Two named-company workflow profiles added: CNX current-use disclosure and Aker BP implementation announcement. The anonymous failure is a separate note. Expand is existing evidence; no duplicate project or result is added.
Three targeted Luna research tracks, followed by editor checks of original sources and existing dossiers. This is not an exhaustive review of all earnings calls or private deployments.
CNX Resources / Company describes current use; start date unknown
Safety cameras that flag unsafe conditions
CNX says it uses AI video analytics to identify unsafe conditions, coach employees in real time and inform training. HSE++ also analyzes earlier incidents and safety observations to find trends.
How it works
Video analytics flags unsafe conditions for coaching and process improvement. Separately, HSE++ predictive analytics uses historical incidents and prior safety observations. The source does not name a model, camera provider or implementation stack.
What was measured
No AI-attributed incident reduction, false-positive rate, site or camera count, implementation cost or net return is disclosed.
When
The current company page describes use, but gives no publication date or deployment start. Its 2025 safety milestones do not date the AI workflow.
What remains unknown
Company-reported use, without independent validation or a quantified effect. Do not attribute station safety milestones to AI. Fleet telematics on the same page is not identified as AI. The undated page cannot establish an in-window launch.
What to ask next
Which unsafe conditions trigger useful interventions, how many alerts are rejected, and did incident rates change on comparable sites?
CNX Workplace Health and Safety Undated · Artificial Intelligence section; separate Measuring Performance and Safety Milestones sections
Read AI-288Aker BP / Implementation announced; operating results not established
Agents for risk assessment and alarm reviews
Cognite and ABB name Aker BP as the first customer signed on to implement an agentic offering for risk assessment and alarm rationalization.
How it works
The announced collaboration would add Cognite Industrial AI and Data capabilities to ABB Ability SafetyInsight and AlarmInsight. The release describes agents coordinating work across systems to support operational risk assessment and alarm rationalization. It does not establish live control authority or autonomous safety decisions.
What was measured
No achieved result is disclosed for this offering. Faster workflow claims and hundreds of agents are ambitions. Aker BP production efficiency and future production targets are company context, not AI-attributed outcomes.
When
Published 21 May 2026. Aker BP signed on to implement; completion date, site scope and sustained live use are not established.
What remains unknown
Vendor announcement, not proof of scaled operation. Existing root-cause-analysis benefits in AI-049 must not be transferred to this workflow. AI-131 covers Microsoft Copilot and is a separate initiative. No measured safety benefit, cost or net return is disclosed.
What to ask next
Which risk and alarm reviews are live, what actions can agents take, and where are human approval and safety-system boundaries?
Read AI-289Read AI-049Anonymous operator in Oman / Vendor account of a failed output; anonymous customer
One fracture-model output failed; the workflow was narrowed
EarthScan reports that fracture-direction estimates became unreliable on horizontal wells. The team removed that output from automation and kept human interpretation, while using fracture location and depth as an aid.
How it works
The vendor describes a blind evaluation on five horizontal wells in an Oman carbonate dataset. Its team and the operator’s interpreter agreed acceptance bands separately for depth, dip and azimuth. The source has ambiguous descriptions of training and transfer; it does not support a clean matched model comparison.
What was measured
Reported horizontal azimuth mean absolute error: 65.4 degrees, outside the interpreter’s 5–15 degree tolerance bands. Vertical reference was about 10 degrees. Vertical depth, dip and F1 comparators were not from a matched controlled run. No review hours, cash savings or net return are disclosed.
When
Case published 10 September 2026; the blind test was in July 2023, outside the initial screening window. The publication describes subsequent use as a preprocessing aid, without a dated deployment population.
What remains unknown
Anonymous customer, vendor-authored account, five horizontal test wells and confidential data/code. The test was in July 2023; September 2026 is the publication date. This shows a failed output and narrower automation, not an abandoned program. It is not included in the named-company adoption or numerical-result counts.
What to ask next
Does each output meet the engineer’s tolerance on each well type, and how much interpretation work remains after automation?
Expand Energy / Existing evidence made easier to find; no new project
Forecasting, site data and integration at Expand
The Snowflake case also covers ML well forecasts, natural-language drilling questions, SCADA-linked site data and M&A data integration. These details already appear in AI-195’s research dossier.
How it works
Snowpark ML supports production/type-curve forecasting. Drilling ML and operational alerts use centralized data; Cortex Analyst supports natural-language queries. The case describes site data for 3,700 locations. A full drilling dashboard is under development and broader MLOps maturity is on the roadmap.
What was measured
The vendor estimates more than 1,000 annual labor hours in one business unit. Its headline reports 40% less time on data-related M&A tasks; a customer quote also mentions cost. These are not demonstrated AI-only cash savings, and methodology and observation periods are unspecified.
When
Undated vendor page, already reviewed 6 October 2026. It does not establish a new 2026 launch. Current workflows, the developing dashboard and the platform roadmap have different stages.
What remains unknown
Same source and substantive claims already in AI-195. No additional adoption or numerical-result count. Site data coverage is not proof of automatic control at all 3,700 locations. Keep this platform separate from Leucipa in AI-194.
What to ask next
Which business unit uses the forecasts, how was the labor estimate made, and what incremental benefit comes from ML rather than data consolidation?