool). The latter is a model-quality / feature-engineering limitation, not a miss: its abstract describes a proposed successful prediction technique and deployment option. [2026 rod-lift pilot](https://onepetro.org/SPEOGWA/proceedings-abstract/26OPES/26OPES/D011S009R002/799204); [RVP study](https://onepetro.org/OTCASIA/proceedings-abstract/24OTCA/24OTCA/D031S030R003/542028). Not counted as failures.

**Bottom line of this follow-up:** primary technical literature adds specific operational constraints (data QC, gauge availability, manual setpoint approval, disturbance filtering, simulated/lab training domain, and explicit scenario-generalization limits), but the focused searches did not find a named Devon, EQT, ConocoPhillips, ExxonMobil, Shell, or Equinor project shown to have been abandoned or materially missed expectations. Costs, shutdown dates, false-alarm burdens, or stop rationales remain unestablished for these leads.

## 16 October follow-up: reported use-case non-retention

Control Global quotes TotalEnergies engineer Thomas Klein saying a Port Arthur drum-level model achieved 75% prediction accuracy, but the drum-cycling use case was not retained. Drum switching depended on conditions outside the model’s scope. The work provided advice, not autonomous control. Original conference reporting, labeled Sponsored. We could not retrieve the presentation or a later company statement. The earlier company release confirms the broader pilot, not the decision to drop this use case. Keep these figures separate from its five-event result.

[Honeywell AI pilot aids coker unit operations at TotalEnergies refinery](https://www.controlglobal.com/show-coverage/honeywell-users-group/article/55383668/honeywell-ai-pilot-aids-coker-unit-operations-at-totalenergies-refinery) · 2026-06-11 · Drum cycling prediction section; Thomas Klein remarks at Honeywell Users Group Americas 2026

This is original conference reporting of named remarks, not an issuer-authored cancellation notice or independent technical validation. No program-level shutdown or financial loss is established.

## 22 October: what the EY survey can tell us

EY’s March 2025 survey is now indexed in the [report library](../?research=library#research). It was already in our October 5 research; this is not another independent setback case.

27% of surveyed executives said AI was meeting expectations. 85% reported difficulty scaling technology pilots. This covers digital technology broadly. Online survey of 508 employees. Technology questions went to directors and above; exact item bases are undisclosed. Pooled sectors; no O&G-only result, project failure rate, costs or measured cash returns. Total sample 508 is not the denominator for these executive-only items. EY sells consulting services.

[Original EY report](https://www.ey.com/content/dam/ey-unified-site/ey-com/en-us/campaigns/energy-resources/documents/ey-ogc-future-of-energy-external-report.pdf) · Printed pp. 7–8: expectations and scaling; pp. 27–29: methodology and profile (PDF indices 6–7, 26–28).

## 24 October: false alarms and incomplete oversight

These are operational lessons, not two new failed deployments. The MARS team changed its design and reached operational use; Havtil reports oversight gaps without naming a cancelled project.

### False alarms changed what the team could put into use

MARS researchers found that testing on small image chips hid the false-alarm burden in full satellite scenes. On the same 200-granule EMIT test, a single-model variant averaged 3,565 false alarms, 213 detected events and 98 misses across five runs. Its five-model ensemble produced 1,501 false alarms, 207 detections and 104 misses. This is a trade-off, not error-free detection. These are test counts, not a live alert rate or measured labor savings. The paper reports 11 months of operational use; analysts verify candidate plumes before notifications. No full-cost ROI is disclosed. Preprint v2: 24 March 2026; first submitted 11 November 2025. Test uses 105 granules with known plumes and 95 without, containing 311 labeled events. Operational start dates differ by sensor. A successful redesign, not a cancelled operator project. Results span emissions sectors; no new named oil-and-gas operator adoption is established. Raw data and code were not rerun. A separate 17 October conference abstract reports a similar benchmark-to-operation gap; its full chapter was unavailable. Before scaling, how many false alarms and missed events will a team handle on the full live workload?

[Operational machine learning for remote spectroscopic detection of CH4 point sources](https://arxiv.org/html/2511.07719v2) · 2026-03-24 · Version 2, 24 March 2026; sections 3.1, 5.1, 6.2 and Table 2; section 6.4 and Table 5. · Editor read original full-text methods, Table 2 and operational workflow. Author preprint, listed as in review on arXiv; raw data/code not replicated. First version 11 November 2025.
[Reducing False Positives in Operational Remote Sensing: Lessons from Methane Plume Detection](https://doi.org/10.1007/978-3-032-41086-3_46) · 2026-10-17 · Publisher abstract and First Online metadata; chapter pp. 721–736. Full subscription chapter not reviewed. · Editor read original publisher abstract and publication metadata. Subscription full chapter, experimental tables and code not obtained; no purchase or bypass.

### Some operators cannot yet account for all the AI they use

Norway’s petroleum regulator says operators vary widely in how they oversee AI. Many lack a complete inventory, including AI inside supplier products. Some companies reported no AI with safety implications while describing AI in well planning, maintenance or operational advice. Havtil questions whether indirect effects on decisions are being recognized. No respondent count, project denominator, incident rate, cost or return is disclosed on the reviewed page. Published 23 October 2026. Survey fieldwork dates and responding companies are not identified on the page; May 2025 guidance is earlier context. This is a regulator’s interpretation of company responses, not proof of a failed deployment, an enforcement finding against every operator or measured poor ROI. It does not rank companies. Can the operating team name every AI system it relies on, including supplier features, and who owns its review?

[Artificial intelligence (AI) in the petroleum industry is in need of better oversight and management](https://www.havtil.no/en/explore-technical-subjects2/technical-competence/reports-from-projects/2026/artificial-intelligence-ai-in-the-petroleum-industry-is-in-need-of-better-oversight-and-management/) · 2026-10-23 · Varying levels of maturity; An overview is essential for effective risk management; Prudent development and use. · Editor read original regulator publication and date. Underlying responses, named companies, response count and questionnaire not published on the reviewed page.

## 24 October: five SOL rounds on value, extraction and retrieval

The West Brae account is an undated historical trial with a site-specific limited-value assessment. The other two notes are controlled tests, not evidence of cancelled operator AI systems.

### West Brae’s model worked, but the value case was weak

NZTC says Intelligent Plant’s ML tool recommended valve settings at Marathon’s West Brae, but appeared to add little value on that already optimized field. The eight-well field yielded five usable verification points in four weeks and 12 in a two-year retrospective. Stable operation and larger unrelated production changes made small gains hard to establish. The listed 1% potential and £26.2 million industry-value scenario at 10% uptake are not realized trial benefits or net ROI. Trial costs are undisclosed. Undated historical account. Exact trial, publication and end dates unknown; not a new deployment in the rolling 24-month window. No cancellation, negative net cash return or later scale is established. Intelligent Plant has an undated anonymous account with similar trial details that claims benefits; we cannot confirm identical case identity or reconcile it with NZTC’s named assessment. Is there enough improvement left in the chosen asset to measure and pay for this tool?

[Well Intelligence – Optimised Production](https://www.netzerotc.com/case-studies/well-intelligence-optimised-production/) · Unknown · Project overview; Industry value; Lessons learned. Undated historical account naming Marathon and West Brae. · Editor read original project summary and lessons. Exact trial/publication dates unknown; website copyright is not a date. Project assessment, not independently audited cash returns.
[Case Study: Enhancing Production with Well Intelligence AI Model](https://www.intelligentplant.com/case-study-well-intelligence) · Unknown · Challenge and Testing; Historical Data Review; Results and Impact. Anonymous client, no publication date. · Editor read original supplier case. Anonymous account shares four-week/~8,000-bbl/day descriptors with NZTC account; same-trial identity not confirmed. No measured financial result or operator corroboration.

### A well-record model picked the wrong depth field

A 2024 study shows why plausible extraction needs field-level checks. With a simple prompt, Llama 2 70B returned depths within 10 ft for 72 of 150 clean Colorado reports. Revised prompts reached 150/150 on the same corpus. On 10 Pennsylvania scanned records, its best prompt located 7/10 wells within 10 m and extracted 9/10 depths within 10 ft. Errors involved source/OCR signs, coordinate conversion and measured depth versus true vertical depth. The paper also tested Llama 3.1 405B on those 10 scans: 9/10 locations and 10/10 depths. These are threshold-based test results, not a current-model ceiling, exact transcription score or live error rate. Published 30 December 2024. Historical documents and 2024-era models; no prospective deployment date. Unequal 150- and 10-document cohorts; no independent untouched prompt-selection holdout established. Do not attribute all errors to the LLM, or infer failed operator rollout, measured savings or net ROI. Does the test check the right depth, coordinates, units and source quality before values enter an operating record?

[Information extraction from historical well records using a large language model](https://www.nature.com/articles/s41598-024-81846-5) · 2024-12-30 · Methods: LLMs performance evaluation; Results: Prompt engineering, Tables 2–3, chain-of-thought and comparison with other LLMs. DOI 10.1038/s41598-024-81846-5. · Editor read original open-access publisher full text, metric definitions and results. Paper independently cross-checked by two SOL tracks; data/code not rerun. Peer-reviewed publication date 30 December 2024; tests use 2024-era models, not current capability.

### Text search missed information in engineering diagrams

Authors affiliated with Nigeria LNG tested four text-retrieval methods on 60 proprietary oil-and-gas documents: 20 narrative, 20 tabular and 20 piping and instrumentation diagrams. All four struggled with the diagrams. The 60 LLM-assisted queries were human-reviewed; people judged pooled search results. On diagrams, the best mean reciprocal rank was 0.106. That ranks the first relevant passage; it is not 10.6% answer accuracy. Across the full corpus, preserving structure led on MRR/NDCG, while recursive splitting led some MAP/F1 measures. No method won every measure. No cash return or final-answer correctness was measured. Author manuscript posted 25 March 2026; proceedings marked 2026, exact issue day unknown. Small proprietary corpus; diagram-query count and annotator agreement undisclosed. No live-user, hallucination or adoption rate. Employer affiliation does not establish that these tested configurations were deployed at Nigeria LNG. Can the system retrieve the right information from drawings and tables, as well as prose, before anyone relies on its answers?

[Evaluating Chunking Strategies For Retrieval-Augmented Generation in Oil and Gas Enterprise Documents](https://arxiv.org/pdf/2603.24556) · 2026-03-25 · Original 11-page PDF: methods PDF pp. 4–7; Tables 1–3 PDF pp. 8–9, printed pp. 64–65. arXiv first submitted 25 March 2026; exact proceedings publication day unknown. · Editor read original author PDF and visually inspected Tables 1–3. arXiv says manuscript matches CCSEIT 2026 proceedings; raw proprietary corpus and relevance labels unavailable. Author affiliation does not establish deployment or corpus ownership.