Three workflow profiles added, Santos’s existing profile expanded, and two industry reports added. CNX methane was already in the research backlog. Undated vendor claims and potential benefits keep their qualifications.
Three targeted Luna tracks followed by editor checks of original sources and existing dossiers. A bounded pass, not a census of every disclosure.
PetroChina / Company reports scaled use on more than 3,800 wells
A model that diagnoses rod-pumped wells in Xinjiang
CNPC says a production-diagnostics model is used on more than 3,800 rod-pumped wells in Xinjiang. Staff receive an abnormal-condition diagnosis, the evidence behind it and suggested actions.
How it works
A one-billion-parameter Transformer model with contrastive learning uses production data from 78,000 rod-pumped wells across Daqing, Liaohe, Changqing and Xinjiang. Models are fine-tuned for operating scenarios. The first deployed scenario diagnoses well conditions, colors a health display and supplies evidence and recommendations. Online liquid measurement and parameter recommendations are expanding scenarios; broader embodied-agent research is future work.
What was reported
Company reports average condition-diagnosis accuracy above 90%, abnormal-well discovery moving from days to minutes, and wells managed per person increasing from 3.5 to 11.5. It also reports usable data rising from 40% to above 95%. Validation definitions, samples, observation periods and staffing populations are not supplied.
When
15 April 2026 disclosure: more than 3,800 Xinjiang wells in use. It describes coverage of light-oil rod-pumped wells at named plants including plants 1 and 2 and Baikouquan. The 78,000-well figure concerns data from four oilfields, not the live rollout.
What remains unknown
Company-reported results without independent validation or complete cost/return data. Do not treat training-data population as deployment reach, multiply staffing ratios into cash savings, or assume automatic control. A Transformer industrial model is not established as a text LLM or a company communication brain.
What to ask next
How does diagnosis hold up across lift conditions, and which recommendations lead to useful field interventions after false alerts and support work?
Finding methane-release precursors and prioritizing repairs
CNX describes ML on well and equipment signals to identify conditions preceding methane releases. ML-assisted camera and aerial analysis helps prioritize sources, while operating data is reconciled with the emissions inventory.
How it works
Well/equipment performance data is used to identify release precursors. OGI camera video and aerial sensor data help identify sources and prioritize high-emitting components. Integrated operating and emissions-inventory data surfaces anomalies. The surrounding monitoring program includes remote equipment shutdown and in-house leak-tracking/work-order software; the source does not establish autonomous AI shutdown.
What was reported
CNX describes qualitative success in Q1 2026. No AI-specific event count, avoided methane volume, precision/recall, repair-time change, asset population, implementation cost or net return is disclosed. Earlier company-wide emissions reductions are not AI results.
When
Undated current company page, with explicit 2026 program and Q1 2026 progress. No model launch date or deployment population is given.
What remains unknown
This is a verified October 9 research backlog item, separate from AI-288 safety cameras. The company’s 2020–24 emissions reductions and its 2025 reporting-baseline reset cannot establish AI-attributable improvement. Remote control capability is not evidence of autonomous AI authority.
What to ask next
Which release precursors are detected reliably, how many interventions prevent a release, and how much inspection and repair work remains?
Preisser describes an invoice assistant for HG Oil that extracts and codes items, flags anomalies and routes approvals. Low-confidence extractions go to people for review.
How it works
Staff upload invoices; the vendor describes extraction of vendor details, line items, totals and GL codes, expense categorization, anomaly flags and approval routing by thresholds and cost centers. A human exception queue handles low-confidence extractions. A separate vendor product page describes human payment approval and names HG as its reference; that control is not stated in the named case alone.
What was reported
Vendor reports 75% less manual invoice handling, more than 40 staff hours/week redirected and an avoided administrative hire. Its baseline is also described as more than 40 hours/week, but the page does not provide a common measured period or exact time denominator. These are reported labor/capacity claims, not audited cash savings.
When
Both vendor case and operator About page are undated. The operator identifies itself as a Kansas producer; its company-wide well count is not the invoice-system population. The disclosure cannot establish an in-window deployment date.
What remains unknown
Vendor-only AI deployment and benefit claims; HG’s own website corroborates operator identity, not this automation. No invoice volume, error/exception rate, staffing denominator, matched time study, costs or independently validated return. Do not calculate dollar ROI or date the deployment from search metadata.
What to ask next
How many invoices pass without correction, what review effort remains, and does released capacity reduce costs or improve the close?
Read AI-292Santos / Specific workflows now reviewed; achieved benefits and targets separated
Production AI, virtual pressure measurement and drilling controls
Santos’s deck has more specific evidence than our earlier excerpt. It separates production-optimization AI, a pressure model replacing physical gauges, rig-connected drilling automation and stimulation logistics.
How it works
The pressure model estimates bottom-hole pressure instead of relying on physical downhole gauges. A separate drilling-automation column describes rig-control integration monitored from an operating center. The production-optimization AI column reports delivered benefits. The fourth column concerns bulk proppant/fluid handling and is not identified as AI.
What was reported
Production-optimization AI: $28m benefits delivered in 2025 and over $64m potential annual benefit by 2029. Pressure model: up to about $70m potential capex reduction. The two potential dollar figures are gross. Separate columns claim 20% penetration-rate improvement and 80% US frac-installation productivity improvement; these must not be assigned to the pressure model.
When
26 May 2026 investor briefing; the deck reports 2025 production-AI benefits and 2029 potential. The stimulation column compares 2026 productivity with a 2023 baseline. No well/rig population or model-specific measurement period is supplied.
What remains unknown
The pressure-model figure is potential, not achieved savings. No pressure-model accuracy, complete project costs or matched net return. The production-AI program may overlap AI-083/Xecta, so its benefit remains outside numerical-result counts pending scope reconciliation. Drilling integration belongs to its own column; stimulation logistics is adjacent technology.
What to ask next
Where is virtual pressure measurement accepted as a replacement for gauges, and which costs and operating limits determine the realized benefit?
Read AI-139Read AI-083KPMG International / Survey of larger companies; priorities are not realized returns
Technology priorities, with an oil-and-gas subgroup
The report includes 58 oil-and-gas technology leaders within a 258-person energy/extractives sample. It helps compare priorities, but does not measure whether a particular AI project paid off.
How it works
Respondents represent organizations above US$1bn annual revenue across 22 countries and territories. The broader sample also includes mining, chemicals, utilities and renewables. The oil-and-gas subgroup and broader energy findings have different populations.
What was reported
79% of the oil-and-gas subgroup prioritizes AI/automation; the broader energy/extractives figure is 69%, versus 77% across sectors. The report says IT leads AI strategy at 58% of the broader energy group. These are self-reports, not operating measurements.
When
2026 report edition; exact publication day and survey fieldwork dates were not established in this review.
What remains unknown
Small oil-and-gas subgroup; companies below US$1bn are outside the stated population. AI/automation includes multiple technologies. Expected future scaling is not achieved deployment. Do not apply these percentages to our project register or infer net ROI.
What to ask next
Do the priorities translate into sustained use and measured outcomes on comparable assets?
KPMG US / Broader-energy survey; no upstream-only deployment rate
US energy companies’ reported technology maturity
KPMG’s US energy report adds a broader view of the gap between technology ambition and implementation. It covers 100 US energy-business executives within a global survey.
How it works
The report describes 2,500 executives across 27 countries, including 100 from US energy businesses. Its maturity comparisons concern multiple technologies and energy subsectors. It is a separate publication from the global energy/extractives report.
What was reported
52% of energy respondents describe advanced AI/automation implementation, versus 57% elsewhere. For data/analytics the comparison is 45% versus 60%. These are reported implementation stages, not audited business results.
When
2026 report edition; exact publication day not established.
What remains unknown
No separate upstream or oil-and-gas denominator is supplied for these comparisons. Do not merge the US sample with the 258-person global energy sample; overlap is not established. Technology maturity does not establish AI-only cash returns.
What to ask next
Which data and integration bottlenecks prevent useful systems from becoming part of everyday work?