Field notes / Reviewed 10 October 2026
Three field cases
worth studying.
They show three different ways to change production work: recommend a setting, let people handle exceptions, or give local software a narrow control task.
The useful comparison is what changed in the field, who had permission to act and how the result was measured. The percentages below cover different populations, periods and measures. They are not a ranking of returns.
| Operator | Who acts | What was measured |
|---|---|---|
| BPX Energy | People accept daily recommendations | 6% average incremental oil on adopting wells; adopter count unknown |
| Burk Royalty | Software adjusts and protects; people handle exceptions | Eight months of workload, electricity and production measures |
| Expand Energy | Local software shuts in and actuates the choke | 70–139% cumulative gas gain over 63–83 days |
These cases connect data to a specific operating action. None demonstrates a general-purpose AI running a company, and none isolates a language model’s contribution. Burk provides useful partial cost evidence, but the full cost bridge still matters before comparing returns.
BPX Energy / AI-284
Daily lift recommendations across a 441-well Permian pilot
A useful example of recommendations that account for field constraints. Ask how many recommendations were accepted, what changed in the field and how incremental production was calculated.
What it changes
Xecta combines physics-based well-flow models with AI and historical and daily production data. It recommends gas-lift injection rates, ESP frequencies, lift-system transitions and acid-job candidates, while considering shared compressor and injection-pressure constraints.
Who acts
Operators receive daily recommendations. The paper does not establish autonomous actuation or disclose how many wells adopted the recommendations.
Where and for how long
Permian asset with 441 gas-lift and ESP wells. BPX is bp’s US onshore business; this is a distinct field case, not a count of bp-wide deployment.
What the authors report
Authors report average 6% incremental oil on wells that adopted recommendations, plus more than 20% lower average field gas-lift use. The number of adopting wells is not disclosed.
What the economics tell us
Lower gas-lift use could reduce resource demand, but no dollar saving, implementation cost or net return is disclosed.
What we still need to know
The 6% result applies to wells adopting recommendations, not all 441 wells. The gas-lift figure is a separate resource-use measure. Duration, adopter count, comparison method, costs and net return are not disclosed. The operator/vendor-coauthored abstract is not independent validation. The source is dated 2026; an exact publication or deployment day is not established.
Read the sources
These are operator/vendor accounts. Reviewing a source confirms what its authors disclosed; it does not independently validate the result. Conference sources retain year-only dates where the exact day is unknown.
- Daily lift recommendations across a 441-well Permian pilot
2026 · SWPSC 2026051 abstract: method, 441-well asset and reported results
Burk Royalty / AI-285
Pump-by-exception monitoring and control at a West Texas lease
An unusually concrete small-site account of changes in field workload and operating cost. Its next test is whether the calculation holds after all hardware, software, installation and support costs, and across more leases.
What it changes
Amplified combines controllers, sensors, tank and facility monitoring, cameras, cloud AI/ML and mobile tools. Automated setpoint optimization and protective shutdowns respond to conditions such as parted rods, abnormal pressure and tank overfill. People handle the exceptions.
Who acts
Software adjusts setpoints and can shut equipment down for protection; staff review alerts and respond to exceptions. This is an integrated instrumentation, control and work-process change, not a stand-alone chatbot.
Where and for how long
West Texas satellite lease. The full paper describes five sucker-rod producers and one injector; its HTML abstract says six producers and one injector. The source population conflict remains unresolved.
What the authors report
Over eight months, authors report 43% lower electricity consumption and daily strokes; electricity expense of $660 to $376 per well-month; routine field time of 126 to about 18 hours per month (86% lower); and 26% lower hot-oiling cost. Production stayed around 19 BOPD versus an expected 4.25% annual decline. Authors calculate about $65,000 incremental profit at $65/bbl WTI and capital recovery in slightly over one month.
What the economics tell us
Some operating costs are disclosed: electricity expense falls from $660 to $376 per well-month, alongside hot-oiling and workload measures. Authors calculate ~$65,000 incremental profit over eight months at $65/bbl WTI and payback slightly over one month. The capital amount and full recurring-cost bridge are undisclosed, so this is partial cost evidence, not a validated all-in ROI. Potential avoided spill-cleanup cost is an event estimate, not recurring savings.
What we still need to know
The full paper says five producers plus one injector; the abstract says six producers plus one injector. Routine field time is measured in hours, not a count of interventions. Production aggregation and the counterfactual decline need clarification. Profit and payback are author-calculated, not independently validated; actual capital and an itemized total-cost bridge are absent. Benefits apply to the integrated system; the AI-only contribution is not isolated. Future profit and wider-fleet extrapolations are projections, not achieved results. Exact publication and deployment days are not established.
Read the sources
These are operator/vendor accounts. Reviewing a source confirms what its authors disclosed; it does not independently validate the result. Conference sources retain year-only dates where the exact day is unknown.
- Pump-by-exception monitoring and control at a West Texas lease
2026 · SWPSC 2026052 abstract; linked full paper PDF pp. 3, 7–9 (site configuration, eight-month results, economics and projections) - Field Optimization Reimagined: full paper
2026 · PDF pp. 3, 7–9; figure on p. 8
Expand Energy / AI-286
Autonomous liquid unloading on nine Haynesville gas wells
A concrete example of local control with cloud learning. Ask how the baseline accounts for unloading cycles and decline, what safety permissions apply, and whether gains persist through a full operating year.
What it changes
An edge IIoT gateway uses surface pressure, temperature and flow data with physics-based critical-velocity logic. It decides shut-in timing and actuates the choke locally. Cloud machine learning predicts shut-in duration; local control does not require continuous cloud connectivity or operator oversight.
Who acts
The system makes local shut-in and choke decisions. The sources do not describe a language model or a general-purpose agent; this is specialist production control.
Where and for how long
Nine Haynesville wells across eight pads. The related July 2026 JPT account reports results over 63–83 days; it describes the same deployment, not an independent replication.
What the authors report
SLB’s JPT account reports 70–139% cumulative gas-production gains over 63–83 days on nine wells across eight pads, with daily gains up to 350 Mscf/d. More than 80 MMscf additional gas per well-year is an estimate, not a year of observed production.
What the economics tell us
The reported short-period production gains and estimated annual gas volume are not net financial returns. Full implementation and running costs are not disclosed in the reviewed primary passages.
What we still need to know
Short field-test results do not establish a year-long or fleet-wide outcome. More than 80 MMscf per well-year is explicitly estimated. The conference abstract and SLB-authored JPT account describe the same deployment. The reviewed primary passages do not disclose full hardware, software, integration or ongoing support costs or net ROI. Annual financial figures from an author-uploaded manuscript are held pending publisher-source access. The conference source is dated 2026; its exact publication day is not established.
Read the sources
These are operator/vendor accounts. Reviewing a source confirms what its authors disclosed; it does not independently validate the result. Conference sources retain year-only dates where the exact day is unknown.
- Autonomous liquid unloading on nine Haynesville gas wells
2026 · SWPSC 2026001 abstract; JPT 1 July 2026, Haynesville Field Results - Edge Computing and IIoT Enable Autonomous Artificial Lift and Well Optimization
2026-07-01 · Field Results: Haynesville Basin
Related technology context / EOG
A useful result that we cannot attribute to AI.
EOG’s August deck reports 5% improved Utica base performance from its production-optimizer suite after the Encino acquisition. It does not mention AI or machine learning. EOG’s earlier description of proprietary ML tools does not establish AI causality for this result.
The deck does not define base performance or disclose the well count, period, method, costs or counterfactual. This has been added to EOG’s chronology; it is not an additional counted AI outcome.
Read the August deck · slide 14 · Open EOG’s project history
And failed deployments?
This bounded research pass did not find a new substantiated named abandonment. Historical model misses and inaccessible or anonymous leads remain research leads. That is a limit of the public evidence, not evidence that deployments always work.
Read the research gaps and setbacks · Download the updated register