Energy AI Observatory · report generated 2026-10-08 · register updated 8 October 2026
AI is doing operating work. The evidence says more about what runs than what it earns.
A collection readout of company-use profiles, measured outcomes, and disclosure limits. The strongest signal is a gap between described operating use and comparable disclosed results—not a measured industry failure rate.
196company-adoption profiles in this public-source register
133profiles classified operating or scaled
54profiles with a counted numerical result
Exhibit 1 · 196 profiles
Where company-use profiles sit
Top six primary analyst-coded departments and the remaining 48 profiles. Bars share a 0–48 profile scale.
IT, data & enterprise transformation36
Production & field operations33
Maintenance & reliability25
Drilling & completions22
HSE & emissions18
Refining, LNG & process operations14
Other48
Show chart data
Department profile counts| Department | Profiles | Denominator |
|---|
| IT, data & enterprise transformation | 36 | 196 |
| Production & field operations | 33 | 196 |
| Maintenance & reliability | 25 | 196 |
| Drilling & completions | 22 | 196 |
| HSE & emissions | 18 | 196 |
| Refining, LNG & process operations | 14 | 196 |
| Other | 48 | 196 |
These are profile counts, not unique companies, spend, adoption rates, or comparative success rates. One primary department is assigned per profile; other functions can also be involved. The full 14-department breakdown is in the CSV.
Exhibit 2 · shared 196-profile denominator
Operating stage and numerical evidence overlap
Operating or scaled
133 / 196
Counted numerical results
54 / 196
Show chart data
Overlapping operating-stage and result counts| Set | Profiles | Denominator | Share |
|---|
| Operating or scaled | 133 | 196 | 67.9% |
| Counted numerical results | 54 | 196 | 27.6% |
| Both | 46 | 196 | 23.5% |
Three separate bars use the same 196-profile denominator. They are overlapping sets, not a funnel. Of the 133 operating/scaled profiles, 87 have no counted result; of 54 counted-result profiles, 8 remain at pilot. Stage describes the disclosure, not independent confirmation of current reach or impact. A missing counted metric does not establish no benefit. “Counted result” spans unlike units and evidence designs.
Exhibit 3 · 54 result profiles
Who reports the counted results?
Attribution shows who reported an outcome, not source reliability or independent assurance.
Operator-attributed3259.3%
Vendor-attributed2138.9%
Independent-attributed11.9%
Show chart data
Result attribution counts| Reported by | Profiles | Denominator | Share |
|---|
| operator | 32 | 54 | 59.3% |
| vendor | 21 | 54 | 38.9% |
| independent | 1 | 54 | 1.9% |
32 operator · 21 vendor · 1 independent. Results span different units, periods, denominators, and maturity.
Implications
What operators and suppliers can take from the evidence
Operators
Tie deployment to an operating decision and a measured baseline.
The register has 133 profiles classified operating or scaled and 54 with counted numerical results; 46 appear in both sets. That pattern supports a disclosure-level takeaway: described use is more common than a comparable result. For a rollout, name the accountable operator, baseline, exposure denominator, override/decision rights, and value measure before scale-up.
Example / counterexample: ExxonMobil/XTO reports closed-loop optimization on 1,300+ Permian wells and about 2% average oil uplift (AI-235); the control phases and cohort details still need care.
Suppliers and service partners
Make the operating handoff part of the product.
Suppliers should fit into existing systems, support the result after launch, and show full costs. The case register includes 21 vendor-attributed outcomes among 54 counted results; attribution identifies who reports a result, not independent assurance.
Example / counterexample: Ecopetrol reports portfolio investment and benefits plus a separate cumulative ROI measure (AI-147), but periods and bases are unmatched. It illustrates the data investors need and the bridge still missing.
Evidence readers
Keep attribution, reach, and assurance separate.
Of 54 result-bearing profiles, 32 are operator-attributed, 21 vendor-attributed, and one independent-attributed. These source labels do not grade the truth of a claim. Pair every result with its source, denominator, period, counterfactual, and review status.
Example / counterexample: Petrobras reports 10,000 tCO₂/year lower emissions from AI industrial-flare control (AI-146), but its investor slide omits plant count and calculation method. Snam/Engineering Group reports strong task-specific extraction scores (AI-208), not independently verified operating economics.
Selected records
Examples that sharpen the interpretation
AI-235 · Reach counterexample · production
ExxonMobil
ExxonMobil / XTO describes closed-loop gas-lift optimization across 1,300+ Permian wells and reports about 2% average oil uplift in its operator-authored SPE abstract.
A material operating example, with counterfactual and control-boundary detail still limited; JPT article is member-exclusive beyond its accessible summary.
- Source date
- 2025-10-01
- Locator
- Accessible abstract and first paragraph; lines 592–620
Open original sourceAI-146 · Reach counterexample · emissions
Petrobras
Petrobras attributes 10,000 tCO2/year lower emissions to real-time AI industrial-flare control.
Issuer presentation does not disclose baseline, plant count, measurement period, or calculation method; it does not name Smart Tocha as the source of the full result.
- Source date
- 2025-11
- Locator
- Slide 158, Accelerate value generation through digital technologies
Open original sourceAI-147 · Economics counterexample · partial bridge
Ecopetrol
Ecopetrol reports US$1.88m AI investment and US$3.50m benefits in 2025, plus an estimated 7% cumulative ROI since 2021.
The annual portfolio investment/benefit and cumulative ROI use different periods and bases. No all-in project cost or cash-conversion bridge is disclosed.
- Source date
- 2026-03-04
- Locator
- 2025 sustainability annex, PDF page 209 (printed page 209); issued 4 March 2026
Open original sourceAI-208 · Independent-attribution example · task test
Snam
Snam / Engineering Group’s technical preprint reports 95% component identification and 90% correct assignment to drawing sections.
This is a task-specific document extraction evaluation, not independent verification of operating savings or field impact; complex diagrams remain a limitation.
- Source date
- 2025-11-05
- Locator
- Sections 3.1, 4 (Tables 1–3) and 5; abstract uses different summary accuracy figures
Open original sourceMethod and limits
Canonical Energy AI Observatory research-data.json; report generated 2026-10-08; register last updated 8 October 2026. The collection is a consolidated public-source screen, not an exhaustive global census. Original screened cohort: 26 September 2024–26 September 2026. Accepted follow-ups: 5 and 6 October 2026. Classification and research review: 8 October 2026. Each profile is classified by record type, one primary department, deployment stage, outcome count, attribution, and source evidence. Outcome_counted excludes overlapping Aramco/Equinor portfolio aggregates. Provenance grades describe source provenance, not independent validation.
No profile is coded as disclosing a complete project cash-cost / collected-cash reconciliation. Ecopetrol reports US$1.88m AI investment and US$3.50m portfolio benefits for 2025 and a separate 7% estimated cumulative ROI since 2021; its periods and basis do not form a matched project-level net cash bridge. This is an evidence gap, not evidence of uneconomic AI.
Opportunity crosswalks and modeled industry dollar pools are not used as actuals in these exhibits. Download the 196-row record file or method and computed counts.