A better model still has to make the job better.

Perenco reports faster drilling across three wells

The work

SLB reports a 19% increase in rate of penetration in Perenco’s 2025 offshore Congo campaign.

How it works

DrillOps and Neuro automated drilling and directional work; SLB describes an automatic safe-state response to a pack-off event. General override and approval rules are not supplied.

What was reported

Three wells; 97% of on-bottom operations automated. These are speed and control measures, not cash return.

When

2025 three-well campaign. Exact publication day and campaign dates unknown.

What remains unknown

Raw drilling-speed baseline, matched offset population, complete costs and AI-only contribution unknown.

What to ask next

Did the gain hold on more wells after accounting for geology, equipment and full deployment costs?

Perenco enhances drilling efficiency with advanced automation solutions
Unknown · Summary and Results; three-well 2025 campaign.
Editor read the relevant original publisher page on 25 October 2026. Results remain source-reported; no independent outcome validation.

Invoice handling fell from about five minutes to under one

The work

SpendConsole reports invoice processing falling from about five minutes to under one in automated scenarios at Wafi Energy Pakistan.

How it works

AI document capture, rules and SAP workflows handle routine invoices; finance staff handle exceptions. Autonomous payment approval is not established.

What was reported

The vendor separately reports 50–70% productivity improvement and full go-live in about 12 weeks.

When

Publication and operating calendar dates unknown; results observed within weeks of go-live. Twelve weeks describes implementation, not the evaluation period.

What remains unknown

Invoice volume, automated-scenario share, precise measurement window and full costs unknown. This is the Pakistan Shell-brand licensee, not Shell-wide adoption.

What to ask next

What share of invoices qualifies, and how much staff time is actually released after exception handling?

From manual accounts payable to AI-powered finance in 12 weeks
Unknown · Client profile; solution; results; automated-scenario qualification.
Editor read the relevant original publisher page on 25 October 2026. Results remain source-reported; no independent outcome validation.

A chemical-treatment redesign cut consumption without identified AI

The work

Apollo reports 65% lower chemical consumption and $210,000 lower monthly H2S treatment cost at an unnamed southeastern New Mexico facility.

How it works

A flow-meter-linked controller, new chemistry and a relocated injection point changed treatment. Key operator personnel received online write access.

What was reported

Zero out-of-spec H2S events during a 30-day observation. The monthly cost claim is not annualized here.

When

Thirty-day observation; publication and operating calendar dates unknown.

What remains unknown

Population, throughput, cost calculation, installation costs and longer-term durability unknown. New chemistry and equipment prevent attributing the result to software alone.

What to ask next

How much could treatment redesign achieve before adding AI, and what does it cost to sustain?

HSV-01 chemical treatment case study
Unknown · Solution and Results; 30-day observation.
Editor read the relevant original publisher page on 25 October 2026. Results remain source-reported; no independent outcome validation.

Targeted chemical placement is another route to less waste

The work

A Weatherford-hosted technical-paper abstract reports 65% less chemical use per well and 52% lower operating expense in capillary-injection field trials.

How it works

Simulation-guided placement, a capillary string and an injection valve deliver chemicals to the treatment point. No AI is identified.

What was reported

Results are from the public abstract of SPE-227643-MS; the full paper was not reviewed.

When

Paper dated 16 September 2025. Trial dates and evaluation duration unknown.

What remains unknown

Operator, geography, well population, raw baseline, cost perimeter and installation costs unknown. This does not establish all-in ROI or scaled deployment.

What to ask next

Does targeted placement outperform surface optimization after installation and maintenance costs?

Maximizing chemical efficiency with rigless precision-based capillary injection
2025-09-16 · Public abstract, SPE-227643-MS; field-trial results.
Editor read the relevant original publisher page on 25 October 2026. Results remain source-reported; no independent outcome validation.

Ecopetrol reports IT gains while its wider AI program still faces data limits

The work

Ecopetrol says using AI in IT operations raised critical-service availability to 99.52% and generated more than US$1.5 million in savings.

How it works

The same page says the AI ecosystem maturity review found limits in data access and autonomy. A broader operating model with integrated AI remains a 2027 goal.

What was reported

Company-reported availability level and savings; starting availability, service population, cost perimeter and savings period are not specified. An adjacent 40% digital self-service resolution rate is separate and is not explicitly attributed to AIOps.

When

The report covers 2025; exact publication day and AIOps start/evaluation dates are unknown. The 2027 target is prospective.

What remains unknown

No all-in cost, cash reconciliation, comparison design or complete net ROI. The US$1.5 million-plus may overlap the existing AI-147 portfolio benefits; do not add them. Neither 160 digital initiatives nor more than 100 operating solutions is a count of AI deployments.

What to ask next

Which services improved, what changed against the prior baseline, and how much of the saving remains after running costs?

Ecopetrol 2025 Integrated Management Report — digital transformation and AIOps
2026 (exact publication day unknown) · Printed p.314 / physical PDF p.315 of 482, Transformación Digital. AIOps, self-service and AI maturity are separate passages.
Editor downloaded the original report, read the complete relevant page and visually inspected it on 25 October 2026. Reuses existing source SW6-004; earlier claim-level reviews remain separate.

YPF’s fuel-blending tool supplies estimates to refinery operators

The work

A Boston Consulting Group (BCG) project lead describes Ekons Matrix AI in use at YPF’s La Plata refinery in Argentina.

How it works

AI and process-simulation rules estimate the properties of streams being mixed into fuel. The estimates support control-room staff; autonomous control is not established.

What was reported

The supplier says blending execution time could fall by up to 50%. This is a potential maximum, without a measured achieved average, comparison cohort or observation period.

When

The interview describes application starting March–April 2026, about six months of implementation and continuing adjustments. A syndicated copy is dated 11 September 2026; the retrieved original does not display its exact date.

What remains unknown

Original interview with the supplier project lead, not an operator filing or independent result. No realized saving, quality/error measure, cost or net ROI. No stream, tank, blend, user or throughput denominator. Further rollout is an ambition. Separate from the earlier YPF cataloging and inspection profile.

What to ask next

Can YPF show achieved blend-time and quality results over enough batches to separate the tool from other operating changes?

YPF refinery blending — interview with BCG project lead Roberto Rodrigo
Unknown · Sections Del laboratorio a la operación en tiempo real and Argentina, el primer caso de implementación.
Editor read the original interview. The exact date is not displayed in its retrieved body; a syndicated copy is dated 11 September 2026. Original named supplier interview, not operator confirmation or independent performance validation.

Dixstone selects shared project and fleet software; go-live remains unverified

The work

IFS says Dixstone selected IFS Cloud to manage project finances and fleet maintenance.

How it works

The intended system brings budgets, schedules, resources and asset data together. The release proposes predictive maintenance for rigs, vessels and platforms.

What was reported

No measured operating result, cost or return. Ten countries and more than 2,200 professionals describe the stated company footprint, not active users, sites live or AI adoption.

When

Selection announcement dated 9 December 2025. Go-live and implementation milestones are not disclosed.

What remains unknown

Vendor announcement, not proof of operating AI or delivered savings. No model, asset rollout, approval boundary or net ROI. Dixstone is a Perenco affiliate consolidating offshore services; this is not a Perenco E&P field deployment.

What to ask next

Which assets and workflows are live, what still needs implementation, and where does AI actually change the work?

Dixstone selects IFS Cloud for offshore services
2025-12-09 · Dated release: selection, project controls and fleet maintenance paragraphs; About Dixstone separates affiliate status from Perenco E&P.
Editor read the original IFS release dated 9 December 2025. A vendor selection announcement with a customer quote; future capabilities and company footprint are not completed rollout or outcomes.

PTTEP reports faster purchase-order processing with X.Bot

The work

PTTEP says Procurement X.Bot combines AI and automation in purchase-order processing.

How it works

The annual report describes improved processing, decision support and compliance. It does not identify model, vendor, user population or approval permissions.

What was reported

Operator reports five fewer processing hours per purchase order and 50% less pending-order backlog within six months. These are system-level process measures, not AI-only cash savings.

When

Report covers 2025; exact publication day, go-live and six-month comparison dates unknown. A 9 July 2025 release independently describes procurement AI use without naming X.Bot.

What remains unknown

No PO count, baseline hours, starting backlog, comparison design, implementation costs or cash return. AI and conventional automation are not isolated. Adjacent 172-supplier event, green-spend and vessel-fuel figures do not measure X.Bot.

What to ask next

Which steps changed, who approves orders, and do the time gains persist after all operating costs?

PTTEP 56-1 One Report 2025 — Procurement X.Bot
Unknown · Section 9.4.2, printed p.284 / physical PDF p.273 of 557. Procurement X.Bot paragraph; adjacent vessel and green-procurement figures are separate.
Editor downloaded original 557-page operator PDF, read and visually inspected the complete relevant page on 25 October 2026. Report covers 2025; exact publication date and six-month evaluation dates unknown. Operator claims, not independent outcomes.

PTTEP organizes OpEx Day 2025, advancing operational excellence
2025-07-09 · Press release, procurement AI sentence in operational-excellence examples.
Editor read original issuer release and displayed 9 July 2025 date. It supports procurement AI use, not X.Bot identity or its measured results.

BKV rebuilt its field-data system around how operators work

The work

BKV consolidated four inherited SCADA/CygNet environments. Its Barnett platform went live 19 November 2025.

How it works

Field staff reviewed terminology and screens. Peer-led training and field ride-alongs supported the change.

What was reported

BKV reports 2.5 million data points standardized, over 250,000 obsolete points removed, polling 60 to 15 minutes, and no production downtime. Acquisition integration reportedly fell from 6–8 months to 6–8 weeks.

When

Go-live 19 November 2025. Report letter is dated September 2026; exact report publication day unknown.

What remains unknown

Issuer-reported measures; no costs, controlled evaluation, cash benefit or AI use established. About 1,100 Bedrock wells were positioned for onboarding, not confirmed live at the go-live date.

What to ask next

Does the common data model survive the next acquisition, and how much ongoing support does it need?

BKV 2025 Sustainability Report — Project Runway
Unknown · Governance: Project Runway; Informed by field operations; Collaboration beyond implementation.
Editor read original operator page on 25 October 2026. Publication day not displayed; report landing page CEO letter dated September 2026. Metrics are issuer-reported.

BKV 2025 Sustainability Report — letter date
Unknown · Letter from CEO footer: September 2026.
Editor read original report landing page; letter month is not exact report publication date.

Harbour documents the exceptions in its supplier invoice workflow

The work

Harbour’s June 2025 Coupa guide covers purchase orders, confirmations, invoices and status tracking.

How it works

It documents access roles, drafts, receipt matching, disputes and credit/reissue paths. The guide points to ex-Wintershall Dea materials; a separate channel uses SAP Business Network.

What was reported

No measured time, error, cash saving or ROI result. Free supplier access does not establish the operator’s total cost.

When

Guide version 4, June 2025; exact day, system start and entity rollout unknown.

What remains unknown

Do not extend the guide’s all-new-PO wording to every Harbour entity or infer that Coupa replaced SAP. Exception instructions do not prove high failure rates.

What to ask next

Where do invoices stall, which exceptions require people, and which supplier channel applies?

Harbour Energy Coupa Supplier Portal FAQ, version 4
2025-06 · PDF p.1 version 4, June 2025 and PO process; pp.2–4 permissions, invoices, receipts and disputes; p.5 refers to ex-Wintershall Dea vendor guides.
Editor read original 5-page operator PDF. Exact revision day, go-live and legal-entity scope unknown. The linked vendor page separately retains SAP Business Network.

Harbour vendor relations — distinct supplier channels
Unknown · Separate sections: vendors using SAP Business Network; ex-Wintershall Dea vendors.
Editor read original operator page on 25 October 2026. Undated current channel guidance; SAP start November 2022 is older context, not new AI.

ConocoPhillips shows how suppliers enter orders and invoices

The work

ConocoPhillips’ December 2024 guide chose GEP SMART for supplier registration, PO acknowledgement and invoicing.

How it works

Suppliers register, accept PO terms, create invoices or service confirmations and inspect remittance. Current guidance offers portal, integration and Enverus submission routes.

What was reported

Faster approvals and better order accuracy are intended benefits; no measured result or ROI.

When

Updated December 2024; January 2025 transition target. Undated live guidance still uses that future-tense wording; completion by business unit unknown.

What remains unknown

No transaction population, costs or actual rollout completion. Supplier fee waivers do not establish operator cost. An available Enverus route is not exclusive-platform adoption.

What to ask next

Which units and invoice routes are live, and who owns the exceptions?

ConocoPhillips GEP SMART Introduction
2024-12 · PDF p.1 updated December 2024; p.3 transition target; p.5 intended benefits; p.6 supplier invoice flow.
Editor read original 7-page issuer PDF. January 2025 is a target, not verified unit-by-unit completion. Exact revision day unknown.

ConocoPhillips technical upgrades to supplier systems
Unknown · GEP portal/EDI and Enverus routes; supplier training and implemented SSIS support.
Editor read original live supplier page on 25 October 2026. Page retains future-tense January 2025 language, so it cannot verify completion date. No publication date displayed.

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

The work

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.

How it works

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.

What was reported

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.

When

Undated historical account. Exact trial, publication and end dates unknown; not a new deployment in the rolling 24-month window.

What remains unknown

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.

What to ask next

Is there enough improvement left in the chosen asset to measure and pay for this tool?

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
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.

Read the trial and the different accounts

A well-record model picked the wrong depth field

The work

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.

How it works

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.

What was reported

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.

When

Published 30 December 2024. Historical documents and 2024-era models; no prospective deployment date.

What remains unknown

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.

What to ask next

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
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.

Read the extraction test and sample limits

Text search missed information in engineering diagrams

The work

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.

How it works

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.

What was reported

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.

When

Author manuscript posted 25 March 2026; proceedings marked 2026, exact issue day unknown.

What remains 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.

What to ask next

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
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.

Read what the retrieval test measured

False alarms changed what the team could put into use

The work

MARS researchers found that testing on small image chips hid the false-alarm burden in full satellite scenes.

How it works

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.

What was reported

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.

When

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.

What remains unknown

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.

What to ask next

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
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
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.

Read the test and operating limits

Some operators cannot yet account for all the AI they use

The work

Norway’s petroleum regulator says operators vary widely in how they oversee AI. Many lack a complete inventory, including AI inside supplier products.

How it works

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.

What was reported

No respondent count, project denominator, incident rate, cost or return is disclosed on the reviewed page.

When

Published 23 October 2026. Survey fieldwork dates and responding companies are not identified on the page; May 2025 guidance is earlier context.

What remains unknown

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.

What to ask next

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
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.

Read the oversight findings

PDO has moved ESP prediction from pilot to operating use

The work

In its 21 April 2026 results, Weatherford reported that Petroleum Development Oman had moved ESP Predictive Analytics from pilot to operational use. The system runs within ForeSite Well Management System and combines it with PetroVisor’s machine-learning framework.

How it works

The on-premise installation provides failure forecasts, run-life estimates and real-time pump-performance insight while meeting the customer’s cybersecurity requirements. The release does not say whether forecasts trigger automatic changes or how people review alerts and approve maintenance.

What was reported

No pump or well count, predictive accuracy, warning lead time, run-life improvement, avoided downtime, cost or ROI is disclosed.

When

Disclosed 21 April 2026, found in this 24 October review. The release says operating use had begun; its exact start date and duration are unknown. This is missed evidence from April, not a newly announced October deployment.

What remains unknown

A named operating deployment is supported by Weatherford’s own statement. There is no independent PDO confirmation in this review, asset/basin or pump-population denominator, stated maintenance approval boundary or measured economic result. Regional expansion is a supplier aspiration, not established regional scale. Do not attribute Weatherford’s consolidated results to this use case.

What to ask next

Ask which pumps and assets are live, how forecasts are validated, who acts on alerts, and whether failures or maintenance costs improved after full implementation costs.

Weatherford Announces First Quarter 2026 Results — PDO ESP analytics
2026-04-21 · Technology Highlights → Production and Intervention; Oman / Petroleum Development Oman paragraph.
Editor read original issuer HTML and its displayed 21 April 2026 publication date on 24 October. Supplier statement; no independent operator confirmation. Exact operating start date unknown.

A rod-lift retrofit recovers energy on one U.S. well

The work

Weatherford reports the first commercial installation of its ForeSite Power Regenerative System retrofit on an existing U.S. rod-lift well. It captures and reuses energy generated during each pumping cycle.

How it works

The supplier says the retrofit avoids replacement of the installed drive. This is energy recovery equipment; the release does not identify AI or a learning model in this installation.

What was reported

Weatherford reports over 50% energy savings on this installation. Metering method, comparison baseline, test duration and operating conditions are unspecified. No capex, total operating cost or payback is reported.

When

Disclosed 20 October 2026 in Q3 results. Exact installation/measurement dates unknown. One unnamed well, not a fleet result.

What remains unknown

Anonymous single-installation supplier claim. Do not treat the percentage as company-wide energy savings, production uplift, net cash ROI or an AI result. The word ForeSite links a brand family, not proof this equipment used the PDO predictive model.

What to ask next

Ask for metered before/after energy under comparable duty, uptime and load; retrofit cost and maintenance burden; and the well conditions under which the result transfers.

Weatherford Announces Third Quarter 2026 Results — rod-lift energy recovery
2026-10-20 · Technology Highlights → Production and Intervention; first U.S. ForeSite Power Regenerative System retrofit paragraph.
Editor read original issuer HTML and displayed 20 October 2026 publication date on 24 October. One unidentified U.S. well; supplier result, not independent validation.

What holds up in SLB’s data and AI investment thesis?

The work

SLB has a credible route to earning more from AI: it can put specialized models into software and operating workflows that customers already use. The public evidence supports a platform, engineering and governed-access advantage. It does not support treating every customer’s downhole data as an unrestricted SLB-owned training pool.

How it works

Menon describes a seismic model trained on public and SLB multiclient datasets, with customers able to fine-tune for their own assets. Lumi connects governed data; domain models handle technical formats; Tela provides interaction and workflow orchestration. SLB describes subscription upgrades, metered usage, tailored workflows and marketplace revenue sharing. These are commercial mechanisms, not disclosed AI-only earnings.

What was reported

FY2025 Digital revenue was $2.660bn and adjusted EBITDA $928m (34.9% margin). Q3 2026 Digital revenue was $693m, adjusted EBITDA $228m and ARR $1.04bn. These are division-wide figures, not AI revenue or customer ROI. The CFO’s 2030 $1.8–2.0bn adjusted EBITDA and approximately $2bn ARR are ambitions, not results.

When

Original investor-day scripts: 17 June 2026. FY2025 baseline released 23 January 2026; Q3 results released 23 October 2026. Terms dates unknown. Investor interview framing published 20 October 2026.

What remains unknown

SLB disclaims ownership of Customer Data in its published online-service terms, while allowing specified service/product-improvement uses. Embedded Gen AI prompts and outputs require customer consent for training. Those clauses do not settle rights to every historical field dataset, derivative or negotiated contract. Digital includes software, operations, data licensing and professional services, with no AI-only profit bridge. Digital Operations overlaps Core segment reporting and is eliminated in consolidation. No demonstrated undervaluation or complete customer net cash ROI follows from these disclosures.

What to ask next

Ask for AI-specific paid usage and retention, incremental profit after compute and support, training rights by data category, held-out-basin model performance, customer-specific control boundaries, and a valuation using post-elimination earnings. For the full evidence map and interactive revenue chart, read the SLB briefing linked from the homepage.

SLB third-quarter 2026 results — Digital economics and category definitions
2026-10-23 · Digital results; Supplementary Information questions 10–13; Digital customer highlights.
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation.

SLB full-year 2025 results — Digital financial baseline
2026-01-23 · Supplementary Information question 12, Digital FY2025 table.
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation.

Shashi Menon: The Race to Scale Digital Operations and AI
2026-06-17 · PDF pp. 3–5: data access and control; p. 11: seismic training; pp. 14–15: monetization.
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation.

Stephane Biguet: Proven Growth. Accretive Returns.
2026-06-17 · PDF p. 4: customer base; pp. 7–10: commercial models; p. 11: 2030 ambitions.
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation.

SLB Online Services Terms and Conditions
Unknown · Definitions: Customer Data and Interaction Data; sections 8(a)–(b), 9(a), 10(a).
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation. Effective/publication date unknown. Public standard terms do not establish the terms of every negotiated customer agreement.

SLB Online Services Service Specific Terms
Unknown · Generative AI Services: definitions; Connected Gen AI A(i), A(iv); Embedded Gen AI B(iii).
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation. Effective/publication date unknown. Public standard terms do not establish the terms of every negotiated customer agreement.

Robert Mullin interview: 3.5% of the S&P, 20% of the Cash
2026-10-20 · Host introduction and linked interview; exact SLB interview timestamp not established.
Original host introduction reviewed; confirms the investment framing and interview identity. Full interview passage and exact timestamp not independently checked. An Eisman episode containing the same argument remains unconfirmed.

Trygve Randen: Our Digital Advantage
2026-06-17 · PDF pp. 2–3: governed platforms; pp. 5–6: hosting and trust; p. 7: operational footprint.
Editor read the accessible original publisher text on 24 October 2026. Statements remain publisher-reported; no independent performance validation.

Read the source review

A maintenance alert only helps if a crew can act on it.

The work

Halliburton describes LOGIX Unit Vitality detecting early problems in cementing equipment so maintenance teams can investigate and plan repairs.

How it works

Sensors feed more than 400 equipment parameters into cloud models. Staff check warnings, arrange parts and schedule corrective work. The product page separately describes transferring approved job designs from VIDA and iCem into controllers; this is not proof of autonomous maintenance.

What was reported

The supplier reports estimated downtime costs avoided of $125,000–$150,000 and 91% better maintenance productivity in four deployments. It does not define the productivity measure or say whether the dollar range is per deployment or combined.

When

Article published 22 October 2026; product page undated. Deployment dates, customers, assets, basins and equipment population are undisclosed.

What remains unknown

No baseline, evaluation period, false-alarm rate, cash reconciliation, full cost or independent validation. Estimated avoided downtime is not booked savings or net ROI. No named customer adoption or new counted result. Separate from LOGIX drilling results in AI-016.

What to ask next

Did the warning lead to the right intervention, and did avoided downtime exceed inspection, repair and system costs?

How predictive intelligence reduces downtime and improves equipment reliability
2026-10-22 · A simple question led to a new approach; How predictive reliability works; What early risk identification delivers.
Editor read the complete original Halliburton article and displayed publication date. Supplier-selected examples; operator names, operating dates and measurement methods are not provided.

LOGIX Unit Vitality — cement equipment health monitoring
Unknown · Automated cement equipment health assessment; AI-driven models; Real-time monitoring closes the loop.
Editor read the original product page. Exact publication date unknown. Same supplier as the article, not independent validation; product capabilities are not proof each feature was used in the four reported deployments.

Read the work and its limits

ONGC’s shared enterprise system predates the current AI wave.

The work

ONGC says Project ICE connects finance, materials, HR, production revenue and other functions on SAP S/4HANA.

How it works

The operator reports more than 500 locations, 24,000 users and 23 modules. It describes a three-site disaster recovery setup and a SAP RISE journey over the next two to three years; that roadmap has no section-specific date.

What was reported

These are reported platform-scale figures, not active AI users, measured uptime, productivity, cost savings or ROI.

When

Undated page checked 24 October 2026. The program began with SAP R/3 in 2004. S/4HANA transition date and the roadmap’s starting date are unknown.

What remains unknown

Historical non-AI context only. AI-enabled innovation is a roadmap aim; no specific live AI workflow or decision authority established. A shared platform does not by itself prove clean data, adoption quality or AI readiness.

What to ask next

Can teams retrieve the same trusted records across functions, and which transactions still require manual reconciliation?

ONGC Technology — Project ICE
Unknown · Project ICE, five paragraphs. Site-wide last-updated stamp does not establish when this section or its roadmap was published.
Editor read the original ONGC Project ICE section. Legacy SAP R/3 rollout is dated 2004; section publication date and S/4HANA transition date are unknown. Website-wide last-updated label is not a section publication date.

Read the work and its limits

AI checks supplier documents. Petrobras staff make the decision.

The work

Petrobras describes AVAL IA Wave 2 reviewing supplier-registration documents before a technical evaluator checks them.

How it works

The evaluation team records its decision and feeds it back to the AI. Petronect handles registration; the notice does not identify the AI model or developer.

What was reported

No accuracy, workload, completed-case count, cost, saving or ROI reported.

When

Undated notice checked 23 October 2026. Forms are scheduled to open from September 2026, with documents due 30 days after each questionnaire opens; actual opening and evaluation dates are not confirmed.

What remains unknown

Technical evaluators keep the final check and decision. The notice says the pilot may pause for adjustments; it does not say it failed or was suspended. No LLM, autonomous approval or full rollout established.

What to ask next

Does the first AI pass reduce total review work while preserving the quality of supplier decisions?

Cadastro de Fornecedores — AVAL IA Onda 2 pilot notice
Unknown · Nova Fase do Projeto Piloto AVAL IA: O que muda / Prazos dessa fase. Portuguese original; September 2026 form schedule is not the publication date.
Editor read the original Petrobras Supplier Channel notice in Portuguese. Exact publication date unknown; the notice describes Wave 2 and a form-opening schedule from September 2026. Current completed-case volume and pilot results are not supplied.

Read the pilot and its limits

Azule connects the well plan to drilling automation

The work

SLB describes Azule using one digital workflow from well design through rig execution offshore Angola.

How it works

DrillPlan passes the plan into DrillOps automation. The supplier case identifies AI-driven operations; it does not isolate AI from engineering automation.

What was reported

Supplier reports 95% of engineering tasks automated, planning time of 192 versus 31 hours and 75% fewer manual handoffs. A 2,000-meter section had zero reported nonproductive time and 16% higher net rate of penetration than the field benchmark. These are different measures, not a common financial return.

When

Q3 release published 23 October 2026. The related case study is undated. One well and a 2,000-meter section are described; exact operating dates unknown.

What remains unknown

No well identifier, wider well population, benchmark construction, review-effort adjustment, quality comparison or full cost. Execution is described as autonomous, but approval, override and fallback rules are undisclosed. No LLM or AI-only contribution is established; zero NPT for this section is not a long-term reliability rate.

What to ask next

Does passing an approved plan into rig execution reduce total work once checking, exceptions and overrides are counted?

SLB third-quarter 2026 results — selected operating and digital highlights
2026-10-23 · PDF p. 16: Digital / Azule Energy paragraph; same case and key measures.
Editor read the original issuer earnings-release HTML and relevant original PDF text. PDF p. 14 and p. 16 are physical page numbers (zero-based viewer P13/P15). Direct local PDF download failed; no call transcript reviewed. Supplier-reported customer claims, not independent validation.

Azule Energy pioneers digital well planning and rig automation in Africa
Unknown · Location; Results: Achieving operational excellence; Products Used: DrillOps AI-driven workflows.
Editor read the original supplier case study. No exact publication date displayed; the 2026 URL path and search crawl dates are not publication dates. Same supplier as the Q3 release; not independent corroboration.

Read the source and limits

A chemistry-control result is useful, even though the operator is unnamed

The work

SLB reports AI injection control and anomaly detection on nearly 1,400 producing Permian wells, with remote recalibration and weekly reviews.

How it works

Analyst takeaway: measure dosage accuracy alongside cost, and ask who reviews exceptions.

What was reported

Reported on-target injection: 63% to 92%; overtreatment: 17% to 3%; $4.4 million in net costs avoided in the first nine months. The supplier does not define the net-cost calculation or disclose a full-cost ROI.

When

Published 23 October 2026. First nine months of the program; exact start/end dates not stated.

What remains unknown

Unnamed operator: no company attribution, core adoption or result-count addition. No controls, cost reconciliation, model error rate or approval/override policy. Do not infer a specific Permian producer or treat the figure as audited net cash ROI.

What to ask next

What costs were included, and did lower overtreatment persist without compromising corrosion or scale protection?

SLB third-quarter 2026 results — selected operating and digital highlights
2026-10-23 · PDF p. 14: Technology Highlights / United States production chemistry paragraph.
Editor read the original issuer earnings-release HTML and relevant original PDF text. PDF p. 14 and p. 16 are physical page numbers (zero-based viewer P13/P15). Direct local PDF download failed; no call transcript reviewed. Supplier-reported customer claims, not independent validation.

Read the source and limits

Eni adds a concrete on-premise subsurface workflow

The work

SLB says Eni expanded petrophysics and geomechanics AI workflows using an on-premise Lumi platform.

How it works

Analyst takeaway: the location of the data platform matters when deciding what information a new assistant or model can access.

What was reported

No task-level benefit, asset population or full cost quantified in this update. Eni’s existing portfolio metrics and targets remain separately sourced; none is allocated to Lumi.

When

Published 23 October 2026; exact deployment and expansion dates undisclosed.

What remains unknown

Supplier statement, without independent operator corroboration of this specific expansion. Model type, user count, data permissions and decision authority are undisclosed. No inference that this uses an LLM or captures all company communications.

What to ask next

Which petrophysics or geomechanics decisions changed, and what data and review boundaries apply?

SLB third-quarter 2026 results — selected operating and digital highlights
2026-10-23 · PDF p. 16: Digital / SLB and Eni paragraph.
Editor read the original issuer earnings-release HTML and relevant original PDF text. PDF p. 14 and p. 16 are physical page numbers (zero-based viewer P13/P15). Direct local PDF download failed; no call transcript reviewed. Supplier-reported customer claims, not independent validation.

Read the source and limits

BSP awards an on-premise software contract

The work

BSP awarded SLB an on-premise contract for Petrel and Techlog, with digital and AI capabilities across subsurface and production work.

How it works

Analyst takeaway: a broad contract establishes buying intent and implementation scope; it does not show which AI tasks employees complete.

What was reported

No measured operating or financial result disclosed. Contract scope is not a user population or proof of completed deployment.

When

Published 23 October 2026. Award date, implementation timetable and active-use dates unknown.

What remains unknown

Supplier reports an award. Live AI workflows, algorithms, asset coverage, approvals and full cost remain unverified. BSP is separately identified; this award does not establish a Shell-wide rollout.

What to ask next

Which workflows will go live first, and what acceptance test separates installed software from useful work?

SLB third-quarter 2026 results — selected operating and digital highlights
2026-10-23 · PDF p. 16: Digital / Brunei BSP software contract paragraph.
Editor read the original issuer earnings-release HTML and relevant original PDF text. PDF p. 14 and p. 16 are physical page numbers (zero-based viewer P13/P15). Direct local PDF download failed; no call transcript reviewed. Supplier-reported customer claims, not independent validation.

Read the source and limits

SLB repeats ADNOC’s rig-center scope, without a new rollout count

The work

SLB’s Q3 release repeats that ADNOC’s AI-enabled operations center covers more than 120 onshore and offshore rigs.

How it works

Analyst takeaway: this is another source for an existing footprint, not evidence that a new set of rigs was added.

What was reported

No additional rollout count or newly measured benefit. Earlier ADNOC claims retain their own source, period and limitations.

When

Published 23 October 2026; earlier operator disclosure was 4 August 2026. No new commissioning date.

What remains unknown

Same supplier relationship and already covered footprint; not independent validation. Rig oversight and coordination do not establish autonomous drilling control on every rig. Do not transfer an individual well’s performance to the rig-center fleet.

What to ask next

What is monitored centrally, what action can the platform take and who approves a change on a rig?

SLB third-quarter 2026 results — selected operating and digital highlights
2026-10-23 · PDF p. 16: Digital / ADNOC Real-Time Operations Center paragraph.
Editor read the original issuer earnings-release HTML and relevant original PDF text. PDF p. 14 and p. 16 are physical page numbers (zero-based viewer P13/P15). Direct local PDF download failed; no call transcript reviewed. Supplier-reported customer claims, not independent validation.

Read the source and limits

Vista uses AI monitoring to adjust completion designs

The work

Vista says its own AI monitoring tool helps engineers make completion-design decisions in real time in Vaca Muerta.

How it works

The tool identifies frac hits and helps adjust stimulation to protect parent wells. The investor-day slide says it has been used since March 2025.

What was reported

Vista reports less than 1% of completion stages lost in 2025 and one runaway stage avoided per well on average. These are issuer-reported technical outcomes. The same slide prints a cost-saving unit as ~150 $M per well; that dollar amount is unresolved and excluded from comparisons. No net ROI is established.

When

Investor day: 12 November 2025. Implementation since March 2025. Results labeled 2025; exact evaluation dates, well count and later continuation not disclosed.

What remains unknown

No matched baseline, well or pad population, effect-isolation method, tool costs or independent validation. Approval and execution rights are not disclosed. The schematic stage numbers are not a project population. The cost-saving unit is ambiguous even in the original image; no dollar benefit is normalized or counted.

What to ask next

What changed in completion decisions, and do the avoided stages still pay after tool costs and engineering review?

Vista Investor Day 2025
2025-11-12 · Slide 19 / PDF p. 19; title, implementation date and results column. Date corroborated by SEC exhibit heading.
Editor downloaded and read the original issuer PDF, inspected slide 19 visually and cross-checked the SEC exhibit. Company-reported, not independently validated.

Vista Investor Day 2025 — SEC exhibit
2025-11-12 · Exhibit heading: November 12, 2025; same slide 19 presentation.
Original SEC exhibit read. A second access path to the same issuer presentation, not independent corroboration of outcomes.

Read the source and limits

Saipem installs a predictive-maintenance pilot on one drillship

The work

Saipem says it installed an AI predictive-maintenance system on the Saipem 12000, the first drillship in its fleet to receive it.

How it works

With ADC Energy, the pilot reads equipment data to flag anomalies and anticipate failures. A separate BIP project for diesel generators on Saipem 7000 was still being implemented and awaiting tests.

What was reported

No measured downtime reduction, failure forecast quality, intervention count, cost saving or ROI disclosed. One installed pilot is not a fleet deployment.

When

Announced 23 February 2026. Pilot installation described; operating evaluation period and later fleet rollout unknown. Saipem 7000 testing was prospective.

What remains unknown

The issuer explicitly calls the 12000 project a pilot. Fleet extension and the separate 7000 project remain plans. Human maintenance approval, automatic execution, fallback rules and measured outcomes are undisclosed.

What to ask next

Did the alerts lead to useful maintenance interventions, with fewer false alarms and enough avoided downtime to cover full costs?

Saipem introduces an AI-based predictive maintenance system onboard the Saipem 12000
2026-02-23 · Page date; Saipem 12000 implementation, ADC Energy, pilot and separate Saipem 7000/BIP paragraphs.
Editor read the full original issuer announcement. Installation is described; no measured reliability result or fleet rollout independently verified.

Read the source and limits

Worley puts four internal tools into engineering and bid work

The work

Worley lists four deployed tools for vendor comparison, engineering-document checks, internal knowledge search and customer-response drafting.

How it works

AI.VendorSelect supports technical-bid evaluation; AI.HoldsFinder extracts design holds; AI.Assist searches proprietary engineering knowledge; Smart Response Generator helps draft responses to queries and expressions of interest.

What was reported

Four tool names are a program scope, not four independently measured deployments. No active-user count, measured time saving, accuracy comparison, full cost or return is disclosed.

When

FY2025 results presentation dated 27 August 2025. Individual go-live dates, usage duration and later adoption unknown.

What remains unknown

These are Worley’s internal tools across energy, chemicals and resources. No upstream-only user cohort or client field deployment is established. NVIDIA, Microsoft and Dell appear as general partner logos, not confirmed partners for each tool. No autonomous vendor selection, design approval or customer communication is established.

What to ask next

Which work do employees complete with each tool, and how much review remains before a bid, design or response is approved?

FY2025 full-year results presentation
2025-08-27 · Slide 31 / PDF p. 31: recently deployed Advanced Development Lab solutions. Date on cover.
Editor downloaded and read the original issuer presentation and visually inspected slide 31. Named internal tools, not measured client or field outcomes.

Read the source and limits

Whitecap describes the data foundation it is building around AI

The work

A Shakudo customer story describes Whitecap’s governed data platform and business-led AI prototypes, with IT bringing selected work into production.

How it works

The account describes relational data and operating telemetry, on-premise language models and selective frontier-model access. Human review remains part of the approach.

What was reported

The page’s less-than-one-hour analytics headline conflicts with narrative describing an endpoint still being built toward. It is not counted as achieved savings or ROI. Company production and team size are not an AI user population.

When

Vendor index dated 23 April 2026; story itself undated. Platform use described, but project start and evaluation dates are unknown.

What remains unknown

Vendor account with a named customer executive, without independent operator corroboration. Predictive analytics and agentic work are prospective. The conflicting time claim, model quality, active use and full costs remain unresolved.

What to ask next

Which prototypes have become routine work, and what does the production team measure before approving wider use?

Whitecap Resources customer story
2026-04-23 · Key Results; govern-first platform, AI operating model, human review and endpoint being built toward. Page is undated; date comes from vendor customer index.
Editor read the original vendor customer story. Named customer executive is quoted; operator-side corroboration and outcome methods not verified.

Shakudo customer stories — Whitecap entry
2026-04-23 · Whitecap customer-story listing dated April 23, 2026.
Original vendor customer index read; verifies listed publication date, not a deployment or benefit.

Read the source and limits

Technip Energies reports progress, without proving 55 AI deployments

The work

Technip Energies says its transformation plan had launched more than 55 digital and AI initiatives and delivered more than 20 minimum viable products by the end of 2025.

How it works

The plan began in 2024. An AI-for-all campaign launched in 2025 aims to upskill staff; neither completed training nor routine use by all employees is established.

What was reported

The counts mix digital and AI work. Launched initiatives and delivered MVPs are not equivalent to production deployments, active users, measured savings or net ROI.

When

2025 annual report published in 2026; exact publication day unverified. Progress is explicitly at end-2025. The URL timestamp is not treated as a publication date.

What remains unknown

Only the indexed original p. 41 excerpt was reviewed; direct PDF access returned 403. Full-report definitions and later rollout remain unreviewed. No per-initiative deployment status, active-user denominator, measured benefit or full-cost return verified. The earlier DAP source dates that program to 2025; the annual report dates a broader transformation plan to 2024. Their exact relationship is unresolved.

What to ask next

Which MVPs survived evaluation, reached routine use and produced a measured result after implementation and review costs?

Technip Energies 2025 Annual Report
2026 (exact publication day unverified) · Printed p. 41, Value creation / We strive for excellence: digital transformation plan and end-2025 output counts.
Editor read the search-indexed original annual-report excerpt. Direct PDF returned 403; full report unreviewed. URL timestamp is not used as an exact publication date.

Read the source and limits

A neural network struggled to predict pressure in this test

The work

An SPE study compared a random forest and a neural network on 206 historical pressure samples. The neural network performed poorly on the reported test split.

How it works

The authors used 10 inputs and an 80/20 training/test split. They describe a small, noisy dataset; the abstract does not establish testing on entirely unseen wells.

What was reported

Neural-network mean absolute error: 209.55 psig; random forest: 93.73 psig. Test R² was 0.05 versus 0.81. R² is not a percent-correct accuracy score. No operating loss or deployment cost is reported.

When

Published 10 August 2026. The ProBHP data are attributed to sources from 1975 and 1986; collection dates, well count, geography and test chronology are unspecified.

What remains unknown

Only the publisher-indexed abstract was reviewed; direct access failed. No raw data, full validation protocol, deployed system, stop decision or later outcome verified. This does not establish a failure rate for industry AI projects.

What to ask next

Would performance hold on entirely new wells, with the full error distribution and an operating decision attached?

Evaluation of Influential Parameters for Prediction of Flowing Bottom Hole Pressure in Dry Gas Wells Using Machine Learning
2026-08-10 · SPE-234889-MS; publisher-indexed abstract: dataset/split and Random Forest/ANN comparison paragraphs. Published 10 August 2026; conference 3–5 August 2026.
Editor independently read the search-indexed original publisher abstract and metadata. Direct OnePetro/DOI opens failed; local page fetch returned 403. Full paper, raw data and complete methods not reviewed. No claim of field deployment or independent replication.

Read the model-test limits

Diamondback wants someone to challenge the economics of its AI projects

The work

An official AI Lead – Value and Economics posting calls for someone to test AI business cases and distinguish actual impact from correlation.

How it works

The role would set baselines and comparison methods, challenge benefit assumptions, compare forecasts with realized value, and recommend whether projects continue, scale, change course or stop. It works with Finance, product owners and the AI portfolio lead.

What was reported

No project results, filled-role evidence, implemented measurement framework, cost or return is disclosed.

When

The posting was available when checked on 22 October 2026. Its publication date and whether the role has been filled are unverified.

What remains unknown

A job description establishes intended responsibilities. It does not show that these controls exist today or that any project was stopped. Existing Diamondback operating initiatives retain their own evidence.

What to ask next

Who can challenge an AI business case, and what evidence would make the company stop funding it?

AI Lead – Value and Economics · R100885
Unknown · Job Responsibilities: valuation framework, baselines/comparison groups, benefits attribution, continue/scale/pivot/stop recommendations. Official public jobPostingInfo payload retrieved.
Editor retrieved and read the full official public Workday job payload. Human-facing page returns no extracted text; search copy corroborates responsibilities. Publication date not verified; metadata startDate is not a verified hire or operating date.

Read the hiring signal and its limits

Italgas’s voice agent can help with appointments; the reported gains need more checking

The work

A vendor-hosted talk describes Italgas testing a voice agent on a subset of customer calls, with people handling complex requests.

How it works

A real-time language model uses governed knowledge and authenticated CRM tools for appointment work. Databricks supports retrieval and tracing; Azure hosts the voice model. Broader channels are planned.

What was reported

Speakers claim 50% fewer dropped calls and 40% more automated resolutions in beta. Test size, period, metric definitions, quality, costs and net return are absent.

When

Undated talk. Full call coverage was planned for September, with no verified year or completion. It is not confirmed as a fresh announcement.

What remains unknown

The transcript’s earlier baseline figures do not reconcile. The summary omits the beta boundary. Do not reconstruct absolute outcomes, infer full rollout or treat these claims as measured upstream results.

What to ask next

Does the bot complete the right task, escalate safely, and reduce total effort after corrections?

Building Production Voice Agents with Databricks and AI Gateway
Unknown · Full transcript: 02:04 and09:38 baseline conflict;10:10–16:57 architecture and CRM;19:28–20:15 limited beta, planned September rollout and claimed changes;21:05 human operators.
Editor read vendor-hosted named-speaker transcript, not just its summary/FAQs. Original video link fetch failed. Transcript baseline numbers conflict; exact talk/publication date and full rollout remain unverified.

Read the adjacent gas-distribution case

TC Energy puts people between equipment alerts and field action

The work

TC Energy describes an equipment-surveillance process built around specialists and field teams. FAST is its consolidation plan with Accenture and AVEVA, separate from Pipeline IQ and compressor-dispatch optimization.

How it works

Alerts pass through triage, specialist investigation, corrective action and documentation. The proposed platform joins existing tools, AVEVA and AWS. SageMaker model development is marked development-only; language assistants and deeper SAP coordination sit under future work.

What was reported

The presentation credits existing surveillance with over $100m/year in avoided failure costs. FAST’s additional $250m over ten years is a forecast. Neither is measured LLM value or all-in net ROI.

When

Undated AVEVA World 2026 presentation, internally dated 29 May. Page 16 schedules Canada for Q3 2026 and US/Mexico for Q4; page 23 describes a first release in under seven months. The 16 October opinion does not resolve completion.

What remains unknown

The deck’s release wording is inconsistent; scheduled quarters do not prove go-live. Benefits lack a disclosed counterfactual, cost schedule, currency designation and measurement period. Do not add the forecast to the annual claim or assign either to FAST’s future AI features.

What to ask next

What reached production, who acts on an alert, and what remains after false alarms, specialist time and deployment costs?

Practical AI in Action: TC Energy’s Path to a Unified, Scalable Asset Reliability Platform
Unknown · PDF pp.14–16: advisory workflow, upcoming FAST and release roadmap; p.17: architecture; p.20: forecast; p.22: future AI work; p.23: release wording
Original 27-page presentation downloaded and read by editor; pp.14,16,22 visually inspected. Deck internally dated 29 May 2026; exact publication date unknown.

How the oil and gas industry is putting data first
2026-10-16 · Posted date; Improved reliability through AI monitoring section
Original vendor opinion read by editor. Named-customer program claims are not independent outcome validation or release confirmation.

Read the operating detail

Shell’s pump analysis stalled until an operator joined the team

The work

Shell describes an initially unsuccessful search for pump-disruption precursors at Perdido. A field operator helped the data team refine its code and find a chemical signature.

How it works

The team first drew a blank, later plateaued and considered stopping. The operator supplied equipment knowledge. Shell then reports a signature before 70% of disruptions, with one example about two days beforehand.

What was reported

The 70% is a reported precursor association in the analyzed events, not demonstrated live prediction accuracy. Event count, false alarms, held-out validation, intervention results, uptime benefit and full costs are undisclosed.

When

Undated webpage and video; operating and evaluation dates unknown. Already mentioned in the coverage log, now structured as a readable case. It is not a new 2026 deployment.

What remains unknown

The transcript does not establish a live alert or control workflow, prevented disruptions or realized savings. The two-day example is not an average warning time. Initial stalled analysis is not proof of an abandoned or failed field deployment.

What to ask next

Could the team reproduce the result on later events, with all false alerts recorded, and show what operators changed because of it?

The Data Detectives — Solving the 10-year-old problem at Perdido
Unknown · Artificial intelligence examples / Perdido video transcript; initial blank, plateau, operator input and refined precursor signature
Original operator webpage and relevant video transcript read by editor. Page/video publication and operating dates are not established; historical or undated context only.

Read the operating detail

Quantum adds a relational layer to its deal-evaluation data

The work

Databricks describes Quantum using Lakebase, Unity Catalog and Lakeflow to give deal teams a shared, governed view of geological, production and financial data.

How it works

A Postgres interface and PostGIS geospatial support connect familiar tools, including Tableau and Spotfire. Schemas, stored procedures, triggers and access controls help teams use consistent datasets. This is data infrastructure; the story does not establish an AI decision-maker.

What was reported

Vendor/customer-reported scale: 1.5 billion records across six vendor feeds and proprietary data, over 100 redundant tables removed, and over 11,000 monthly queries across a 150-person firm. Queries are not active users, completed deals or investment returns.

When

Undated source; adoption and evaluation periods unknown. Extends existing sponsor platform AI-052, separately from QB Energy’s field-planning workflow AI-054.

What remains unknown

No named asset, portfolio-operator rollout, task-time or accuracy baseline, investment-approval rule, full cost or measured return. The story’s investment-performance language is not evidence of investment alpha or AI attribution.

What to ask next

Can the team trace a deal assumption to its source and show whether shared data reduces reconciliation and review work?

Transforming energy deal evaluation with Lakebase
Unknown · Building a consistent foundation; Lakebase relational capabilities; adoption and FAQ paragraphs
Full original vendor customer story read by editor. No publication date shown; search age and footer year are not rollout dates. Vendor/customer claims, not independent validation.

Read the operating detail

A gas-power project queue is not committed gas demand

The work

The IEA’s 14 October report explains why data-centre power announcements need to be checked against project maturity, equipment delivery and actual grid access.

How it works

It distinguishes islanded, mixed and temporary bridge-power arrangements. Connection queues can include duplicate or speculative requests, while operating loads often ramp up in stages. A later grid connection can change how onsite generation is used.

What was reported

The report describes roughly 210 GW of planned onsite gas generation globally, around 190 GW at proposal stage and 17 GW under construction. These are rounded project-capacity figures, not additional additive totals, operating load or contracted gas volumes. No residual operating capacity is inferred.

When

Published 14 October 2026. Gas-project figure uses IEA analysis of Global Energy Monitor data from August 2026. A distinct October edition, not a new announcement today or an amount to add to the April IEA report.

What remains unknown

Project announcements do not establish completion, utilization, fuel burn or firm offtake. The report’s abstract covers 34 policy jurisdictions and about 90% of global data-centre capacity; those are policy-study scope measures, not an oil-and-gas AI-adoption sample.

What to ask next

Before relying on a project’s gas demand, ask for its customer commitment, generation equipment, permits, grid-access terms and phased load schedule.

Powering Artificial Intelligence: Strategies for integrating data centres into electricity systems
2026-10-14 · Dated landing page; executive summary; PDF pp.6–10, 36–47 and 58–62; gas-generation pipeline on p.60
Original IEA landing page, executive summary and relevant report passages read. Date verified on landing page. Policy and project-pipeline analysis, not operator AI adoption or contracted gas demand.

Devon confirms ChatDVN is part of daily work

The work

Devon’s CEO says its firewalled ChatDVN tool was a standard part of daily work by May 2026, after years spent organizing internal data.

How it works

The call distinguishes access to company data, AI-assisted coding and interpretation of operating data, and a next phase of process redesign. It does not establish that every example runs inside ChatDVN. Separate autonomous lift controls remain in AI-006.

What was reported

No active-user count, task-quality comparison, ChatDVN-specific cash benefit or full cost is disclosed. The call does not corroborate the older interview’s 15–30% productivity claim or AI-101 training count. Devon’s $1bn business-optimization target is not ChatDVN savings.

When

Call dated 6 May 2026; CEO describes three years of operation, an approximate history rather than an exact start day. This is newly linked original evidence, not a new launch.

What remains unknown

Daily use supports an operating workflow, not verified enterprise-wide active adoption. No comprehensive communications capture, decision-outcome memory, execution permissions or AI executive authority is established. Training and productivity claims from the earlier trade interview retain their separate, unverified status.

What to ask next

Which jobs do employees finish through ChatDVN, and does the full review process improve time, accuracy and follow-through?

Devon Q1 2026 earnings call — internal ChatDVN workflow
2026-05-06 · PDF and printed p. 4; Clay Gaspar prepared remarks, three waves of AI
Issuer-hosted corrected 20-page transcript retrieved HTTP200. Editor read relevant p.4 and visually inspected it; management claims, not independent performance validation.

Read the operating detail

Automated cementing has moved beyond its first job

The work

SLB says its cementing technology had passed 100 remote and automated jobs by July 2026, following first-quarter commercialization.

How it works

The release also names an automated dual-slurry operation on Shell’s Velox-1 in Egypt. The 100-plus figure combines remote and automated jobs; it is not 100 fully autonomous wells or a Shell-only fleet count.

What was reported

Job count shows reported operating reach. The release gives no controlled cycle-time, cement-quality, safety or cash-return comparison for this cohort.

When

Issuer release 24 July 2026; first-quarter 2026 commercialization described. Individual job dates and the remote/automated split are undisclosed.

What remains unknown

Automation and remote operation do not establish machine learning or LLM use. Human supervision, control envelopes, fallback and override rules are not specified. The same release’s AI Production Studio is a pre-field demonstration environment, not proof that those capabilities ran on Shell’s well.

What to ask next

Which steps execute automatically, what still needs the crew, and how does job quality compare under similar well conditions?

SLB Announces Second-Quarter 2026 Results
2026-07-24 · Core / Technology Highlights; automated cementing paragraph and Shell Velox-1 follow-up
Original issuer release opened and relevant technology-highlight paragraphs read by editor. Remote and automated categories are not broken out; no AI method specified.

Read the operating detail

Eight fields use physics-based asset modeling

The work

PETRONAS reports Integrated Asset Management across eight fields, combining physics-based models with near-real-time data for surveillance and production optimization.

How it works

The report says the workflow helps teams identify production opportunities faster. Its own passage does not attribute AI to IAM. Nearby PROTEAN and ERMAI AI descriptions concern different systems.

What was reported

The issuer says IAM improved production efficiency and value realization in 2025, without giving an amount, baseline, comparison group, attribution method or full costs. Eight is a field count, not eight models or platforms.

When

Operating period 2025. Source register dates the 2025 integrated report 27 February 2026; exact posting day was not re-audited. Printed p.55 / PDF p.29 newly inspected.

What remains unknown

No field names, user population, human approval or automatic control permissions are disclosed. The chapter’s broad AI heading cannot establish IAM’s method. Do not transfer PROTEAN’s equipment footprint or ERMAI’s time comparison to IAM.

What to ask next

What production decisions changed, and what incremental result can be separated from other field improvements?

PETRONAS Integrated Report 2025 — IAM
2026-02-27 · PDF p.29 / printed p.55; Turning Data into Production Upside
Original 10,849,024-byte PDF retrieved HTTP200; editor extracted and visually inspected spread54–55 (PDF29). Publication day retained from existing source register; exact posting day not re-audited in this pass.

Read the operating detail

PTTEP describes an in-house engine for specialist agents

The work

PTTEP says it launched X.brain, an internal AI engine grounded in its own operating knowledge and data.

How it works

The November 2025 feature describes support for analysis and recommendations. Eight specialist agents across subsurface, drilling, production, supply chain and corporate work were planned; the article does not establish that all eight went live.

What was reported

No X.brain user count, adoption rate or measured return is disclosed. The same article separately reports LogSight interpretation falling from two days to one hour and DORA drilling savings described only as “millions of US dollars annually”. Those are product-specific claims, not demonstrated X.brain results.

When

Published 3 November 2025. Launch date, operating population and subsequent rollout status unknown. Missed historical evidence, not a new October 2026 announcement.

What remains unknown

No proof of comprehensive communications capture, decision-outcome memory or autonomous executive authority. DORA combines machine learning and physics-based simulation; it is not shown to run on X.brain. No asset scope, test population or full costs support the DORA/LogSight comparisons.

What to ask next

Ask which agents people actually use, what data they can reach and which decisions still need a named human approver.

PTTEP’s AI Vision for a Smarter, Sustainable Future of Energy
2025-11-03 · Page date; X.brain introduction; eight-agent plan; DORA and LogSight examples
Full original operator page read by editor; Thai date 3 November 2568 converts to 2025-11-03. Company claims, not independent validation.

Read the operating detail

Five diagnosed equipment faults, with people confirming the fixes

The work

An undated MNVR and Energy Solutions case describes ODIE supporting seven gas generators at the Muskrat microgrid in North Dakota’s Bakken Basin.

How it works

Across six weeks, three cases came from agent alerts and two from operator questions. The authors say all five diagnoses were confirmed on inspection. Confirmed diagnoses and operator corrections are retained for future use. The product page keeps decisions with operators; automatic equipment control is not established.

What was reported

The authors jointly estimate $37,015 in avoided costs across the five selected faults. These are counterfactual estimates, not measured cash savings or net ROI. A $25,500 relay case includes the hypothesized cost of replacing the wrong controller; do not add that controller’s price again.

When

Publication date unknown; checked 21 October 2026. Six-week calendar window unknown. PDF filename is not a verified date. Adjacent gas-to-power operations, not an upstream adoption count.

What remains unknown

Selected equipment faults are not failed AI deployments. No total alert population, missed-fault count, false-positive rate, control group or full system cost. Some repairs were scheduled or could be planned; not all five were completed. Roughly 5,200 signals means data signals, not assets.

What to ask next

Can the team show every alert, inspection and resulting work order—including alerts that turned out to be wrong?

Five Failures Across Five Systems
Unknown · PDF pp. 1–3: seven generators; five fault cases; estimated costs; six-week period and retained diagnoses
Original three-page vendor/partner PDF retrieved HTTP200, full text read and pages 1–2 visually inspected. No publication date; filename is not date evidence.

ODIE — Operations Diagnostic Intelligence Engine
Unknown · Human in the loop / operators make decisions section
Original product page retrieved HTTP200 and read by editor. Undated product description; not a field performance validation.

Read the operating detail

Quantum’s planning gain includes a separate algorithm improvement

The work

Microsoft says algorithm changes in an updated version of Quantum’s field-planning tool reduced solve time by 50–65%.

How it works

The integer-linear-programming optimizer selects feasible pads and wells. A Copilot Studio agent explains outputs. The numerical improvement belongs to the optimization algorithm; the source does not attribute it to the language model.

What was reported

50–65% shorter solve times in an updated version, said to compound over hundreds of runs. The source does not specify absolute times, comparable hardware or the evaluation period. Do not multiply this with the separate three-weeks-to-20-minutes comparison.

When

Microsoft story dated 30 April 2026; later operator identification remains QB Energy from Quantum’s 2026 report. This is an enrichment of AI-054, not a new deployment.

What remains unknown

Vendor/customer reporting with no full-cost return or independent performance validation. The version comparison is a different baseline from the manual-workflow comparison. Engineers retain plan review; no capital approval is delegated. The reported solver improvement is not attributed to the language model.

What to ask next

Separate the solver, data preparation and assistant when testing which change actually improves the planning decision.

Field development planning at Quantum Capital Group
2026-04-30 · Field development planning solution; Improved speed, scale, and precision; updated algorithm paragraph
Full original Microsoft customer story read again on 21 October. Updated-version comparison is vendor-reported; no independent validation.

Read the operating detail

Twelve wells moved from recommendations to automatic adjustments

The work

U.S. Energy and ROAM-AI describe automatic ESP adjustments on 12 Permian wells. Engineers define the operating limits and handle exceptions.

How it works

Physics-aware models use live data to adjust setpoints within approved frequency, pressure, temperature and pump-curve limits. The authors say thousands of adjustments were executed automatically. No LLM is identified; exact override and shutdown procedures are undisclosed.

What was reported

The authors report 12.2% higher production against a 90-day pre-activation baseline, and a return exceeding four times initial cost within six months. Full costs and return definitions are absent; this is not verified net ROI. Six months was too early for a definitive pump-failure interval conclusion.

When

Started December 2025. July 2026 article; exact publication day unknown. Six-month results, not a year or the operator’s full fleet. This older case was missed in earlier research.

What remains unknown

Operator/vendor coauthored. No controlled comparator or disclosed adjustment for decline, workovers or other changes. Do not import the article’s separate unnamed 250-plus-well operator results into this 12-well case.

What to ask next

Ask for the all-in cost and benefit bridge, the baseline calculation and the log of exceptions, overrides and rejected adjustments.

The Bottleneck Isn’t the Model. It’s the Execution.
2026-07 · July 2026 cover story; A New Optimization Philosophy; Eric’s Inside View; author biographies
Full original publisher article retrieved (HTTP200) and read by editor. Operator/vendor coauthored, not independently validated. Month precision only; vendor’s July20 promotion is not the article’s exact publication date.

Read the operating detail

Rubiales adds a dated operating footprint to CADE

The work

Ecopetrol says its ISA-built energy decision platform managed 173 MW at Rubiales from March 2026, across five generation centres and the national grid.

How it works

CADE supports choices of electricity source, timing and volume. Sixty workers were trained; that is not a user-adoption count. The earlier report describes a Rubiales dispatch-model MVP being built in 2025; the newer platform coverage does not independently confirm that model’s commissioning or AI control authority.

What was reported

The company describes the covered supply as 20% of group demand. Another 931 MW and 80% coverage are plans, not achieved scope. MW is operating power coverage, not energy saved, emissions avoided or ROI.

When

Issuer announcement May 4, 2026; Rubiales coverage reported from March 2026. Older evidence newly checked, not a new October launch.

What remains unknown

No measured saving, implementation cost, LLM/model detail or human approval/override rule. The name “autonomous” and platform capacity do not establish autonomous AI dispatch. Group-wide generating capacity is a different population.

What to ask next

Which decisions are actually automated, and what changes in delivered electricity cost after all platform and operating costs?

Grupo Ecopetrol inaugura centro para optimizar su gestión energética
2026-05-04 · May 4 heading and Bogotá dateline; Rubiales operating coverage, ISA partnership, projected expansion and training paragraphs
Full original operator page retrieved (HTTP200) and read by editor. Coverage and plans remain company-reported; no independent cost or performance validation.

Read the operating detail

A driverless truck stayed available for 64 days

The work

Kodiak reports a 64-day uptime streak for one driverless truck in Atlas’s Permian Basin fleet, plus five trucks with streaks above 45 days.

How it works

The truck and autonomy software stayed powered on and available for service. This is physical driving automation, not a language-model assistant. No one was in the cab; remote support and intervention boundaries are not quantified.

What was reported

As of June 30: 35 Atlas-owned and operated driverless trucks and over 40,000 cumulative paid driverless hours. These are supplier-reported fleet figures, not the uptime of every truck. No fleet availability rate, savings or net ROI.

When

Announced October 20, 2026. The 64-day streak ran June 30–September 1. The fleet and paid-hours figures are as of June 30; 100 trucks by mid-2027 and a year-end highway launch remain plans.

What remains unknown

Available does not mean continuously driving or hauling. No mileage, dispatch utilization, collision/exposure comparison or cost denominator disclosed. Kodiak AI is separate from Kodiak Gas Services (AI-193).

What to ask next

How much paid hauling did those available hours produce, and what did remote support, maintenance and autonomy cost per load?

Kodiak Driver Sets New Uptime Record By Running 64 Consecutive Days on an Unsupervised Driverless Truck
2026-10-20 · Opening uptime definition and dates; Atlas fleet as-of paragraph; future fleet and highway plans
Full original supplier release retrieved (HTTP 200) and read. Supplier-reported availability and fleet figures; no independent reliability, safety or financial audit.

Read the operating detail

Engineers may want the knowledge more than another AI product

The work

SPE and i2k announced extended automated classification of OnePetro content. Separately, SPE says customers often wanted its technical content inside tools they already used.

How it works

i2k describes metadata tagging using more than 20 domain knowledge bases. SPE’s original JPT account says EnRG-LLM remains licensable and identifies Halliburton Landmark as its first content-license customer in an undated follow-up paragraph.

What was reported

No measured retrieval accuracy, time saving, token saving, license revenue, usage population or net ROI. General claims about i2k’s geotechnical agents do not establish that SPE uses every advertised workflow.

When

Classification release: October 20, 2026. JPT article: September 10 publication header, with an undated following-conference licensing update. Neither dates completed classification rollout.

What remains unknown

The complete issuer release was verified on syndication pages; the Business Wire original timed out. This is not an abandoned EnRG-LLM product. The two announcements do not prove that classification replaced the model.

What to ask next

Our reading: make trusted content useful inside existing tools before assuming customers want another assistant. Ask about content rights, retrieval quality and cost per completed task.

SPE extends i2k Connect information-classification collaboration
2026-10-20 · Opening announcement; classification paragraph; SPE CEO statement
Complete issuer-authored Business Wire release read on an accessible syndication page. Distributor original timed out; identical release also retrieved from AOL. No independent outcome validation. Distributor URL: https://www.businesswire.com/news/home/20261020770911/en/

Industry Executives Offer Practical Advice on How To Use Modern AI Models Wisely
2026-09-10 · SPE’s AI Journey; licensing-demand discussion; following-conference Landmark paragraph; cost and metadata discussion
Full original publisher article retrieved and read 20 October. September 10 publication header; later following-conference paragraph is undated. Direct current text takes precedence over stale search extracts.

Read the operating detail

Bumerangue’s appraisal contract includes drilling automation

The work

Halliburton’s August contract for bp’s Bumerangue appraisal campaign includes LOGIX automation and remote operations offshore Brazil.

How it works

The integrated drilling and evaluation scope connects planning, execution and evaluation. The release says LOGIX will be deployed; it does not identify an LLM, individual control permissions or human override arrangements.

What was reported

No number of wells, completed automated wells, operating comparator, quantified benefit, contract value or net ROI disclosed.

When

Original announcement August 25, 2026; checked October 20. This is older evidence newly added to this review, not a launch announced today.

What remains unknown

A contract is not proof of completed field use. Do not transfer LOGIX results from Eni’s Indonesia well or another customer to Bumerangue. Recorded as a named-operator extension of existing LOGIX coverage, not a second platform.

What to ask next

Which drilling tasks will run automatically, who can intervene, and how will bp compare outcomes with a credible baseline?

bp awards Halliburton integrated contract for Bumerangue field appraisal in Brazil
2026-08-25 · Opening and contract-scope paragraphs; LOGIX sentence
Full original supplier release retrieved (HTTP 200) and read. Future-tense scope; no completion or measured result established.

Read the operating detail

Oil India’s plan connects field operations with corporate systems

The work

Oil India describes ten DRIVE 2.0 workstreams, including a control centre, a digital oilfield covering 380 producing wells, subsurface agents, maintenance and corporate workflows.

How it works

The proposed systems would connect operating data with engineering and business records. Rig-planning approvals and safety notifications to designated people are described; AI decision authority and safeguards are not.

What was reported

The 380 wells are planned scope, not a verified connected population. No measured uplift, savings, implementation cost or net ROI.

When

Publication and DRIVE 2.0 start dates unknown. Header: “11/09/2026 09:35 AM”; footer: “20/10/2026 11:28 AM”. These update strings do not date deployment. The 2019 start belongs to earlier DRIVE.

What remains unknown

One undated issuer plan, now accessible after an earlier blocked route. No completion dates, live user population or model providers established. Ten workstreams are not ten operating AI deployments.

What to ask next

Which parts have been commissioned, and who can approve or override their recommendations?

Digitalization in OIL — DRIVE 2.0 programme
Undated · Digitalization in OIL; DRIVE 2.0 §§1–10; §2 380-well scope; §4 approvals; §7 notifications; header and footer update strings
Original issuer page read by editor 20 October 2026. Public /digitalfootprint path accessible; earlier /index.php/digitalfootprint attempt was blocked. Header and footer update strings are not verified publication or launch dates.

Read the ten-part plan by department

An AI finding is not permission to change an invoice

The work

SPE’s flow-measurement editorial describes AI support for investigating discrepancies in the quantities used for commercial settlement.

How it works

Rules can flag frozen signals; ML can prioritize anomalies; language tools can organize procedures and evidence. Settlement changes still need validated calculations, preserved records and approval.

What was reported

No deployment count, recovered revenue or ROI. Measurement uncertainty multiplied by volume and price is not a demonstrated loss.

When

Published 12 October 2026. Illustrative workflows; no named asset, operating period or installed system.

What remains unknown

Author affiliations include operators and suppliers, but this is not an ExxonMobil or service-company deployment announcement. Shared inputs can make checks agree while retaining the same bias. No evidence that AI changed billed quantities.

What to ask next

Can an investigator reproduce the correction and see who approved it?

From Measurement to Commercial Confidence: An Operator’s View of Custody Transfer—and a Responsible Role for AI
2026-10-12 · Meter Accuracy Is Not System Credibility; Custody Transfer is an Organizational System; Automation and AI Can Assist, but Accountability Cannot Be Automated; author biographies
Original publisher HTML and relevant passages read by editor 19 October 2026; guidance/review, not independently validated operating outcomes.

A model that fits one field may not travel well

The work

A stimulation-AI review highlights weak external testing, small datasets and inconsistent reporting in the published research it examines.

How it works

The authors discuss validation on independent fields, time-series leakage, physical constraints and integration with field systems. Their readiness assessments are judgments about applications, not audited operator rankings.

What was reported

The quantitative subset contains 23 studies, not 23 deployments. No independently established project return; broad cost and benefit estimates are not admitted as benchmarks.

When

Published 15 October 2026. Literature focus: 2020–2025, with earlier foundational studies; much of the underlying work predates the rolling window.

What remains unknown

Review conclusions depend on heterogeneous studies and incomplete reporting. We have not re-audited every underlying experiment. No named abandonment or industry failure rate; performance averages cannot rank companies or predict another field’s results.

What to ask next

Was the final model tested on a different field, with tuning kept separate from the test?

A systematic and critical review of artificial intelligence applications for enhancing well stimulation operations and sustainable reservoir management
2026-10-15 · Abstract; §§2.2–2.3 review selection/quality; §5.8 heterogeneity; §§6.1.1–6.1.5 validation, integration, readiness and economics; §7.4 denominators
Original publisher HTML and relevant passages read by editor 19 October 2026; guidance/review, not independently validated operating outcomes.

A good pilot did not make maintenance AI easy to deploy

The work

Murphy’s two-platform, 24-month project deployed 46 rotating-equipment models. Missing data delayed the first model until six months after kickoff.

How it works

Incomplete maintenance records and generic anomaly alerts created investigation work. Engineers and operations staff reviewed alerts; retraining reportedly reduced false positives. Verified alerts fed maintenance notifications, later through an API.

What was reported

No quantified downtime savings, project cost or net ROI. A known-failure pilot’s 100–121-day warning was a replay, not demonstrated live lead time.

When

Article: 20 January 2025. Alert records begin January/March 2023, before this report’s rolling window; exact project start/end dates unknown.

What remains unknown

The reviewed 2023 alerts are older context; current operating status is unknown. The authors cite insufficient supplier equipment expertise and reported mitigation—not abandonment. Platform names, supplier and precision/recall are undisclosed; no autonomous maintenance established.

What to ask next

How many alerts lead to useful work, and what does reviewing the rest cost?

Framework for AI- and ML-Based Predictive Maintenance for Offshore Rotating Equipment
2025-01-20 · Methodology; Observations; Discussion: Project Planning and Management, Data Prerequisites, Domain Knowledge, Model Development, Performance, and User Experience; Figs. 4–6
Original technical account read by editor 19 October 2026. Text and qualitative figure discussion reviewed; numeric chart values not transcribed. Not an independent outcome audit.

Remote crane control still has a person at the controls

The work

NOV reports controlling a full-size crane from about 20 km away and delivering one Aura-equipped remote-operated crane to an unnamed customer.

How it works

Aura brings camera feeds, load information, equipment diagnostics and environmental data into the human operator’s screen. Customers are expected to start onboard before moving control ashore.

What was reported

No measured cost, safety or crew reduction. Two further cranes were being built for delivery by early 2027; these are planned deliveries.

When

Article: 20 August 2026. Supplier test: Kristiansand to Søgne, Norway; exact test and first-delivery dates unknown.

What remains unknown

No AI/ML identified. Delivery does not establish shore-controlled customer operations or autonomous lifting. Customer, vessel, operational hours, interlocks and fallback arrangements are undisclosed.

What to ask next

What must be proven onboard before a customer can safely move the operator ashore?

Aura moves remote crane operations from concept to reality
2026-08-20 · Remote-access paragraph; Kristiansand/Søgne test; first delivery; two planned deliveries; phased customer adoption
Original supplier account read by editor 19 October 2026. Supplier demonstration and delivery are distinguished from customer operating use.

One survey vessel brings several offshore jobs together

The work

Oceaneering says an Ocean Intervention II survey campaign began in April 2026 for an unnamed international operator offshore Trinidad and the U.S. Gulf.

How it works

The scope combines geophysical and geotechnical work with surface-vessel-supported autonomous underwater surveys. U.S. Gulf pipeline inspections include non-contact cathodic-protection measurement.

What was reported

The supplier says combining work on one vessel and crew can reduce mobilizations, exposure and emissions. No measured savings, baseline or net ROI disclosed.

When

Announcement: 30 June 2026. About six months of work expected, including roughly four in Trinidad and two in the U.S. Gulf; not verified completed durations.

What remains unknown

Autonomous survey is not proof of ML or LLM use. The vessel still has a crew. Operator identity, human supervision, intervention rules, achieved campaign length and comparative performance are undisclosed.

What to ask next

Does combining survey tasks cut total vessel days once weather, rework and supervision are included?

Oceaneering Secures Multi-Discipline Cross-Regional Offshore Survey Contracts for Ocean Intervention II
2026-06-30 · June 30 release: opening contract/program paragraphs; Trinidad/U.S. Gulf scope; simultaneous operations; supplier quotation; 2025 vessel upgrades
Original supplier release read by editor 19 October 2026. Begun campaign is distinguished from planned scopes, durations and unverified completion.

Safety learning needs common definitions and traceable answers

The work

Havtil’s account of a 40-person petroleum workshop describes shared incident-learning work and concerns about unreliable or untraceable AI answers.

How it works

The slides show PDF extraction, a shared map of incident concepts and a searchable database. The approach supports review; it does not replace root-cause investigation for high-severity incidents.

What was reported

No accuracy score, incident reduction, cost or ROI. Participant concerns are general experience, not a measured evaluation of the demonstrated tool.

When

Workshop: 28 October 2025. Article: 7 January 2026. The tool was developed in 2024; exact development and slide-posting dates unknown.

What remains unknown

Forty representatives are not forty companies or a survey sample. No named failure or agreed follow-up project is identified. Listed framework user-test participants do not establish AI deployments. This is distinct from the 2024 responsible-AI overview (INS004).

What to ask next

Can reviewers trace each extracted cause to the original passage and correct a classification?

Can artificial intelligence help improve learning from incidents?
2026-01-07 · Workshop date/attendance; How AI can help; Quality and trust; Obstacles; The way forward
Original regulator account read by editor 19 October 2026. Workshop experience is not a representative survey or a tool benchmark.

Bruk av kunstig intelligens til læring etter hendelser
2025-10-28 workshop; posting date unknown · 36-page deck: physical/printed p. 15 workflow; p. 20 contributing-factor framework and high-severity root-cause-analysis boundary
Original PDF text read; pages 15 and 20 rendered and inspected. Web screenshot failed; local PyMuPDF rendering succeeded. Event date is not a verified posting date.

TotalEnergies uses generative AI to check trading data

The work

AWS describes a TotalEnergies Trading & Shipping system that matches structured and unstructured transaction data.

How it works

The workflow supports transaction traceability and input-error detection. AWS lists Bedrock and ECS; the model, connectors and exception-handling process are unspecified.

What was reported

A named customer manager reports more reliable controls and fewer financial losses, without amounts, error rates, a baseline or an observation period.

When

AWS page undated. Implementation date, trading-desk scope, user count and transaction volume unknown; an in-window launch is not established.

What remains unknown

Vendor-published customer testimony, not independent validation. No costs or net ROI. Matching records does not establish authority to place trades, release payments or change books. No specific model provider can be inferred from the Bedrock service description.

What to ask next

Which mismatches need a person’s approval, and how much error-review work does the system remove?

TotalEnergies puts Gen AI to work on its trading operations
Undated · Overview; Olivier Collon customer quotation; AWS Services Used
Original AWS customer account read by editor 19 October 2026. No visible publication date; a secondary 2024-10-28 clue could not be confirmed from an original dated source.

OMV tests field-development decisions with analytical agents

The work

OMV’s Norwegian CCS pilot uses decision theory and geostatistics with Stanford and TerraAI.

How it works

Seismic, geological and well data feed scenario planning for drilling, injection and monitoring. Engineers challenge assumptions and validate recommendations. OMV explicitly distinguishes this analytical system from generative AI.

What was reported

OMV reports over 25% higher project value, around 50% less quantified subsurface risk and simulations up to 1,000 times faster. These are early internal planning claims, not realized production or net ROI.

When

Norwegian CCS planning pilot; article published 14 April 2026. Test dates, asset name and project count unknown.

What remains unknown

No metric definitions, comparator details, validation sample or complete costs. The wider six-year Stanford sponsorship is not this pilot’s cost. Pilot remains separate from NAIC document search (AI-132).

What to ask next

Do better plans still outperform after engineering review and actual field execution?

How AI is powering better decisions in the race to net zero
2026-04-14 · Sections: 1,000 times faster; Creating value while reducing risk; A discussion partner, not a decision maker
Original publisher text read by editor 18 October 2026; claims remain source-reported.

Woodside’s maintenance estimate has a traceable source

The work

Woodside’s AI page supports the previously unresolved 15% maintenance-hours estimate at Angel.

How it works

Maint Intel compares maintenance history with reliability targets to recommend intervals. A separate January article identifies AWS support and the North West Shelf launch.

What was reported

The AI page describes potential 15% lower annual maintenance hours over five years, not hours already avoided. The January article reports model processing cut from five days to under two hours; that is not a field-labor saving.

When

Tested at Angel and launched at North West Shelf. January article: 5 January 2026. AI page last updated 27 May 2026; original publication date and estimate period unknown.

What remains unknown

The earlier audit checked the January article alone. No independent validation, achieved maintenance-hours result or full costs. Global rollout was planned for 2026; completion remains unverified.

What to ask next

Were the recommended intervals adopted, and did reliability hold while maintenance hours fell?

Artificial Intelligence at Woodside
Updated 2026-05-27; publication day unknown · Maintenance section; last-updated stamp at foot of page
Original publisher text read by editor 18 October 2026; claims remain source-reported.

Right fix at right time – AI delivers for Woodside
2026-01-05 · Angel pilot / North West Shelf launch, AWS collaboration and model processing-time paragraphs
Original publisher text read by editor 18 October 2026; claims remain source-reported.

NOV connects two agents to recommend maintenance

The work

NOV describes its agent interfacing with a Norwegian operator’s agent at a sulphate-removal unit.

How it works

Real-time equipment diagnostics on the Max Platform support maintenance recommendations. The release does not establish automatic execution or identify an LLM.

What was reported

No measured accuracy, downtime reduction, savings, cost or ROI disclosed.

When

Release: 28 July 2026. Implementation start and duration unknown.

What remains unknown

Supplier-only account; operator, unit and user population unnamed. Approval, override and safety rules undisclosed. Separate service-rig monitoring in the same release is not evidence of AI.

What to ask next

Who checks the recommendation, and which system can authorize maintenance work?

NOV Reports Second Quarter 2026 Earnings
2026-07-28 · Significant Achievements: Norwegian sulphate-removal unit; separate remote service-rig monitoring paragraph
Original publisher text read by editor 18 October 2026; claims remain source-reported.

Presidio separates field recommendations from controller adjustments

The work

Presidio’s Q2 call distinguishes DOUG recommendations, weekend visit routing and automatic plunger-cycle adjustments.

How it works

DOUG alerts field engineers; pumpers receive a prioritized map. Controllers on 24 wells continuously adjust cycle timing. Management says field judgment remains with staff; controller safeguards are unspecified.

What was reported

Reported gas rates on the controller population: 2,680 to 2,946 Mcf/day, roughly 10%. Weekend production reportedly rose 2.5%. These overlap the existing 2.3% program-uplift claim and must not be added to it.

When

24 controller wells within Presidio’s operated base. Call: 12 August 2026; controller pre/post periods and weekend baseline unknown. Platform scope is a separate population.

What remains unknown

Already-reviewed source, not a new deployment. No matched control, decline adjustment or full costs. Existing $4.5m annualized program revenue is not net cash. Platform coverage (~2,000 call / ~2,300 later goal slide) is not controller coverage.

What to ask next

What happened on comparable wells, and what net benefit remained after controller and field-work costs?

Presidio Production Company Q2 2026 Earnings Call Transcript
2026-08-12 · Chris Hammack remarks, printed page 3: weekend coverage, 24 plunger boxes, human judgment; Will Ulrich page 2 for aggregate uplift/revenue
Original publisher text read by editor 18 October 2026; claims remain source-reported.

SANAD helps drilling engineers find the procedure they need

The work

KOC describes SANAD, a drilling knowledge assistant developed with SLB and piloted by its North & West Kuwait Drilling Group.

How it works

The pilot used 14 North Kuwait procedures, safety material and well records. KOC says its standards team checked every pilot output. OpenWELL context was tested; wider integration is still planned. The engineer retains the decision.

What was reported

A later engineer interview reports one Lower Burgan well finished 13 days early and SOP-related savings of about KD 97,000 per well. It does not isolate SANAD’s contribution; neither figure is counted as an AI result or ROI.

When

Development reportedly began March 2025. KOC describes a 2025–26 pilot; its April–June 2026 issue has no verified posting day. The interview was published 29 September 2026. Field-test dates and well ID are not supplied.

What remains unknown

No active-user count, output error rate, full cost or matched baseline is disclosed. Potential benefits across 47 wells are an extrapolation, not an achieved rollout. CAPEX, stage-gate and production modules remain roadmap items. This is separate from G Agent scheduling in AI-085.

What to ask next

Can engineers trace each answer to the right procedure, and does it improve decisions after review time and errors are included?

SANAD AI: The Generative AI Drilling Assistant
2026-Q2 · PDF pages 28–31 / printed 26–29; 14 SOPs on PDF 30; validation, ownership and roadmap on PDF 31
Editor read the original 52-page PDF and rendered pages 30–31. Issue period is April–June 2026; exact posting date unknown.

KOC’s Sanad uses AI to unlock drilling knowledge
2026-09-29 · Timestamp; lower-Burgan field test, SOP-related savings, RAG description and human responsibility
Editor read the attributed interview. No original KUNA page located in the bounded search. Savings are engineer-reported, not independently validated.

Energy data is becoming something customers’ agents can query

The work

S&P Global Energy describes giving internal and customer agents conversational access to structured energy datasets.

How it works

Domain experts curate narrow dataset agents. A reusable MCP interface combines them across commodities. The account says table permissions and audit controls carry through, and experts can test answers against benchmarks.

What was reported

A customer executive reports setup shrinking from development cycles measured in months to days. No sample, measured accuracy, active-user count, cost or ROI is supplied.

When

Published 25 September 2026. Implementation and external-customer launch dates are not disclosed.

What remains unknown

This is a vendor-published customer account, not an independent quality or security assessment. Benchmarks are described without scores. Access to research data does not establish authority to trade, approve investment or control field equipment.

What to ask next

Who owns the data definitions, and which benchmark and permission checks must pass before a new agent can use them?

From Data to Dialogue: S&P Global Energy’s Genie and MCP architecture
2026-09-25 · Date; per-dataset curation; Unity Catalog permissions; FastMCP composition; customer quote and benchmark loop
Editor read the original vendor customer account. Architecture and benefits remain vendor/customer-reported.

PetroReconcavo connects its field network and back-office systems

The work

PetroReconcavo reports implementing SAP and activating six LTE towers in Bahia and Rio Grande do Norte.

How it works

The ERP work connected production data and material tracking. The field network supports monitoring and prepares for IoT and automation. A separate data-lake passage describes GenAI analytics, reports and assistants without naming a tool or defining live use.

What was reported

The report lists more than 2,200 training hours for SAP. Faster response and stronger controls are qualitative claims; no measured savings, cost or ROI is disclosed.

When

Portuguese report published 14 July 2025, confirmed by issuer notice. It covers 2024; exact go-live dates are unknown, so an in-window deployment is not established.

What remains unknown

No partner, user population, AI approval rules or autonomous control is specified. Tower and training counts are implementation scope, not economic outcomes. Company-wide emissions and social results are unrelated and are not attributed to this technology.

What to ask next

Which field and finance tasks improved after the systems went live, and what did people still have to do manually?

Relatório de Sustentabilidade 2024
2025-07-14 · PDF / printed page 78: SAP, Projeto Waves and data lake
Editor retrieved the original 100-page Portuguese PDF after encoding the plus sign in its download parameter; page 78 read and rendered.

2024 Sustainability Report · Notice to the Market
2025-07-14 · PDF pages 1–2; publication announcement and English-edition timing
Original issuer notice read; Portuguese publication 14 July, English edition separately planned for 25 July 2025.

INPEX changed its core system and the processes around it

The work

INPEX reports that SAP Cloud ERP Private began full operation in Japan in January 2026.

How it works

With ABeam’s support, it aligned business processes to SAP standards instead of filling gaps with custom add-ons.

What was reported

The release says development was shortened, without giving elapsed time or a comparison. Reporting and workload improvements are objectives; no measured savings or ROI are disclosed.

When

Joint release published 24 June 2026; Japan go-live January 2026. Other countries’ rollout dates are not supplied.

What remains unknown

AI appears as compatibility and a future efficiency objective, not a deployed function. Users, departments, costs and approval rules are unspecified. No connection to the separate Copilot/management-meeting agents in AI-294 is established.

What to ask next

Which processes changed, and can the separate AI tools access those records with appropriate permissions?

INPEX commences operations of SAP Cloud ERP
2026-06-24 · PDF pages 1–2; Japan go-live and Fit-to-Standard approach on page 2
Editor read the official three-page PDF and inspected page 2. ERP use is established; AI operation and Copilot integration are not.

A better modeled development plan is a hypothesis to test in the field

The work

An OriGen.AI-authored study models development choices for one unnamed North American reservoir represented by 10 completed wells and four reservoir units.

How it works

The abstract describes generative AI proposing well locations and trajectories, physics-informed production surrogates, and metaheuristic optimization. It does not establish an LLM or chat agent.

What was reported

Authors report about 24% higher expected NPV and roughly 3% surrogate error. These are modeled comparisons, not realized production, cash flow or ROI.

When

Published 3 November 2025, according to the publisher’s indexed metadata. Field-data and test dates are not supplied.

What remains unknown

Editor checked the publisher-indexed abstract; direct page access failed and the full paper was unavailable. Error definition, test split, economic baseline, costs, human review and operator identity remain unverified. One field does not establish repeatability.

What to ask next

Would the preferred plan survive an independent simulator, economic assumptions and engineering review before capital is committed?

Field Development Plan Optimization Under Uncertainty Using Physics-Informed Neural Networks and Generative AI: A Real-Field Case Study in North America
2025-11-03 · SPE-229333-MS, abstract; publisher metadata: Published November 03 2025; conference 3–6 November 2025
Publisher-indexed metadata and complete abstract inspected by editor. Direct OnePetro page failed; full paper, figures and validation protocol not reviewed.

CNRL puts an annual savings figure on hydrogen optimization

The work

CNRL says AI helps optimize hydrogen in its oil-sands upgrading operations.

How it works

The presentation names the task but does not explain the model, inputs, vendor or approval process.

What was reported

The company reports about CAD 8 million in annual savings. This is a gross savings claim, without a disclosed all-cost ROI.

When

Presentation dated 7 November 2025; exact posting and deployment dates unknown. File path is dated December 2025.

What remains unknown

No site, baseline, measurement period or attribution method is given. The CAD 107 million total on the slide mixes savings and revenue from several projects; it is not an AI total. Existing AI-279 reliability results are separate.

What to ask next

How much hydrogen was avoided, against which baseline, and what does the system cost to run?

Investor Open House · 7 November 2025
2025-11-07 · PDF page / printed slide 50: hydrogen; slide 51: separate reliability program; PDF page 97: currency convention
Editor read original source; company/vendor claims are not independently validated.

An “autonomous” drilling product was used in advisory mode

The work

Baker Hughes describes an advisory-mode UAE carbonate-drilling case supported from its remote operations centre.

How it works

AI assisted torque-and-drag work. Despite the product name, the case explicitly says advisory mode; who approved parameter changes is not stated.

What was reported

The vendor reports 22 alerts, 22 hours saved against a previous well, 5% higher penetration rate and no stuck pipe or fluid losses.

When

PDF revised April 2026, with a 2023 copyright. Original publication and field-work dates are unknown; this may be older operating context. The September 2026 webpage date was not independently verified.

What remains unknown

The customer and well are unnamed. A prior-well comparison does not isolate AI from the wider automation service, crew or geology. Costs, accepted recommendations and repeatability are undisclosed; no ROI is established. The deployment date is unknown; the PDF revision does not establish a 2026 deployment.

What to ask next

Which recommendations changed drilling practice, and did the time saving repeat on comparable wells?

Kantori drilling advisory case · UAE
2026-04 · Single-page case sheet; Solution, Results and footer revision 04/2026
Original PDF read and rendered. HTML case page returned 403 to editor. Revision date is not a field-deployment date.

PDO contracted for a shared industrial data and workflow system

The work

PDO selected Kongsberg Digital’s Kognitwin to connect operational and enterprise data across operations and maintenance.

How it works

The release calls it AI-integrated, but does not identify an AI task, model or authority to act.

What was reported

Safety, production and cost improvements are goals. No measured benefit, contract value or ROI is disclosed.

When

Announced 8 September 2025, with an ambition to start deploying the twins that year. No later completion evidence was verified in this pass.

What remains unknown

No fields, facilities, users or workflow counts are identified. The company-description paragraph about gas operations does not define deployment scope. A passed target date alone does not establish delay or failure.

What to ask next

Which facilities are live, which tasks use AI, and what do people still approve?

Kongsberg Digital and PDO digital-twin agreement
2025-09-08 · Release timestamp 8 September 2025; scope and deployment-ambition paragraphs
Editor read original source; company/vendor claims are not independently validated.

PTTEP’s CEO describes AI in drilling, alongside planning and production systems

The work

At its March AGM, PTTEP described deployed AI and automation across subsurface, drilling and production work, including AI that monitors drilling data for early risk detection.

How it works

The other examples are automated planning of well locations and wellhead platforms, and systems optimizing production across many wells. Their specific AI methods are not identified.

What was reported

Management says benefits are reaching financial performance, but gives no isolated AI result, baseline, cost or ROI.

When

Meeting held 30 March 2026; minutes posted 10 April, confirmed by the cover letter. Individual deployment dates are unknown.

What remains unknown

No fields, vendors, well count, model performance or approval/control boundaries are disclosed. The separate USD 328m new-business budget is not AI-only. This broad account does not establish the scope or success of the later offshore advisory pilot in AI-067.

What to ask next

Which fields use these systems, what do people approve, and what operating result survives a matched comparison?

Minutes of the 2026 Annual General Shareholders’ Meeting
2026-04-10 · PDF page 13 / printed page 12 of 25, response to question 1.7; cover letter on PDF page 1 confirms publication date
Editor read the official translated minutes and inspected the rendered passage. Company self-report, not independent outcome validation.

Skills and data remain barriers to using AI in energy operations

The work

The IEA’s 2026 follow-up to Energy and AI separates data-centre electricity demand from using AI to improve energy operations. Its company survey identifies internal digital skills as the largest adoption barrier.

How it works

Chapter 9 also discusses fragmented data, data protection and equipment digitalisation. Figure 9.4 draws on the IEA Industry Employment Survey 2025.

What was reported

All data centres used an estimated 485 TWh in 2025; the report projects roughly 950 TWh in 2030. Those are not oil-and-gas productivity results or AI-only electricity demand.

When

Published 16 April 2026, verified on the official page. Survey vintage: 2025; forecast horizon: 2030.

What remains unknown

Broad energy-sector evidence; the reviewed chapter supplies no survey respondent count or upstream-only subgroup. Projections are not realized benefits or ROI. This is a distinct follow-up to the 2025 report INS001, not a replacement.

What to ask next

Before buying another model, which missing skills, data connections or equipment would prevent people from using it?

Key Questions on Energy and AI
2026-04-16 · Official report page: Published 16 April 2026; PDF page 10: demand outlook; PDF/printed page 84: Figure 9.4; pages 80–85: adoption barriers
Editor read the official page and relevant full-PDF passages, including the rendered Figure 9.4. IEA report text is CC BY 4.0; Annex A has a separate noncommercial share-alike licence.

Key Questions on Energy and AI — full report
2026-04-16 · PDF page 10; PDF/printed page 84, Figure 9.4; adoption discussion pages 80–85
Relevant original PDF passages read and chart page visually checked. Survey respondent count not supplied in the reviewed chapter; no upstream-specific failure or adoption rate inferred.

TotalEnergies reportedly dropped one drum-cycling use case

The work

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.

How it works

Drum switching depended on conditions outside the model’s scope. The work provided advice, not autonomous control.

What was reported

The report does not define the accuracy denominator or test set, or quantify cost or loss. Other pilot work continued; the entire program was not reported abandoned.

When

Drum work started July 2025; remarks reported June 2026. Exact decision date unknown.

What remains unknown

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.

What to ask next

Did the model help operators choose when to switch drums, rather than merely predict drum level?

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
Editor read the original conference report and both partners’ earlier releases. The decision to drop this use case rests on reported named remarks, not a company-authored outcome report.

TotalEnergies and Honeywell pilot AI-assisted control room
2025-11-11 · Initial DCU pilot and preliminary five-event result paragraphs
Original issuer release read. Establishes the existing pilot and operator corrective action; it predates and does not corroborate reported drum-use-case non-retention.

ExxonMobil has two FlexRobotics systems deployed; seven more are planned

The work

H&P reports two FlexRobotics systems deployed on its rigs for ExxonMobil. It expects seven more over the following twelve months, for nine in total.

How it works

The robots automate repetitive rig-floor activities. The announcement does not identify an AI/ML mechanism, task-level approval, supervision or stop authority.

What was reported

Two is the deployed count; nine is the intended total. No quantified safety, time, cost or net return is disclosed.

When

Published 7 October 2026. Initial deployment dates, basin and rig identities are not supplied.

What remains unknown

An ExxonMobil executive supports expansion intent; the joint account is not independent outcome validation. It does not establish installation of the separate AI camera add-on in AI-116. Keep as robotics context, outside AI-adoption counts.

What to ask next

What changes in crew exposure and connection performance after deployment, under comparable rig conditions?

H&P & ExxonMobil Expand Deployment of FlexRobotics Technology
2026-10-07 · Opening release paragraphs: first two systems, seven additional units over next 12 months; ExxonMobil quotation
Original H&P release read by editor; customer quotation corroborates intent, not quantified results.

Earlier formation images are now reported from more than 50 wells

The work

SLB says Retina has imaged more than 50 wells across 18 locations during its first year of commercial deployment.

How it works

The instrument acquires formation images at the bit while drilling, at roughly 2 mm resolution, to inform well placement and development decisions.

What was reported

Deployment and imaging specifications are not measured recovery, drilling-time savings or ROI. No comparative economics is disclosed.

When

Published 15 October 2026; individual deployment dates and customer identities unknown. Publication time cannot be placed after the previous update.

What remains unknown

No AI/ML mechanism or automatic control is identified. Keep as instrumentation context; do not transfer OQ Oman’s separate drilling-tool results to Retina.

What to ask next

Does earlier imaging change a placement decision or avoid a cost, compared with the previous workflow?

Retina At-Bit Imaging System Surpasses 50-Well Milestone in First Year of Deployment
2026-10-15 · Published date; Global adoption and Global deployment sections
Editor read the original SLB feature; performance and scope remain vendor-reported.

Repsol completes an initial phase of an autonomous glycol-plant optimizer

The work

Repsol says it completed the initial phase of its first Autonomous Optimizer with Control, developed internally for one section of a glycol plant.

How it works

The report calls the optimizer AI-based. It does not identify the plant, model, controlled variables, operating limits, approval steps or fallback.

What was reported

No measured plant result, cost or net return is disclosed.

When

2025 annual report, released 19 February 2026. Project and commissioning dates are not supplied.

What remains unknown

Completing an initial phase does not establish routine autonomous operation or a plant-wide rollout. Keep separate from Repsol’s multiagent, workforce and legal programs.

What to ask next

What did the initial phase demonstrate, and what must operators approve before it can control the process?

2025 Group Annual Financial Report
2026-02-19 · PDF page 195 / printed Management Report page 45, section 5.5: AI-based autonomous optimizers
Editor retrieved the original PDF, read the relevant text and inspected the rendered page on 15 October 2026.

Well designs and drilling data are being connected in both directions

The work

The vendors have developed an integration between Oliasoft WellDesign and Corva Real-Time AI and say they are preparing customer pilots.

How it works

Design trajectories, casing and pressure assumptions can inform drilling workflows. As-drilled trajectories, operational data and nonproductive-time events return to planning and model calibration.

What was reported

No pilot customer, well or rig count, measured benefit, implementation cost or ROI is disclosed.

When

Announced 12 October 2026; pilot start dates are unspecified. Newly found evidence, not a disclosure after the previous update.

What remains unknown

Model feedback is not evidence of automatic rig control. Two vendor versions of the same release do not provide independent validation; action permissions are unspecified.

What to ask next

Can the pilot show fewer repeat design errors or faster decisions, with engineers able to review the changes?

Corva and Oliasoft Create a Continuous Learning Loop Between Well Design and Drilling
2026-10-12 · Joint release: engineering-to-operations, operations-to-engineering, and customer-pilot paragraphs
Editor read both vendors’ original releases on 15 October 2026; neither supplies customer-pilot outcomes.

Corva and Oliasoft Create a Continuous Learning Loop Between Well Design and Drilling
2026-10-12 · Joint release: engineering-to-operations, operations-to-engineering, and customer-pilot paragraphs
Original partner release read; same joint announcement, not independent validation.

A forecast can look accurate and still miss the operating change that matters

The work

A UNISIM-IV study finds that production-history-only forecasts miss changes driven by operating events, interactions between wells and constraints.

How it works

Its BiLSTM illustration omits operating context from the baseline. The authors emphasize input design and constraints over a more elaborate architecture.

What was reported

Aggregate forecast error can hide important operating errors. Numerical effect sizes and test splits were not verified.

When

October 2026 issue; exact online publication date unknown. It cannot be dated after the previous update.

What remains unknown

Publisher abstract and indexed section text only; full paper unavailable. This is a synthetic benchmark, not a named failed field deployment or an economic result.

What to ask next

Does the forecast see interventions and constraints, and how does it behave when the operating regime changes?

Data-driven production forecast for oil and gas reservoirs - Advantages and limitations
2026-10 (issue month; exact online date unknown) · Volume 265, article 214569; DOI 10.1016/j.geoen.2026.214569; publisher abstract and indexed sections, no PDF page locator
Editor independently read publisher-indexed abstract and section text; direct retrieval failed, and no full paper or numerical tables were obtained.

Keep an inferred asset record distinguishable from an original record

The work

A February lab examined a GenAI pilot that structures legacy asset records, fills gaps and builds an asset hierarchy.

How it works

Participants called for separating authoritative records from AI inferences, testing across asset conditions, showing uncertainty and preserving change histories and human oversight.

What was reported

No asset count, measured reliability improvement, cost or ROI is published.

When

Lab held 25 February 2026; summary published 29 April 2026.

What remains unknown

An anonymized discussion, not a system audit, observed failure, regulatory endorsement or compliance advice. Energy-infrastructure context; no petroleum operator is identified. Findings reflect participants’ views, not Ofgem policy.

What to ask next

Can an engineer tell what came from a record, what was inferred, and who accepted the change?

AI Reg Lab: High-level findings for licensees and stakeholders — February 2026
2026-04-29 · PDF page 1: publication date; page 2: pilot and lab scope; page 3: themes; page 4: disclaimer
Editor read the original four-page PDF on 15 October 2026. © Crown copyright 2026; document reused under the Open Government Licence.

GEORGE helps production engineers work across thousands of wells

The work

SPE reports that Chevron’s Peter Nesom described GEORGE, a production-engineering assistant used for analysis, optimization and automation tasks on more than 5,000 wells.

How it works

The report describes an assistant working with production engineers across several problems. It does not specify its models, tools, data connections, approval steps or control-system permissions.

What was reported

More than 5,000 wells is reported use scope. No production uplift, time saving, cost or net return is supplied.

When

Remarks at the 25 August 2026 plenary; SPE report published 4 September. Launch date and active-well denominator unknown.

What remains unknown

Original conference reporting of named operator remarks, not an operator-authored deployment paper. The event program corroborates speaker and session only. Keep as reported use awaiting technical corroboration, outside additional adoption/result counts. Do not infer an LLM architecture, autonomous field control or a company-wide knowledge system.

What to ask next

Which engineering tasks does GEORGE complete, what must engineers approve, and what changes in field performance?

Maximizing Every Barrel: Artificial Lift Embraces Automation
2026-09-04 · Chevron section: Peter Nesom remarks following the image caption; 25 August 2026 plenary, reported 4 September
Original SPE/JPT report read by editor 15 October 2026; named operator remarks reported by the publisher, not issuer technical validation.

SPE Artificial Lift 2026 opening plenary program
2026-08-25 (event date; page publication date unknown) · Opening plenary: Peter Nesom, Chevron Global Upstream Production Engineering CoE
Official event page read; corroborates speaker identity and session, not GEORGE scope or performance.

A reservoir assistant is tested as an interface to technical systems

The work

A Petrobras-coauthored paper proposes a GPT-based assistant for reservoir-management systems, intended to let engineers use natural-language requests instead of navigating technical interfaces.

How it works

The abstract describes model training and tests of domain-specific requests, including attempts to avoid sensitive or irrelevant responses. It does not identify the model provider, connected applications, tool permissions or production deployment.

What was reported

The authors report favorable test behavior and training loss approaching zero. No held-out task-success denominator, active users, field result or economics is disclosed. Training loss is not proof of reliable operating decisions.

When

Published 21 October 2025, OTC Brasil. Test dates and subsequent rollout unknown.

What remains unknown

Publisher-indexed original abstract reviewed; full paper not retrieved. Retained as research, outside operating adoption and outcome counts. No evidence of permission to change reservoir plans or execute downstream actions.

What to ask next

Does it perform useful tasks reliably on unseen requests, preserve access rights and leave an auditable record?

Virtual Assistant Using Generative Pre-Trained Transformers on Reservoir Management Context
2025-10-21 · OTC-36185-MS; publisher-indexed author list and abstract, test-scenario paragraphs
Original publisher-indexed abstract read by editor 15 October 2026; direct page retrieval failed and the full paper was not retrieved.

ESP recommendations pass safety checks before reaching the controller

The work

ExxonMobil authors describe a closed-loop ESP pilot on 26 Midland wells. Commercial physics-based software recommends changes; internal checks test facility constraints and pump limits before execution.

How it works

Recommendations that pass the checks can be written automatically to field devices from different controller makers, without a person approving every change. The abstract does not identify machine learning; an AI keyword does not establish the deployed method.

What was reported

The authors describe less surveillance and execution work, without numerical time, production or financial results. No complete cost or net return is supplied.

When

Paper published 18 August 2025; pilot dates unspecified. A September 2026 conference report describes wider Permian rollout underway, without a completed-fleet count.

What remains unknown

Original publisher-indexed abstract reviewed; full paper not retrieved. It establishes physics-based control and a validation gate, not an LLM controlling pumps. This is separate from AI-235’s gas-lift ML and its oil-uplift figures.

What to ask next

What happens when data is stale, a recommendation fails a check, or operating conditions move outside the validated range?

Closed-Loop ESP Optimization to Maximize Asset Value
2025-08-18 · SPE-225261-MS; abstract, system checks and 26-well pilot paragraphs
Original publisher-indexed abstract read by editor 15 October 2026; direct page retrieval failed and the full paper was not retrieved.

Maximizing Every Barrel: Artificial Lift Embraces Automation
2026-09-04 · ExxonMobil / Mark Agnew remarks, closed-loop ESP rollout and facility constraints paragraphs
Original SPE/JPT conference report read. Later rollout is reported underway; no completed-fleet count or quantified ESP benefit. Separate from the 2025 pilot abstract.

Aethon’s CIO describes the foundations for better data use

The work

Aethon CIO Kang Chen reports a full cloud migration and technology overhaul, along with new data-quality, analytics and workforce-training initiatives.

How it works

The profile also describes cybersecurity benchmarking. It names no cloud provider, data domains, field applications or AI model.

What was reported

No quantitative operating or financial result is supplied.

When

20–26 February 2026 issue. Individual project dates and completion of the data initiatives are unknown.

What remains unknown

Named executive self-report in an advertising supplement, not independent technical validation. Cloud migration and upskilling do not prove AI readiness or deployment. Keep as company transformation context, outside AI adoption and outcome counts.

What to ask next

Which field and business data can teams actually reuse, and what applications have they put into daily work?

Dallas CIO ORBIE special advertising supplement
2026-02-20 to 2026-02-26 · PDF page 10 / printed page 10, Kang Chen profile; supplement dated 20–26 February 2026
Original advertising-supplement PDF text read by editor 15 October 2026; named executive self-report, not independent validation.

One oilfield connects reservoir, production and equipment data

The work

CNOOC says Qinhuangdao 32-6 in Bohai uses more than 26,000 data-acquisition points in one platform for reservoirs, injection, production and equipment. It diagnoses production abnormalities and generates optimization plans.

How it works

More than 26,000 acquisition points feed the platform. The company places this field example within its broader AI+ and data-efficiency program. The field passage describes connected instrumentation, diagnostics and plan generation, but does not identify the model or how much of the improvement comes from AI. Automatic plan generation does not establish automatic execution or operator approval.

What was reported

CNOOC reports 30% less time for reservoir-potential studies, more than 90% early-warning accuracy for critical equipment, and more than RMB 10 million lower annual operating and maintenance costs. These are integrated-program results, not isolated AI benefits or net ROI.

When

2025 reporting period; report released 9 April 2026. Start date, observation periods and deployed user count are not specified. More than 26,000 points measures instrumentation, not people or wells.

What remains unknown

Company-reported results. Study-time baseline, warning-accuracy denominator, cost calculation, implementation cost and independent validation are absent. No disclosed execution authority or safety boundary. This expands the already-named Qinhuangdao component of AI-145; it is not a second project or a new counted numerical-result profile. AI-082 remains separately unresolved.

What to ask next

Which plans do engineers accept, what do false alarms cost, and how much of the reported saving survives integration and support costs?

2025 Environmental, Social and Governance Report
2026-04-09 · Printed page 23 / PDF page 28, Qinhuangdao 32-6 case; report index dates publication 9 April 2026
Parent retrieved original PDF and checked the relevant page on 15 October 2026; results remain source-reported.

A single model helps decide where the molecules should go

The work

Suncor describes one linear-programming model connecting its oil-sands operations, Edmonton refinery, domestic markets and global export sales. During an upset, teams can change the crude slate, move diluent and route intermediate streams.

How it works

The model supports regional choices across connected assets and markets. Dave Oldreive calls it a big AI tool; the disclosed method is linear programming. No machine learning or LLM is established. Management describes integrated teams maximizing value, but does not specify who approves individual model recommendations.

What was reported

The stated 10,000–15,000 barrels/day is intermediate-stream routing capability. It is not incremental AI production, savings or measured uplift. Record throughput and sales are not assigned to the optimizer.

When

Present-tense operating description on the 5 August 2026 call, discussing Q2 adjustments. This passage was already in our earnings-call ledger; it is now easier to find in the operating-technology view.

What remains unknown

No architecture beyond linear programming, counterfactual, model-specific benefit, implementation cost or detailed control boundary. Keep it outside new AI-adoption and result counts. Suncor haulage and truck dispatch remain separate existing records.

What to ask next

Can a team optimize the whole asset network rather than improving one unit at the expense of another?

Second Quarter 2026 Financial Results Call
2026-08-05 · PDF page 10 (page index 9), Q&A lines 331–358, Dave Oldreive and Rich Kruger; call dated 5 August 2026
Parent read the relevant original issuer transcript passage on 15 October 2026; results remain source-reported.

Purchasing automation has its own value case

The work

Petrobras reports automating purchase requests to speed up purchasing and avoid costs. The same slide also describes AI projects, but this purchasing box does not identify AI.

How it works

The slide names request automation without explaining the software, process steps, business-unit coverage, approval rules or exception handling. We retain it as purchasing automation, separate from the neighboring well, forecasting, flare and control-room AI examples.

What was reported

Petrobras reports 190,000 items automated per year and average avoided costs of R$111 million per year. Avoided cost is not the same as cash collected or net savings after implementation and support costs.

When

Business plan published 27 November 2025. The slide gives annual figures without a measurement period, start date or rollout denominator.

What remains unknown

Company-reported values without a disclosed baseline, cost method, implementation cost or independent validation. The slide does not establish whether the estimate is realized or forecast. Do not count this as AI adoption or add it to the separate AI benefits.

What to ask next

Which purchasing steps disappear, which approvals still need people, and how is avoided cost calculated?

2026–2030 Business Plan
2025-11-27 · Printed slide 158 / PDF page 158, purchase-request automation box; separate from the AI and digital-twin boxes
Parent retrieved original PDF and checked the relevant page on 15 October 2026; results remain source-reported.

Employees build agents for research, rules and management meetings

The work

INPEX uses Copilot for research and shared agents for company rules, regulatory searches and contract knowledge. A management-meeting agent checks uploaded materials and suggests questions.

How it works

SharePoint, OneDrive, Office and Teams provide work context. Employees can create simpler agents; advanced Studio agents using external connectors or triggers require a cost-and-access approval process. Training videos and departmental ambassadors support use.

What was reported

Microsoft reports 1,450 paid licenses, over 1,000 agents and 130 research requests, most completed the same day versus previously weeks. A survey found 44% perceived less task time. INPEX’s estimated annual value of over JPY2bn covers company AI use; it is not Copilot-only cash savings.

When

Published 1 September 2026; Copilot became the core platform in 2025. Exact measurement windows, active-agent count, survey size and value calculation are not supplied.

What remains unknown

Vendor story with named customer speakers. Report throughput does not establish report accuracy. License/agent counts do not measure sustained use. The company-wide value estimate may overlap other AI work, including Ichthys, and has no disclosed net-return method.

What to ask next

Which agents do people keep using, and does better research or meeting preparation change a decision?

INPEX transforms business processes with Microsoft 365 Copilot and AI agents
2026-09-01 · Research tool; Building a foundation for change; Looking beyond estimated annual value
Relevant original source read during the 14 October evening review; source claims are not independently validated.

A field technician turns equipment reports into turbine alerts

The work

A Williams technician built an anomaly detector that reads emailed equipment reports, compares them with operating baselines and flags changes in temperature, oil pressure or vibration.

How it works

The Microsoft story names Copilot Cowork and Outlook distribution-list reports. Staff also create agents for procedures and safety paperwork. The technician receives alerts and can investigate; automatic equipment control is not described.

What was reported

More than 90% organizational AI adoption is repeated from earlier coverage. No detector-specific uptime, false-alert, safety or financial result is supplied.

When

Published 9 September 2026; the existing profile retains Williams’s June company-wide adoption disclosure. The detector’s start date and asset count are unspecified.

What remains unknown

Midstream operating example, not an upstream production result. Vendor account without disclosed detector validation, alert threshold, escalation protocol or before/after equipment comparison. Organization-wide adoption is not this detector’s user share.

What to ask next

Do alerts catch faults earlier than normal monitoring, and how many alerts require unnecessary investigation?

Williams transforms frontline work with Copilot, gaining 90% adoption
2026-09-09 · Bringing AI to the frontline first; Accelerating operations with Copilot
Relevant original source read during the 14 October evening review; source claims are not independently validated.

Egyptian drilling teams expand an edge-based warning system

The work

NOV says Drilling Beliefs & Analytics expanded from a two-rig trial to double-digit rigs in Egypt’s Western Desert. An RTOC engineer monitors several rigs, with crews acting on warnings.

How it works

Bayesian belief indices use drilling parameters to flag dysfunction. Edge processing supports use with limited connectivity. An overpull warning led staff to correct drill-off procedures; a separate autozero warning led a crew to reset bit weight.

What was reported

NOV estimates $120,000 avoided in the stuck-pipe case. A separate autozero case reports 8.2 hours and about $26,000 saved. These are distinct from the earlier $75,000 bit-balling estimate in AI-158; they are not added together.

When

Article published 11 March 2026. Egypt introduction dates to 2023; the source does not date each intervention or disclose exact rig/well counts.

What remains unknown

Operator names remain anonymous. Event-level vendor estimates lack an independent counterfactual or full implementation costs. No autonomous control or fleet-wide cash ROI is established.

What to ask next

How often are warnings actionable, and does the fleet show lower downtime after accounting for differences between wells?

NOV’s Drilling Beliefs & Analytics advances digital operations in Egypt
2026-03-11 · 11 March 2026 article: trial expansion, stuck-pipe example and Bit Weight Autozero example
Relevant original source read during the 14 October evening review; source claims are not independently validated.

Closed-loop drilling and pressure management on one Indonesian well

The work

Halliburton describes LOGIX coordinating drilling, tripping and managed-pressure drilling on an Eni deepwater exploration well offshore Indonesia.

How it works

Rig equipment, automated well placement, downhole hydraulics and pressure controls operate in a connected closed loop. Human approval, handoff and safety-interlock arrangements are not detailed.

What was reported

Supplier reports over 15% better efficiency, without defining the efficiency measure, baseline or absolute time/cost effect.

When

9 July 2026 announcement; one exploration well supports the reported result, not a fleet-wide rollout.

What remains unknown

The release does not identify the AI/ML elements. Retained as operating technology, outside the AI-adoption and result counts.

What to ask next

Which decisions are automated, and how much useful well-delivery time changes under comparable pressure conditions?

Eni and Halliburton achieve industry first with closed-loop rig automation in deepwater Indonesia
2026-07-09 · 9 July 2026 release: deployment, integrated controls and reported efficiency
Relevant original source read during the 14 October evening review; source claims are not independently validated.

An energy investor backs sites, power and connectivity for AI demand

The work

EnCap’s Energy Transition arm backed Quantica to combine land, renewable and grid power, and network connectivity for data-center sites.

How it works

Integrated site development serves hyperscale and AI-compute customers. The source describes an existing project pipeline without quantifying it.

What was reported

No completed-project, deployed-load, AI productivity or financial result is reported. Leadership’s prior 15GW experience is not Quantica’s delivered capacity.

When

Platform launch announced 28 July 2025; rollout and commissioning dates are unspecified.

What remains unknown

AI is a customer-demand driver here, not a technology shown improving a portfolio operator’s work. Keep this capital-allocation evidence separate from AI adoption.

What to ask next

Which sites have secured power and customers, and how much announced capacity becomes operating load?

EnCap launches Quantica Infrastructure
2025-07-28 · 28 July 2025 launch release; platform scope and existing pipeline
Relevant original source read during the 14 October evening review; source claims are not independently validated.

Powered-site development is a separate kind of AI exposure

The work

NGP identifies Cloverleaf as a portfolio company since 2024, developing powered sites for large data centers. Its current profile reports over 5GW developed and sold.

How it works

Site selection, power access and community work support data-center customers. The CEO describes AI demand, market saturation and local collaboration as business conditions.

What was reported

Over 5GW developed and sold over the last two years, according to the undated sponsor page. This is not commissioned capacity, active AI load or an AI-use result.

When

Undated page reviewed 14 October 2026. Its relative two-year period cannot be anchored to a known publication date.

What remains unknown

No site/customer count, utilization, sale economics or AI-specific revenue allocation. The sponsor relationship does not establish AI adoption by NGP’s upstream holdings.

What to ask next

What distinguishes a sold development project from an energized data center serving customers?

NGP portfolio profile: Cloverleaf Infrastructure
Undated · Portfolio Profile / Cloverleaf Infrastructure
Relevant original source read during the 14 October evening review; source claims are not independently validated.

Plans for agent authority are ahead of reported controls

The work

BCG’s survey finds that 42% expect agents to have decision authority by 2030, while 5% report having all the relevant controls today.

How it works

The article covers over 1,300 senior leaders across more than 20 sectors; its companion release specifies 1,330. It argues for organizational ownership, governed actions and business-outcome measurement.

What was reported

42% is a future intention; 5% is a present control-readiness claim. They are not the same measure or an estimate of unsafe deployments.

When

Published 30 September 2026. Survey fieldwork dates and an oil-and-gas denominator were not established in the reviewed material.

What remains unknown

Self-reported cross-sector evidence and maturity classifications, not causal proof or audited project ROI. The reviewed article highlights power/utilities and chemicals but does not provide an upstream-only comparison.

What to ask next

Before granting an agent authority, who owns its decision, what limits its actions, and how can a person reverse them?

The Formula for Agentic AI Value / Applied AI Index 2026
2026-09-30 · Key takeaways; study scope; FAQ on agent autonomy
Relevant original source read during the 14 October evening review; source claims are not independently validated.

Portfolio CEOs report more deployment, without an energy breakdown

The work

Blackstone says nearly half of surveyed portfolio CEOs report AI in production or fully deployed, up from under 30% a year earlier. Eighty percent plan to increase spending.

How it works

The 2Q26 survey covered 117 portfolio companies, including 73 US CEOs, largely in private equity and credit. The 2Q25 comparison covered 87 companies; respondents changed.

What was reported

Deployment stage and planned spending are reported; productivity, cash returns and energy-company adoption are not measured here.

When

Published 5 August 2026. The 2Q26 survey ran 9–30 June 2026; 2Q25 ran 5–19 June 2025.

What remains unknown

Sponsor-reported, nonrepresentative subset with changing composition and no energy subgroup. Blackstone’s own token consumption is a separate internal-use measure, not a portfolio operating result.

What to ask next

Which portfolio workflows are in use, and what changes in operating or financial results after deployment?

Pattern Recognition / AI Adoption Is Accelerating
2026-08-05 · 5 August 2026 section; 2Q26 / 2Q25 CEO Survey footnotes
Relevant original source read during the 14 October evening review; source claims are not independently validated.