What is in use, what is still being built, and what did not help.
One new Aramco workflow; four existing profiles clarified, including supplier-reported Exxon results already in the history; one digital-planning note and two technical publications. No named failed deployment was verified in this pass.
Aramco / Company describes use; additional ML work is underway
Connecting rock samples to a geological knowledge platform
Aramco describes AI-supported core analysis and a shared geological knowledge platform. Samples, images, lab results, interpretations and production outcomes are connected for geoscientists.
How it works
Digitized rock imaging feeds AI pattern recognition and geological knowledge graphs. Teams use the linked information to assess reservoir quality and development opportunities. The same article says additional ML work on integrated datasets is underway.
What was reported
No AI-specific accuracy, turnaround improvement, cost or financial result is disclosed.
When
Published 25 August 2026. The center’s 150+ employees and archive of more than one million feet of samples are facility context, not AI deployment or validation denominators.
What remains unknown
Issuer account with mixed technology maturity. Routine lab turnaround is not an AI-attributed result. No company-wide communications brain, executive oracle, autonomous field control or independently validated benefit is established.
What to ask next
Which geological judgments improve when teams can connect the sample, its analysis and the well’s later performance?
Autonomous drilling in Guyana, with supplier-reported results
SLB’s named ExxonMobil case describes Neuro directional control and DrillOps surface-equipment automation, with shore teams monitoring the work.
How it works
Neuro updates trajectory and requests downlink commands. DrillOps controls pumps, drawworks and top drive, adjusts drilling parameters and responds to drilling dysfunction. Operations are monitored from an onshore center.
What was reported
SLB reports more than 48 km of complex well paths, more than 93% of operations executed autonomously, 51% fewer controlled-ROP downlinks and 75% fewer off-bottom downlinks since implementation.
When
Undated case in the supplier’s 2026 URL directory. Exact start, observation period, well count and definition of the operations denominator are not supplied.
What remains unknown
Supplier-reported operational measures, not independently audited performance or cash ROI. This substantiates the Neuro/DrillOps case, not the separate pending March 2026 first-well claim or its different vendor attribution. Existing source history is retained.
What to ask next
How much well-delivery time changed after automation, and how often did shore teams intervene?
Read AI-072Ecopetrol / CADE described in use; Rubiales dispatch-model MVP under construction in 2025
An operating energy desk, with a dispatch model still being built
Ecopetrol’s annual report describes CADE consolidating live operating information and supporting electricity-supply decisions with advanced analytics and generative AI.
How it works
A Rubiales dispatch-model proof of concept was developed in the first half of 2025. Work on an MVP began in the second half, using electrical signals to generate operating setpoints. That model’s construction is separate from the broader CADE platform already described in use.
What was reported
No measured energy, emissions, reliability, cost or production benefit is disclosed for these AI components.
When
2025 operating and development activity, described in the 2026 annual-report edition. Exact publication day and MVP commissioning date are not established.
What remains unknown
The platform’s operating status does not establish live use of the Rubiales MVP. The name “autonomous” does not establish AI control authority. Model, approval and override details are not given.
What to ask next
Are Rubiales recommendations now used, who approves a setpoint, and how is the result compared with the previous dispatch method?
Read AI-148PETRONAS / Earlier exploration tools deployed; new surface-equipment work remains development
Deployed exploration tools and a separate equipment-development agreement
PETRONAS says AI.SEEK and Global Exploration Basin Screening & Analysis (GEB) have been deployed. The third TriCipta agreement develops a different set of surface-equipment tools.
How it works
The earlier tools unify geoscience information for geospatial analysis. IBM Malaysia and Tridiagonal.AI are partners in the new surface-equipment optimization work, aimed at production and maintenance decisions.
What was reported
No measured exploration, production, maintenance, time or financial result is supplied. Deployment is a stage disclosure, not a quantified benefit.
When
Company release dated 8 July 2026. Earlier exploration tools came from 2025 agreements; the third agreement was newly signed.
What remains unknown
This is already mentioned in AI-143; the update makes the split clearer. Do not transfer the earlier tools’ deployment status to the new equipment workflow or count another project from the same announcement. Asset, user and model-validation populations are undisclosed.
What to ask next
What do teams use AI.SEEK and GEB to decide, and which surface-equipment tools have since reached live use?
Read AI-143Chord Energy / Existing scaled AI workflow; broader operating results kept separate
Rod-lift AI sits inside a broader artificial-lift improvement program
Chord reports ML on about 99% of rod-lift wells and an early roughly 25% improvement in rod-pump run times. Those figures were already in the register.
How it works
The release places rod-lift ML alongside a wider artificial-lift optimization effort. It does not describe the model, vendor or approval/control interface.
What was reported
The broader program reports over 50% improvement in ESP replacement cycle times, over 25% improvement in failure rates since early 2025, and about 1,200 fewer workover rig days in 2025. These are not separately isolated ML effects.
When
2025 results, disclosed 25 February 2026. ESP comparisons start at the beginning of 2025; an exact matched AI evaluation period is not supplied.
What remains unknown
Keep the rod-lift result separate from ESP and workover outcomes. Do not attribute the whole continuous-improvement program or its free-cash-flow benefit to AI. No additional result profile or cash ROI is counted.
What to ask next
What part of fewer failures and workovers comes from ML rather than equipment changes, maintenance and field practices?
Read AI-097TotalEnergies / New digital-planning agreement; no AI deployment established
A 15-year agreement for a shared well-planning environment
SLB announced a 15-year DrillPlan agreement for TotalEnergies’s global upstream portfolio. The work connects subsurface information with well design and engineering.
How it works
Offset-well data, trajectories, casing, completion inputs and planning assumptions are brought into a common environment while retaining customer workflows and internal digital expertise.
What was reported
No rollout population, achieved operating benefit, contract value or return is disclosed.
When
Agreement announced 8 October 2026; implementation and completion dates are not given.
What remains unknown
The release does not identify an AI component. It belongs in operating-technology context, not the AI-adoption count. A 15-year term does not prove the whole portfolio is already live.
What to ask next
Does a common planning environment make it easier to compare alternatives and preserve the assumptions behind a well decision?
Petroleum Research / Comparative study; not a named deployment failure
A more complicated model did not improve these forecasts
In a comparison of four production datasets and three forecast horizons, kernel-based models outperformed neural architectures. More complexity did not deliver an accuracy advantage in these tests.
How it works
The study compares ML proxies for field-scale production under operating scenarios. Its abstract also reports that spatial data coverage appeared more important for generalization than dataset size.
What was reported
No company-wide deployment result, cancellation rate or project cash return is established.
When
Published online 21 January 2026. We reviewed the publisher’s indexed abstract; direct full-article access was unavailable.
What remains unknown
The finding applies to the tested inputs, datasets and horizons. It does not establish that deep learning is ineffective across petroleum tasks. This is technical benchmarking, not a failed operator rollout.
What to ask next
Does the proposed model outperform a simpler baseline on the operating decisions and conditions that matter?
Energy and AI / Literature review; not a field incident or failure-rate study
A good prediction is only part of a reliable control system
A surface-network review finds gaps in uncertainty, governance, transfer across fields and evidence for safe closed-loop use beyond pilots. Sensor drift and rare operating conditions can undermine an otherwise accurate model.
How it works
The authors reviewed prediction, monitoring and control literature from January 2015 through March 2026. They argue for physically defensible recommendations, oversight and protected actuation.
What was reported
This review does not measure incident frequency, failure prevalence or the performance of a named operator’s deployed system.
When
September 2026 journal edition; literature search through March 2026. Relevant publisher text was available through the web index.
What remains unknown
Literature synthesis is not an independent field audit. Its gaps do not show that a particular company’s controls are unsafe or that every deployment remains a pilot.
What to ask next
What detects drift, handles unfamiliar conditions and prevents a recommendation from becoming an unsafe action?