I've spent 20 years in petroleum engineering and the last eight building AI tools for this industry. In that time, I've watched at least three "AI will transform oil and gas" hype cycles come and go. Most produced PowerPoint slides, not production gains.
2026 is different. Not because the hype has stopped — it hasn't — but because the deployments have become specific enough to verify. Named operators, named fields, named results. That's what this article covers: where agentic AI is actually running in upstream operations right now, what results it's producing, and what it means if you're a mid-size operator watching from the sidelines.
No vendor pitches. No "imagine a future where..." Just what's deployed, what's working, and what's still unproven.
1. Drilling & Completions: The Most Mature AI Deployments
Drilling was always going to be the first domain where AI agents proved themselves. The feedback loops are tight (minutes to hours, not months), the data streams are rich (surface sensors, MWD/LWD, mud logging), and the cost of a bad decision is immediate and measurable. That combination is exactly what machine learning needs to demonstrate value quickly.
SLB Tela: The First Commercial Agentic Platform
In November 2025, SLB launched what they're calling the industry's first commercial agentic AI platform for upstream: Tela. This isn't a dashboard or a recommendation engine. It's a system of AI agents designed to execute multi-step workflows across drilling, production, and subsurface domains — pulling data, running analyses, and triggering actions with minimal human prompting.
Whether Tela lives up to SLB's positioning remains to be seen at scale. But the architectural choice matters: they're betting on autonomous agents, not copilots. That's a meaningful signal from the largest oilfield services company in the world.
ConocoPhillips: Hard Numbers from Eagle Ford
The deployment I find most credible is ConocoPhillips' machine learning program in the Eagle Ford. The results are specific: 65% fewer motor failures and a 20% improvement in rate of penetration. These aren't pilot numbers from a controlled test — they're field-scale results from one of the most heavily drilled unconventional plays in North America.
A 20% ROP improvement translates directly to fewer drilling days, which translates directly to dollars. On a $10M well, that's meaningful. Across a multi-hundred-well program, it's transformative.
Nabors + Corva: RigCLOUD as the Operating System
In April 2025, Nabors expanded its partnership with Corva to bring AI drilling analytics into the RigCLOUD ecosystem. The play here isn't a single algorithm — it's platform positioning. RigCLOUD is increasingly becoming the operating system layer that third-party AI applications plug into.
This matters because it solves the integration problem that killed earlier AI drilling efforts. You can have the best stuck-pipe prediction model in the world, but if it can't ingest real-time data from the rig's WITS feed and surface its recommendation where the driller actually looks, it's academic.
2. Production & Operations: Where Agents Get Harder
Production optimization is a tougher environment for AI agents than drilling. The feedback loops are longer (weeks to months), the data is messier (SCADA gaps, sensor drift, commingled production), and the decision space is wider. But there are real deployments here too.
Aker BP + Cognite + NVIDIA: Offshore Anomaly Detection
In March 2026, Aker BP announced a deployment using Cognite's data platform combined with NVIDIA's NV-Tesseract framework for anomaly detection across its offshore Norwegian assets. The system is designed to catch equipment failures and process deviations before they cause unplanned shutdowns.
Offshore anomaly detection is a high-value target. A single day of unplanned downtime on a Norwegian Sea platform can cost $1-5M in deferred production. If an AI agent catches a compressor bearing degradation 48 hours before failure, the ROI case is straightforward.
Baker Hughes: GenAI for Artificial Lift
Baker Hughes has been rolling out "Flow," a GenAI assistant purpose-built for artificial lift field personnel. In June 2025, they partnered with Repsol to deploy a GenAI virtual assistant specifically for plunger lift optimization.
This is an interesting use case because plunger lift is one of those domains where tribal knowledge dominates. The best plunger lift technicians carry decades of pattern recognition in their heads — which wells respond to cycle time changes, which need velocity adjustments, which ones you leave alone. Encoding that into a GenAI assistant that a less experienced technician can query in the field is a practical application of the technology.
The question is whether the system handles edge cases well enough to earn field trust. Anyone who's worked artificial lift knows the difference between a good recommendation for a textbook well and a good recommendation for the weird one that's been slugging intermittently since the last workover.
3. Data Infrastructure: The Unsexy Layer That Makes Everything Work
Here's what most AI-in-oil-and-gas articles skip: the data plumbing. Every deployment I've listed above depends on clean, contextualized, accessible data. The companies investing most aggressively in that layer are the ones whose AI agents will actually work at scale.
TotalEnergies + Cognite: Full-Asset Data Unification
In September 2025, TotalEnergies signed a three-year deal with Cognite to deploy the Cognite Data Fusion platform across all upstream assets. Not a pilot. Not a regional trial. All assets.
This is arguably the most important AI-related investment on this list, even though it's not an "AI deployment" per se. It's the data foundation that makes every downstream AI application possible. TotalEnergies is betting that unified OT/IT data across every well, every platform, and every pipeline will be the competitive advantage — and the AI applications are the second-order effect.
Cognite + Snowflake: Bridging OT and IT
In January 2026, Cognite and Snowflake announced a partnership to unify operational technology (SCADA, historians, sensor data) with information technology (ERP, maintenance records, production accounting) in a single queryable layer.
If you've ever tried to join a well's real-time production data with its maintenance history and its cost allocation, you know why this matters. That join is where most AI projects die — not because the model doesn't work, but because the data doesn't exist in a joinable form.
4. The Executive Signal: Follow the Org Charts and the Capital
Technology deployments tell you what's happening now. Executive decisions tell you what's coming next.
Occidental: Chief of AI
Occidental appointed Patrick Bangert — formerly of Samsung's global AI team, with a PhD from UCL — as their Chief of AI. When a major E&P creates a C-suite AI role and fills it with someone from big tech rather than promoting an internal champion, that's a signal about the scale of ambition.
Devon Energy: No Ambiguity from the CTO
Devon's CTO Trey Lowe has been direct: "Companies that don't adopt AI will get left behind." That's not a conference talking point — it's a strategic posture from one of the largest independent E&Ps in the US.
Devon + Coterra: $58B Merger with AI as a Stated Thesis
The $58 billion Devon-Coterra merger explicitly cited combining AI subsurface modeling capabilities as part of the strategic rationale. When AI appears in merger justifications at this scale, it has moved from R&D curiosity to board-level strategic priority.
Novi Labs: $35M and $50B in Capital Allocation
Novi Labs raised $35M and claims to be guiding over $50 billion in capital allocation decisions for E&P companies. Their models help operators decide where to drill, how to space wells, and how to design completions. If that $50B figure is even directionally accurate, it means AI is already influencing the highest-stakes decisions in upstream — well placement and capital allocation — not just optimizing execution.
5. What This Means for Mid-Size Operators
Here's the part that doesn't get discussed enough.
Every deployment listed above involves either a supermajor (TotalEnergies, Occidental), a large-cap independent (ConocoPhillips, Devon, Aker BP), or a major service company (SLB, Baker Hughes). These companies have dedicated data science teams, enterprise data platforms, and seven-figure AI budgets.
If you're running 200 wells in the Permian with a team of 15, none of that is directly applicable to you.
But the gap is closing, and it's closing from two directions:
From above: The service companies (SLB Tela, Baker Hughes Flow, Nabors/Corva RigCLOUD) are productizing their AI capabilities. You don't need to build your own drilling optimization agent — you need to be a customer of one that works. The question is whether these products will be priced for mid-size operators or only for enterprise contracts.
From below: Open-source and lightweight tools are making basic AI-assisted workflows accessible without enterprise infrastructure. Decline curve analysis, production anomaly detection, wellbore stability screening — these don't require a Cognite deployment. They require clean data and the right algorithms.
The operators who will struggle are the ones in the middle: too small for enterprise AI platforms, too set in their ways for lightweight tools. If you're still running your production optimization in spreadsheets and your decline curves in Harmony without questioning whether there's a better way, the gap between you and ConocoPhillips is going to widen.
The honest assessment: Most mid-size operators don't need agentic AI today. They need to get their data house in order, adopt modern engineering tools, and build the muscle memory of data-driven decision-making. The agents come later — and when they do, the operators with clean data and modern workflows will adopt them in weeks, while everyone else will spend months on data cleanup before they can even start.
Start with the Tools, Not the Hype
At Groundwork Analytics, we build engineering tools that bridge the gap between spreadsheet workflows and enterprise AI platforms. Our free tools — production data cleanup, fluid property calculations, well economics — are used by petroleum engineers who want modern, data-driven workflows without a six-month procurement process.
Dr. Mehrdad Shirangi is the founder of Groundwork Analytics LLC. He holds a PhD from Stanford in optimization and energy systems and has been working at the intersection of AI and petroleum engineering since 2006.