Look at what changes in the work. Then ask what the results actually prove.
Initial screen: 26 September 2024–26 September 2026. Broad register review: 8 October 2026. Selected updates through 26 October 2026; individual sources retain their own dates.
Across nearly 1,400 wells, SLB reports overtreatment falling from 17% to 3% over a 20-month program.
Each square = 1% of wells · performance over 20 months
Sensors feed ML detection. Weekly team reviews support remote adjustments.
Better chemical dosing reduced waste across an existing well fleet.
Changing a setting, prioritizing a visit and planning a development are different jobs.
Devon EnergyChanging well conditions
Adjusts lift settings in real time
1,000 live wells · August 2026
The pilot gain is not a result for the later fleet. Full costs, eligible-well population and detailed control limits are not disclosed.
Read the project
OccidentalField sensors and operating data
Helps prioritize lease-operator visits
~40% of onshore production · August 2025
No route-level time or savings result is given. AI’s contribution is not separated from sensors and automation; dispatch and overrides are unspecified.
Read the project
QB Energy / QuantumTerrain and engineering constraints
A solver proposes plans; an assistant explains
12 areas over nine months · sponsor report
Vendor- and sponsor-reported. Faster planning does not establish better capital returns. The solver’s results cannot be attributed to the language model.
Read the projectSelected examples, not a ranking. More autonomy does not establish more value. Open a project for scope, reported results and limits.
Start with these patterns. Then look at the companies and the work behind them.
Devon describes software that changes lift settings. Oxy uses it to help decide where field crews should focus. These examples change what happens next. sources
Ask who can make the change, what the software is allowed to do and how people check the result.
EOGBuilds internal software
EQTShares work and communicationEOG builds its own software. EQT brings important processes and communication into a shared digital workspace. That may make it easier to put new tools into everyday use. sources
An existing foundation could help. It does not prove that a company gets better AI results.
We have 220 profiles of company AI projects. Sources describe 151 as in operation; 55 of those also report a numerical result. The figures cover different measures and periods.
To know whether a system pays off, we need a fair comparison, all its costs and evidence that the benefit lasts.
Our reading of selected public sources, across 104 named companies or groups. These counts describe profiles in this collection, not industry adoption. Some profiles cover parts of the same program. Select an exhibit or a source to inspect the evidence.
These stand out for what they change: equipment settings, fleet oversight, field work or major planning decisions. Ambition doesn’t tell us which earns the best return.
Our selection from the reviewed cases, with no rank order. The examples below have reported operating use. Reach, decision authority and results are different measures.
It moves from recommending a setting to changing it, around the clock.
Devon reports real-time autonomous lift optimization across 1,000 wells in August 2026. sources
The earlier 2–3% pilot uplift is not a measured result for this whole fleet. Full costs are missing.
It changes how engineers oversee a large fleet, rather than adding another isolated model.
ADNOC reports a deployed AI operations center connecting live rig data across its onshore and offshore fleet. sources
This is oversight, not autonomous rig control. The reported effort and downtime benefits lack detailed baselines.
A numerical solver selects pads and wells under engineering constraints. An assistant explains the plans.
Quantum identifies QB Energy as the operator. Microsoft reports a 20,000-acre scenario taking 20 minutes versus three weeks. sources
The scenario comparison is vendor-reported. The roughly $10bn development program informed is not savings; capital approval remains with people.
Vickery Energy / QuantumIt applies language models and document recognition to ownership, heirship and lease availability.
Quantum reports a West Virginia title-research run in about 40 hours, versus an estimated 40 business days. sources
The comparison is an estimate, not a controlled trial. Accuracy, full costs and replacement of legal title opinions are not established.
It changes a daily operating routine: crews focus on exceptions instead of following fixed routes.
Management reported this production share had moved to routeless operations in August 2025, using sensors, automation and AI. sources
That is a share of production, not wells. The source does not isolate AI’s contribution or report route-level savings.
It makes ongoing adjustments to a continuous industrial process, with permanent use after a pilot.
Aramco reports permanent adoption and 10–15% lower amine and steam use after the October 2024–April 2025 pilot. sources
These are company-reported results at one plant. Evaluation details, full costs and repeatability at other plants are not disclosed.
Look up a company or start with your team.
See their projects, reported results and original sources.
Find examples in production, drilling, maintenance, finance and other teams.
What should each leader do with AI? Briefs for the CEO, CFO, COO and 13 other roles—with peer examples and ideas to test next.
Tools like ChatGPT use large language models, or LLMs. They work well with text. But a model built for a well or a piece of equipment may create more operating value. This chart asks where the value could be, and how much of the work a language model could handle.
Reading information, looking into a problem and drafting a response.
Read the full research report for more charts and the detailed eight-company comparison. Compare companies side by side or work through an idea for your own team.
Selling to operators? Research potential customers or prepare a report with sources.
Who is building a company brain? Compare shared knowledge, executive assistants and agents that turn conversations into work.
Could your company repeat these results? Compare operating conditions, costs and remaining questions.
220 profiles cover company AI projects, across 104 named companies or groups. The full collection has 306 profiles, including products, investments and other related work.
We start with earnings materials, investor presentations, company reports and named customer examples. We also cover products, investments, company plans and energy infrastructure serving AI. We keep those separate from companies using AI in their own work.
We searched and reviewed public sources, but we cannot claim to have found every project worldwide. Some profiles overlap. A publication date may be later than the work it describes. We keep each result with its source, period and limitations. If we found no announcement, that does not mean a company is inactive or a project failed.
