The VisionAI Strategy

AI Is Now a Line Item on the Earnings Call

Three operators have published AI outcomes specific enough to examine (850-plus wells on fully autonomous artificial lift at Devon, 80-90% off select maintenance and inspection tasks in ConocoPhillips Norway, the ApEX platform at Chevron). The structural divide, the four use cases that travel from supermajor to independent, and the path that runs in weeks instead of years.

Michael Atkin, P.EngPublished 22 May 2026Updated 14 October 202612 min read

850+
Devon Energy wells on fully autonomous artificial lift optimization (Q1 2026 earnings call)
15-30%
Devon Energy productivity gain for engineers, software developers and geoscientists on daily tasks (AOGR, July 2025)
80-90%
ConocoPhillips Norway BU reduction on select maintenance and inspection tasks, Global Digital Twin program (JPT, 2023)
at least 30%
Gartner forecast: generative AI projects abandoned after POC by end of 2025, across all industries
~70%
McKinsey: oil and gas companies still stuck in the pilot phase
<15%
BCG: oil and gas respondents who say digital has meaningfully created value
Figures cited in this article. Source and basis as labelled on each.

850-plus wells on fully autonomous artificial lift at Devon. 80 to 90 percent off select maintenance and inspection tasks in ConocoPhillips' Norway business unit. A named generative-AI platform in production at Chevron. AI is no longer only a slide in the strategic plan. In a handful of cases it is a number the CFO is being asked about on the quarterly call. The question for every other operator is whether the bar gets cleared on the operator's own stack.


The Bar Got Quantified in Public

For most of the last decade, "AI in oil and gas" was a tradeshow conversation. The proof points sat behind NDAs. The internal pilots were rolled into broader digital programs and never disclosed at the line-item level. The CFO had nothing to defend, because nothing was being measured at a level the investor could ask about.

That began to change in the 2024 through 2026 reporting cycle, though the disclosures are narrower than the industry conversation suggests. Three operators have published AI outcomes specific enough to examine. Each is sourced below to the operator's own disclosure, and each is quoted at the scope the disclosure actually covers.

Devon Energy. Over 850 wells on fully autonomous artificial lift optimization, with the 2025 pilot reported at about a 2 to 3 percent uplift and the full rollout not yet quantified by the company. Source: Devon's Q1 2026 earnings call, May 2026. Separately, Devon reports a 15 to 30 percent productivity gain for engineers, software developers and geoscientists on daily tasks, and describes ChatDVN as a firewalled internal tool that is a standard part of the daily workflow. Source: The American Oil & Gas Reporter, July 2025, and the Q1 2026 call.

ConocoPhillips. Global Digital Twin program rolled out across business units, with the Norway business unit achieving between 80 and 90 percent reductions for select maintenance and inspection tasks. The mechanism is a digital twin paired with a handheld device: an operator types in a tag number and gets walking directions to the equipment. Source: JPT, October 2023.

Chevron. ApEX, an in-house generative AI platform launched in August 2024 and first used in deepwater exploration in the Gulf of America, runs multiple search agents across more than a million files to streamline subsurface exploration. Chevron attaches no production or percentage figure to ApEX. Source: Chevron's newsroom.

The pattern across the three is the same. Each operator is running a vertical AI deployment against the SCADA, EAM, production accounting, and historian data the operator already owned. None of them paused for a multi-year data lake. None of them paid for a horizontal copilot license and waited for it to figure out the operating problem on its own. The work was vertical, the data was the operator's own.

What the Other 70 Percent Are Doing

Over the same window these disclosures rolled into the earnings transcripts, three of the most-quoted research firms in oil and gas boardrooms published the counter-statistic. The bar exists. Most operators are not clearing it. The structural reasons keep showing up across every industry conversation.

Gartner. At least 30 percent of generative AI projects will be abandoned after proof of concept by the end of 2025, on a forecast published across all industries rather than oil and gas specifically. The agentic-AI version is worse: 40 percent of agentic-AI projects will be canceled by 2027. The model is not the problem. The data the model was asked to learn from is. Source: Gartner press releases July 2024 and June 2025.

McKinsey. Roughly 70 percent of oil and gas companies have not moved past the pilot phase. Only 17 percent of firms realize more than 75 percent of expected savings from digital programs. Source: McKinsey's industry insights on oil and gas digital transformation.

BCG. Fewer than 15 percent of oil and gas respondents say digital has meaningfully created value. The dashboard nobody opens is the most expensive code in the operator's stack. Source: BCG, "Creating Value Through Digital in Oil and Gas."

The generic-LLM trap. ChatGPT, Microsoft Copilot, and the other horizontal copilots cannot see the operator's SCADA, EAM, AFE history, reserves forecast, or historian. Asking a horizontal copilot to optimize the operation is asking it to guess. There is no published US operator failure case that names this as the proximate cause, because the failures get rolled into "the pilot was inconclusive" and shelved. The structural problem is consistent: a model that is not connected to the operator's data cannot produce field-grade decisions.

Three of those four anti-patterns are research-firm statistics. The fourth is a structural truth that every junior engineer who has pasted production data into ChatGPT will recognize. The operators clearing the bar in the section above are clearing it because they avoided all four.

What the Independent Sees on the Other Side of the Disclosure

Every operator below the supermajor tier is being asked the same question by the board right now: what is the operator's number against the AI bar? The honest answer for most independents is some version of "we are running a pilot." That answer was acceptable in 2023. It is not acceptable in 2026, because the operators that have published have moved the bar from "is AI ready for our operation" to "on what asset population, against what baseline." The independent that walks into a board meeting in late 2026 with a pilot-phase update gets benchmarked against a disclosed well count, fairly or not.

Here is the part that should worry the independent and excite it at the same time. AI has barely reached the trades that run oil and gas. The capability is broad, the highest of any occupation in computer and math work, but observed AI usage in engineering and production work sits near the bottom of the curve. The capability is here; the adoption is not. That means the names above are early, not late, and the gap between them and everyone else is still widening. The supermajors crossed it by paying for bespoke internal builds; the independent that waits is letting the gap widen against it without the budget to close it later. The full third-party-cited picture of where operations sits on AI adoption shows how much of the field is still untouched.

The path forward is not to copy the supermajor playbook. The supermajors paid six to eight years of tuition to learn what worked. They built data lakes that they no longer use as the primary AI substrate. They hired chief data officers who manage organizations that are too large for an independent to fund. The lesson the independent should extract is not the architecture. It is the result. The published proof points (ExxonMobil's automated gas-lift optimization on 1,300-plus Permian wells, Devon's fully autonomous artificial lift on 850-plus wells) both ran against the SCADA history the operator already owned. The independent that has the same SCADA history can run the same class of decision against it, without paying the supermajor tuition. The detailed argument lives in The Data Lake Is a 2017 Idea.

Four Use Cases That Travel from the Supermajor to the Independent

The operator disclosures above all reduce to a small set of use cases that operate at any scale, against the data the operator already has. Each one ties back to a WorkSync product line. Each one has either a public supermajor proof point or a deployed reference at a top 25 private producer running on 5,000+ wells.

Pump-by-priority. Predictive maintenance plus autonomous artificial lift, scored by economic impact rather than fixed-interval routing. Devon's autonomous-lift program on 850-plus wells is the public proof. The same loop runs at the top 25 private producer reference deployment of WellOPS, scoring every active well overnight by revenue variance and risk, publishing the ranked list before 6 AM. The four-week pilot version of this is laid out in The 4-Week Pump-by-Priority Pilot.

Variance commentary that writes itself. The agent reads the historian, the production database, and the ledger, drafts the weekly variance commentary, and surfaces the deviation worth investigating. Devon's ChatDVN 3.0 is the public proof of the daily-use case. The independent version is the WorkSync DataHub plus the variance-commentary agent, running on the existing production accounting feed.

Engineering-hour compression. Hydraulic and facilities modeling from a multi-hundred-hour study cycle to a sub-hour iteration. FlowSync runs the engineer's own model against the calibrated network, drafts the MOC language, and cites every change back to the source document. The detailed gas utility case is in The Million-Dollar Model.

Closed-loop work management. Voice-first field capture into the system of record, every observation tagged to the well, the asset, and the cash-flow impact. The pumper's 90 minutes of post-route data entry compresses to zero, and the next day's plan reflects yesterday's actual conditions. The structural argument is in Exception-Based Surveillance.

Every one of those four use cases runs against the operator's existing data. None of them requires a multi-year data lake. None of them is a horizontal-copilot deployment. Each one is the operator's own SCADA, EAM, production accounting, and well file, surfaced to the lease operator, the dispatcher, the planning engineer, and the operations manager, in the form each one already uses.

What Clearing the Bar Looks Like on a Small-to-Mid Independent

The operator deployments above ran on asset bases of 850 to 1,300-plus wells. The deployed reference at the top 25 private producer on WorkSync runs across 5,000+ wells in the Western Anadarko, Permian, and Wyoming basins. The minimum viable footprint is smaller than that.

An operator with a single SCADA system, a production accounting feed, and an EAM tool has enough operating data to clear the bar on at least one of the four use cases above. The constraint at small scale is not data volume. It is the maturity of the field workflow that consumes the ranked decisions. The DataHub integration time on a typical independent stack is under one week, in read-only mode, with no rip-and-replace. The first ranked daily plan publishes inside 30 days. The first closed-loop deployment runs inside 90.

The Impact Guarantee is in writing. Pick the metric in week zero (production uplift, deferment reduction, route-time recovery, study turnaround, whichever metric the CFO will sign for). Run the loop for four weeks. If the metric moves, the operator signs the annual subscription. If it does not, the operator walks away. No license fee. No kill fee. The pilot is planned with the WorkSync implementation team.

The supermajors paid six to eight years of tuition to learn what works. The vertical-AI substrate that produced their disclosures is accessible to the 500-well independent at a fraction of the supermajor cost, without paying tuition the supermajors already paid. The question is whether the operator answers the board's AI question with a pilot update in late 2026, or with a booked outcome.

Five names cleared the bar. The math says most operators are still stuck behind it. The bar is not theoretical, the proof points are not vendor claims, and the path is no longer 18 months.

You are one week away. Start the 4-week pilot.

Frequently asked

Questions this article answers

What does "AI on the earnings call" actually mean?
It means the AI program is disclosed at a level specific enough for an investor to ask about on the quarterly call: a named asset population, a named mechanism, and where available a measured change against a stated baseline. Devon Energy, for instance, disclosed over 850 wells on fully autonomous artificial lift optimization on its Q1 2026 call, with the 2025 pilot at about a 2 to 3 percent uplift and the full rollout not yet quantified. The threshold is no longer "are we doing AI." It is "on what asset population, against what baseline."
Which operators are publicly disclosing AI outcomes in oil and gas?
Three operators have disclosures specific enough to examine. Devon Energy (ChatDVN as a firewalled internal tool in daily workflow, 15-30% productivity for engineers, software developers and geoscientists per AOGR July 2025, and over 850 wells on fully autonomous artificial lift per the Q1 2026 call). ConocoPhillips (Global Digital Twin program, Norway BU 80-90% reduction on select maintenance and inspection tasks, JPT 2023). Chevron (the ApEX generative-AI platform launched August 2024 for Gulf of America subsurface exploration, with no production figure attached to it). Each is sourced to the operator's own disclosure or to the trade publication that reported it.
Why are 30 percent of generative AI projects abandoned after POC?
The Gartner number is structural. The model is not the problem. The data the model was asked to learn from is. Most POCs are run against a horizontal copilot that has no grounding in the operator SCADA, EAM, AFE history, reserves forecast, or historian. The output cannot be cited back to a source document and cannot score a field decision against cash-flow impact. The operators that have published outcomes (Devon, ConocoPhillips, Chevron) all ran vertical AI deployments against their own data, not horizontal copilots.
Can a 500-well independent replicate the supermajor pattern?
Yes. The published proof points ran against SCADA history the operator already owned. ExxonMobil on 1,300-plus Permian wells, Devon autonomous lift on 850-plus wells. The constraint at small scale is not data volume, it is the maturity of the field workflow that consumes the ranked decisions. An operator with a single SCADA system, a production accounting feed, and an EAM tool has enough operating data to clear the bar on at least one of the four use cases (pump-by-priority, variance commentary, engineering-hour compression, closed-loop work management).
What is the difference between a horizontal copilot and a vertical AI for oil and gas?
A horizontal copilot (Microsoft Copilot, Google Antigravity, Anthropic Claude Code, Cursor) is a general productivity tool. It has no grounding in the operator SCADA, well file, procedures, or basin. It cannot cite the source document for an answer. It cannot score a pumper visit by cash-flow impact. A vertical AI is domain-bound and pre-integrated to the operator data. It reads production accounting, SCADA, EAM, GIS, HSE, and engineering drawings in their native homes, in read-only mode, and produces field-grade output the lease operator, dispatcher, planning engineer, and operations manager already use. The supermajors run vertical models. The horizontal copilot is, at best, a back-office tool for meeting summaries.
How fast does WorkSync stand up an AI pilot on a typical independent?
Initial DataHub integration is typically under one week, read-only on the operator existing systems. Ranked daily plans go live within 30 days. Full closed-loop deployment with optimized routing, exception-based dispatch, and nightly retraining completes within 90 days. The Impact Guarantee is in writing: pick the metric in week zero (production uplift, deferment reduction, route-time recovery, study turnaround, whichever metric the CFO will sign for), run the loop for four weeks, sign the annual only if the metric moves. No license fee and no kill fee if it does not. Plan the pilot with our implementation team.
What are the four AI use cases that travel from the supermajor to the independent?
Pump-by-priority (predictive maintenance plus autonomous artificial lift, scored by economic impact, Devon proof on 850-plus wells). Variance commentary that writes itself (agent reads the historian, production database, and ledger, drafts the weekly commentary, Devon ChatDVN proof of daily use). Engineering-hour compression (hydraulic and facilities modeling from a multi-hundred-hour study cycle to sub-hour iteration, WorkSync FlowSync proof on a gas utility deployment). Closed-loop work management (voice-first field capture into the system of record, every observation tagged to the well and cash-flow impact, 90 minutes per route reclaimed across our customer base).
Why is the data lake not on the critical path for AI in oil and gas anymore?
Three things broke the lake economics. Inference cost collapsed (Stanford 2025 AI Index, GPT-3.5-class call roughly 280x lower than launch price). Vertical AI vendors emerged with pre-integrated workflows that compete on the operator data and the domain integration, not on a horizontal model. Modern integration patterns (event streams, read-only API gateways, ABAC layers) made it cheaper to connect to twelve systems in their native homes than to extract them into a thirteenth. The detailed argument is in The Data Lake Is a 2017 Idea.

Next step

4-week Impact Guarantee on a metric your CFO will sign for.

Tell us your system stack and your biggest operational challenge. We’ll show you how WorkSync delivers a prioritized daily plan in your environment, with your data.

Four weeks to stand up under the Impact Guarantee: license fees only when the metrics move. We charge when your number moves.

Related insights

Keep reading