Each square = 1% of wells · performance over 20 months
Share of wells overtreated
$4.4m SLB-reported net savings
First nine months · net-cost calculation undisclosed
- Where
- Nearly 1,400 producing Permian wells
- When
- 20 months; exact dates unknown
- What changed
- Sensors · ML detection · team review · remote adjustments
Read the sources and limits
SLB-reported $4.4m net savings covers the first nine months. Net-cost perimeter unknown; no complete ROI. Not an 82% reduction in chemical volume or total cost.
Open the full research noteReported improvement
Perenco
Perenco reports faster drilling across three wells
19% higher drilling speed
- Where
- Offshore Republic of the Congo; individual assets unnamed.
- When
- 2025 · three-well campaign
- What changed
- 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.
Read the sources and limits
Raw drilling-speed baseline, matched offset population, complete costs and AI-only contribution unknown.
Open the full research noteReported improvement
Wafi Energy Pakistan
Invoice handling fell from about five minutes to under one
~5 min to <1 min in automated scenarios
- Where
- Automated invoice scenarios at Wafi Energy Pakistan; their share of all invoices is unknown.
- When
- Calendar dates unknown
- What changed
- AI document capture, rules and SAP workflows handle routine invoices; finance staff handle exceptions. Autonomous payment approval is not established.
Read the sources and limits
Invoice volume, automated-scenario share, precise measurement window and full costs unknown. This is the Pakistan Shell-brand licensee, not Shell-wide adoption.
Open the full research note65% less chemical consumption
- Where
- Unnamed facility, southeastern New Mexico.
- When
- 30-day observation · dates unknown
- What changed
- A flow-meter-linked controller, new chemistry and a relocated injection point changed treatment. Key operator personnel received online write access.
Read the sources and limits
Population, throughput, cost calculation, installation costs and longer-term durability unknown. New chemistry and equipment prevent attributing the result to software alone.
Open the full research note65% less chemical use per well
- Where
- Unnamed field trials; geography unspecified.
- When
- 2025 paper · trial dates unknown
- What changed
- Simulation-guided placement, a capillary string and an injection valve deliver chemicals to the treatment point. No AI is identified.
Read the sources and limits
Operator, geography, well population, raw baseline, cost perimeter and installation costs unknown. This does not establish all-in ROI or scaled deployment.
Open the full research noteReported improvement
Burk Royalty
Lower electricity expense on an existing lease
Electricity expense per well-month
- Where
- Small mature lease; not Burk’s entire fleet
- When
- Eight months · 2026 paper; exact publication day unknown
- What changed
- Sensors, exception-based field work and control changes.
Read the sources and limits
Author-reported; combined changes, electricity prices and full-cost reconciliation matter. Not an AI-only or fleet-wide return.
Open the full research noteUse case stopped
TotalEnergies
The model predicted drum level; it did not answer the switching decision
75% prediction accuracy; use case not retained
- Stage
- Advisory pilot / one use case reportedly dropped
- Where
- One delayed coker at Port Arthur, Texas; accuracy denominator unknown
- When
- Drum work began July 2025; named remarks reported 11 June 2026
- What changed
- Advisory drum-level predictions left important switching conditions outside the model. Other pilot uses continued.
Read the sources and limits
Sponsored conference report quoting a named site engineer. No accuracy definition, test count or quantified loss. This is one discontinued use case, not an abandoned AI program.
Open the full research noteFive usable live verification points in four weeks
- Stage
- Historical field trial
- Where
- Eight-well West Brae field; not a current deployment announcement
- When
- Undated historical trial; 12 usable retrospective points over two years
- What changed
- ML recommended valve movements. Stable operations and larger unrelated changes made small gains difficult to establish.
Read the sources and limits
NZTC judged additional value limited. The 1% potential and£26.2 mindustry scenario are not realized returns. No cancellation or negative full-cost ROI established.
Open the full research note50% sensitivity; nine missed failures
- Stage
- Retrospective test / later deployment reported
- Where
- 137 pumps overall; 83 failed pumps trained. Remaining failed-test count inferred as 18.
- When
- Conference 29–30 October 2024; underlying records historical
- What changed
- Ensemble predictions use ML, reliability models and engineering rules. Engineers choose interventions; the paper separately reports subsequent daily deployment.
Read the sources and limits
Nine misses and 50% sensitivity imply 18 failed test cases; this denominator is inferred. Five misses involved short circuits. Event definitions depend on maintenance actions. Not a live failure rate or abandonment.
Open the full research noteTechnical limit
UNEP IMEO / MARS
Fewer false alarms came with more missed plumes
Single model3,565 false alarms
Five-model ensemble1,501 false alarms
- Stage
- Model test / operational system after redesign
- Where
- 200 EMIT test granules containing 311 labeled plume events; multiple emissions sectors
- When
- Preprint v2: 24 March 2026; evaluation dates unknown
- What changed
- Five-model ensemble lowered false alarms, but detected 207 events versus 213 and missed 104 versus 98. People verify candidate plumes.
Read the sources and limits
Single-model result is a five-run mean; ensemble is one combined prediction. Test counts, not live alert rates or labor savings. A successful redesign, not a cancelled operator project.
Open the full research noteReported false-positive rate
- Stage
- Provider-reported technical evaluation
- Where
- more than 2,000 attorney-validated answers; more than 200,000 documents in the platform universe
- When
- Publication, operating and evaluation dates unknown.
- What changed
- Regression tests catch version changes before release; rules and structured legal reasoning support title determination.
Read the sources and limits
Anonymous client, undated provider case. No independently confirmed live error rate, realized labor saving, full cost or ROI. Reduced false positives does not eliminate human title review. Exact evaluated-decision denominator unknown; document volume is not that denominator.
Open the full research noteTechnical limit
Snam / Engineering Group
Plant digitization was less reliable on complex systems
42.4% relationship recall at 100 predictions
- Stage
- Research evaluation
- Where
- Gas-plant documents; approximately 1,500 ReMi in supplied collection, not verified live deployments
- When
- Preprint 5 November 2025; evaluation/operating dates unknown
- What changed
- AI reconstructs information and diagram relationships. Authors identify weaker performance on complex plants.
Read the sources and limits
Authors provide technical tests, not realized time savings or full-cost ROI. Do not merge abstract summary accuracy with task-specific tables or label this a cancelled rollout. The 95% component and 90% assignment measures concern separate final tasks.
Open the full research noteTechnical limit
SLB
Overall accuracy concealed poorer coverage of rare drilling risks
89% overall accuracy; 44% equal-class recall
- Stage
- Patent-application technical example
- Where
- One GPT-2 comparator in Table 15; exact test population not established
- When
- Patent application published 5 March 2026; evaluation dates unknown
- What changed
- More customized models perform better in the filing; the warning is to inspect rare-event metrics, not only overall accuracy.
Read the sources and limits
Applicant-authored technical examples, not an independently validated field outcome. Other examples use differing cohorts and later model configurations. No measured reduction in drilling incidents or complete ROI.
Open the full research noteReported improvement
Aramco
Less amine, steam and power in acid-gas removal
10–15% less amine and steam use
- Stage
- Pilot followed by reported permanent adoption
- Where
- Fadhili Gas Plant, Saudi Arabia; one plant
- When
- Pilot October 2024–April 2025; permanent adoption reported, exact date unknown
- What changed
- Yokogawa reinforcement-learning control adjusts acid-gas removal. Aramco reports fewer manual interventions.
Read the sources and limits
Operator-reported; absolute baseline, control group, full costs and independent validation unknown. Around 5% lesspower is a separate input measure; wider-site deployment was being assessed.
Open the full research note