Verified Author Share
Notes from a closed-door dinner with energy, oil and gas, climate-tech and venture leaders. Houston, June 2, 2026.
In physical industries like energy, AI fails when it is treated as an autonomous answer engine instead of a guided instrument working alongside first principles and domain expertise. Across a closed-door dinner of more than a dozen oil and gas, energy transition, climate-tech and venture leaders in Houston in June 2026, the same diagnosis surfaced from operators, founders and investors who had never compared notes: the problem is almost never the model. It is the data integration, the undocumented operational knowledge, the executive who cannot say what business problem the AI is meant to solve, and the temptation to throw compute at a question that better process would answer for a fraction of the cost.
The phrase that organized the night was the inversion of the industry’s favorite slogan. Not “unleash your AI.” Leash it. Give it a clear objective, keep a human checking its work, anchor it in the physics of the asset, and only then let it run.
This pattern is consistent with broader data. IDC research cited across the energy sector attributes up to half of AI project failures to interpretability problems, the difficulty of trusting a recommendation you cannot trace. RAND Corporation found more than 80% of AI projects fail to deliver intended value, roughly twice the rate of non-AI IT projects. The dinner put human faces on a structural problem the research already confirms.
On a Tuesday evening in early June, in a Houston summer that arrived early, we closed the doors and sat down with a group of energy operators, oil and gas veterans, climate-tech founders and the investors who fund them. No slides. No pitch deck. One question to start, and a long table willing to argue.
What we heard was not a debate about which model is best. It was a much harder conversation about why so much of the AI spend in this industry produces motion without value, and what the people closest to the assets have learned about making it real.
This post is our record of that night, shared in the same spirit the room offered it: candid, specific, and without the polish that usually flattens these conversations into a press release. If you were at the table, this is a thank-you and a set of notes. If you were invited and could not make it, this is a window into what your peers are wrestling with when no investor is on the call.
The table was deliberately mixed. Energy transition founders building in hydrogen, carbon capture and produced-water optimization. Operators and advisors with decades inside the supermajors, from upstream to downstream, across Brazil, the Middle East, North America, Europe and Asia. A consulting and engineering procurement leader. Two venture investors, one running an operating-partner role inside a climate fund, the other raising his first fund focused on climate adaptation and resilience. An entrepreneurship professor who spent thirty-five years commercializing technology in oil and gas. A fiber-optic sensing founder who has deployed across thousands of miles of pipeline.
Most of the sharpest observations below were not confessions from the affected companies. They came from founders describing what they see inside the operators they sell into, from investors describing the companies they diligence, and from veterans describing the supermajors they spent careers inside. The dinner was a cross-diagnostic conversation, with each vantage point checking and correcting the others.
That is itself a finding. The clearest picture of an organization’s AI dysfunction usually comes from the people trying to serve it, not from the organization narrating itself.
A few of the figures shared at the table are striking, and we have kept them in. We have also kept them attributed to the people who claimed them, rather than presenting them as settled fact. A founder citing 20% more cash flow for his customers, or thousands of miles of monitored pipeline, is reporting his own results, not an industry benchmark. We have reserved the verified statistics for independent sources, named throughout.
One boundary worth stating plainly: the room’s center of gravity was oil and gas, energy transition and the infrastructure around them. Power, grid and data-center energy came up forcefully, but as adjacent pressure rather than direct representation. Utilities, renewables developers and downstream chemicals were not at the table. Read the patterns below as most reliable for upstream, midstream and energy-transition operations, and as hypotheses elsewhere.
The industry has spent three years being told to unleash AI. One of the host remarks early in the evening flipped it: in real business workloads, the work is not unleashing AI, it is leashing it. Giving it a defined task. Keeping a practitioner in the loop. Constraining it to the problem in front of you.
The phrase kept returning all night, in other people’s mouths, applied to their own problems. By the time dessert arrived, one founder who had opened the evening talking about unleashing AI in his own operation said, in front of the table, that he had changed his mind and wanted to clarify his earlier remark. The way the room had approached the question, he said, had made him feel better about AI than he had in a while.
That reversal is the heart of what we want to share. The leadership conversation in energy has moved past whether AI matters. It has landed on a quieter and more useful question: how do you keep a powerful, confident, occasionally wrong system inside the lines, in an industry where being wrong can fracture a pipeline or blow out a refinery.
Everything below is a variation on that theme.
The first substantive answer of the night set the tone. Asked how AI should be used in energy, a veteran with decades in technology commercialization argued that AI driven purely by data, with no grounding in the physics of the system, is a poor fit for this industry. Fluids, mechanics, geophysics, geology: the field already has rich first-principle models. AI earns its place by complementing them, by solving the hard, nonlinear, multidimensional optimization problems that physics alone struggles with, not by replacing the physics.
He told a story that drew knowing laughter. A consulting firm offered to mine an operator’s drilling data for hidden insights. Weeks later they returned with a finding: the rig stops roughly every sixty feet to change pipe. Anyone who has stood on a rig already knew that. The lesson was not that the AI was useless. It was that AI without domain context rediscovers the obvious at great expense.
The investors at the table reinforced the point from the diligence side. One, with fifteen years digitizing oil and gas before moving to venture, noted that the sector has used machine learning on specific problems for over a decade, from seismic interpretation to drill planning. What worries operators now is agentic AI, which behaves like a black box. In an industry built on traceability and accountability, a system that returns an answer without showing its reasoning is a hard sell. What has worked, in his experience, are hybrid and physics-informed models, where AI accelerates a problem you already know how to solve with physics, cutting the computational load and the time to answer rather than guessing the answer outright.
The independent data tracks this caution. DNV reported that roughly 47% of senior energy professionals intended to bring AI-driven applications into operations, a real commitment paired with deliberate restraint, precisely because model hallucination in a high-stakes setting is unacceptable. Field engineers, as one industry analysis put it, will not trust a recommendation they cannot trace, and that trust deficit is one of the largest brakes on adoption.
Takeaway: In energy, the question is not whether to use AI. It is whether your AI is anchored to the physics of the asset. If it is not grounded in first principles and domain knowledge, it will produce confident output that an experienced operator would never sign off on.
When the conversation turned to whether the obstacle is the technology or something else, a consulting and engineering procurement leader answered without hesitation: it is never the technology, it is always the humans. Good technology fails when it is rolled out without adoption. He pointed to the long history of promising tools that died for lack of uptake, and argued AI is now repeating the pattern.
His most useful contribution was operational. On a project to merge public and proprietary data and figure out where the hydrocarbons were, his team learned that letting the AI integrate the data in isolation produced garbage on the back end. What worked was sitting a data scientist next to a geophysicist and a geoscientist, so a human could sense-check whether the model’s output was a true result or just a mathematically possible permutation that never happens in the real world. Treat AI as a tool, he said, and remember that even a hammer requires a practitioner who knows how to use it. Hand people a hammer and ask them to use it like a saw, and you get nothing.
A former Shell downstream leader, now building a subsurface hydrogen company, extended the point with what he called the rule of five for maintenance: the right people, at the right time, in the right place, with the right tools, and the right paperwork. He has yet to see a single system that covers all five. Most cover one or two. The gap, he argued, is integration and translation, getting the language of one database to talk to the language of another, from inventory to contractor timesheets to the supply chain. That is where AI can genuinely help, but only if it is guided, because left to itself it generates so many possible solutions that the system collapses under them.
The fiber-optic sensing founder told the most complete version of this story. Bringing AI-based monitoring to operators in India and beyond, the technology was never the hard part. The hard part was training control-room officers and field staff who came from petroleum backgrounds and had no exposure to signal processing or programming. His team started with the first principles of how the sensing works, then layered the machine learning on top, deliberately and transparently, building trust with the people doing the daily work. Over four to five years that approach produced tens of thousands of hours of labeled field data, which became the backbone of a physics-informed model competitors could not easily replicate.
This matches what the broader market keeps reporting. The shortage of professionals who understand both AI and energy operations is cited as a primary brake on adoption, and the operators making progress are the ones investing in cross-domain capability rather than buying a tool and hoping.
Takeaway: Budget for adoption, not just acquisition. The decisive investment is putting domain experts and data people in the same room, and training the workforce that will live with the system long after the vendor moves to the next project.
A theme that ran under everything: AI is excellent at ingesting and merging large, messy data sets, and useless at knowing which merge matters.
One story landed hard. A company was studying its own cost structure, and consultants offered to do the work for half the going rate. Before signing, someone looked closer. The marketing line in the top-level books was carrying not just marketing spend but the abandonment liability for the company’s assets. Strip that out, and the apparent marketing budget collapsed, which finally explained why the marketing team had been insisting for years that they were underfunded and could not reach enough people. The numbers had been technically correct and completely misleading. The work AI should do, the storyteller argued, is translate from the language of the books to the language of the business, deciphering how charge codes entered at the asset level get aggregated and distorted at the top. No model does that on its own.
The produced-water and oil and gas software founder framed integration as the precondition for value. His pitch was concrete: an average operator might spend fifty cents a barrel simply to dispose of produced water, across millions of barrels, while nearby communities need water. Drive the recycling cost below disposal cost and everyone wants it. But getting there requires integrating existing data, first-principle models, a reasoning layer and a cash-flow metric, across what he described as a dozen layers of technology touching several large data sets. Integration is not the unglamorous prelude to the AI work. It is the AI work.
Takeaway: If your systems do not talk to each other, AI cannot reason across them. The companies skipping integration to get to the visible AI faster are building on sand, and the cost shows up later as outputs nobody can trust.
The most commercially revealing thread came near the end. A carbon-capture advisor with twenty-five years in oil and gas trading described a recent conversation with three senior executives, a CEO, a chief commercial officer and a chief operating officer, all of them seasoned, none of them able to articulate what AI would actually do for their business. They were being pressed by investors on every call about their AI strategy, and so they were pushing the mandate down through the company without a clear problem to attach it to. A top-down demand for an answer that the top did not have.
One of our hosts confirmed the pattern from the consulting side. He had stopped attending many conferences, he said, because executive after executive would take the stage to declare that AI changes everything, with no content behind it, beaten into the script by investor pressure. His response had become a precondition for any engagement: before scoping a project, run a baseline session so the leadership team actually understands what is and is not possible, after which a real conversation becomes possible and the efficiency of every later discussion improves.
This gap is not anecdotal. EY’s 2026 research notes that deriving results from AI investment has been on the minds of energy C-suite leaders for at least three years, and that the path forward in oil and gas runs through cohesive data and integrated operations rather than another tool purchase. McKinsey’s global survey found that while 88% of organizations now use AI in at least one function, only 39% report any enterprise-wide EBIT impact. The boardroom mandate and the operational result have come unbolted from each other.
Takeaway: A mandate is not a strategy. If your executives cannot name the specific business problem an AI initiative is meant to solve, the initiative will produce activity and burn credibility. Start with a shared baseline at the leadership level, before the budget is committed.
The sharpest disagreement of the night was productive. One host argued that education is the unlock, that a six-hour workshop with executives sets a baseline that makes everything afterward work. A founder pushed back on the word itself. Education, he said, is a luxury you rarely get with a busy executive, and it is not the real job anyway.
His reframing was the most quoted line of the evening: take your stakeholders seriously, do not take them literally. What presents as the problem is often a symptom of something deeper. The work is not teaching someone what a tool can do. It is doing the hard work of understanding what they are actually trying to accomplish, which they frequently cannot articulate themselves. He gave a benchmark from an operations company where hands-on tool time sat at 17%, meaning staff spent the overwhelming majority of their hours on something other than the work everyone assumed they were doing. AI does not even enter the picture until you understand that. Used well, AI helps you find the gap. Used as a starting point, it just makes prettier slides for a problem nobody has defined.
The entrepreneurship professor tied it to value, which he argued the whole industry struggles to measure honestly. It is not enough to ask for bottom-line impact. At what scale, over what horizon? He warned about a specific failure mode: machine learning that books savings month after month, then misses the rare catastrophic event, a refinery blowout, an instability nobody caught, that wipes out years of accumulated gains and leaves you deeper in the hole than before. This, again, is why the system needs a leash. It also why he worried aloud about engineers losing their own judgment as they lean on the tool, walking with it on every project until the underlying skill atrophies.
Takeaway: Lead with diagnosis, not training. The highest-leverage work happens before the AI is chosen, in understanding the real problem precisely enough that you can tell whether AI is even the right instrument for it.
One exchange reframed the entire conversation. Asked how hard it is to enter a market hungry for AI, a hydrogen founder corrected the premise. He is not selling AI. He is selling energy, so that energy can power AI. That is a different business, and a fast-growing one.
He described a North American reality that should concern anyone building in this space. In parts of the Northeast, data centers have become the local villain, and approvals stall unless a project brings its own power, because drawing from the grid drives up costs and undermines affordability for the surrounding community. Clean and affordable, he argued, are the two things people consistently neglect, and they are now the gating factors.
The numbers behind that intuition are stark. The IEA reports that data-center electricity demand grew 17% in 2025, far outpacing the 3% growth in overall electricity demand, while consumption from AI-focused data centers surged roughly 50% in a single year. Its base case sees data-center electricity consumption roughly doubling to around 945 terawatt-hours by 2030, with AI-specific demand tripling. The agency’s executive director put it as bluntly as the dinner did: there is no AI without energy. In the United States, the Department of Energy projects AI could push data centers to as much as 12% of national electricity consumption by 2028, and Goldman Sachs forecasts a 165% increase in global data-center power demand by 2030 against 2023 levels. With grid connections slow, developers are increasingly turning to on-site generation, much of it natural gas, which puts the energy industry squarely in the middle of the AI buildout.
For energy leaders, this is the quiet strategic opening. The AI conversation is usually framed as something happening to your operations. It is also a demand surge happening for your product.
Takeaway: AI is not only a tool to deploy inside the business. For energy companies, it is also a customer. The constraint on the AI economy is increasingly power, and the operators who can supply it cleanly and affordably hold more leverage than the framing of “AI adoption” suggests.
Since two investors were at the table, we asked directly how they separate a company with real AI potential from one riding the wave. Their answers converged.
Technology, both said, is rarely the differentiator. There are enough capable people to build the right architecture for most problems. What they underwrite is the team, whether they have done it before and worked together before, and product-market fit defined broadly, meaning not just whether the technology works but whether there is a real market, a real customer, a credible path to scale revenue, and a viable distribution channel. One pointed to the insurance sector as a cautionary example: the technology was not the challenge and the data was not the challenge, but without a way to distribute the product, the market fragments and nobody scales.
A second concern has sharpened in the past year. With foundation models advancing quickly, investors now worry that a startup’s moat can evaporate when a major lab ships a new capability. The companies that hold up, in their view, are the ones with genuinely differentiated foundations, proprietary first-principle models, hard-won labeled data, real domain depth, the kind of thing that takes years to build and cannot be replicated by an API call.
Takeaway: Whether you are raising capital or evaluating an AI partner, weight the team, the data moat and the distribution path above the model. Those are the things a competitor cannot copy overnight.
The conversation kept circling a handful of tensions that most organizations are currently resolving by doing both, and paying for it in fragmentation and stalled projects.
None of these has a universal answer. The organizations struggling most are not the ones choosing wrong. They are the ones not choosing, while behaving as if they had.
Two hours of conversation produced more clarity on what breaks than on what to do next. That asymmetry is itself the lesson: the discipline lives in the sequence, not the strategy. The order the room kept returning to:
Late in the evening, the conversation turned from operations to something harder. What is sacred? What should AI not touch?
A consulting leader, taking the technical side off the table for a moment, said what makes us human is our creativity, and that should not become the work of AI. The fear he named, half joking, was the future where the machines write the poetry and the humans mop the floors. The room’s host picked it up and inverted it toward something hopeful: a smart city, in her framing, is a city for humans, not for sensors and robots, a place where people can be more creative and live better, not merely a grid of instrumentation. The value, several agreed, is in the human outcome, and the industry’s habit of chasing use cases for their own sake gets that backward. The point is not to find a problem for the AI. It is to find the value for the people.
We do not have a clean answer to that question, and we are suspicious of anyone who claims one. But it is the right question, and the leaders willing to hold it, even while the market rewards speed, are the ones we would bet on for the long run.
Based on the June 2026 closed-door dinner of energy and oil and gas leaders, AI projects in the sector fail most often for non-technical reasons: AI deployed without grounding in the physics of the asset, data that lives in disconnected systems the model cannot reason across, undocumented operational knowledge the AI never captures, and executive mandates with no defined business problem attached. This aligns with independent research. IDC findings cited across the sector attribute up to half of AI project failures to interpretability problems, and RAND Corporation found more than 80% of AI projects fail to deliver intended value, roughly twice the rate of non-AI IT projects.
Physics-informed AI combines machine learning with the first-principle models the industry already trusts, such as fluid dynamics, geophysics and reservoir physics. Rather than guessing an answer purely from data, the AI accelerates a problem you already know how to solve with physics, reducing computational load and time to answer. Energy leaders favor this approach because it preserves traceability, the ability to understand why a system produced a given recommendation, which is essential in safety-critical and capital-intensive operations where an unexplained black-box output cannot be trusted.
With caution. The investors and operators at the June 2026 dinner were skeptical of fully autonomous agentic AI in high-consequence energy operations, precisely because it behaves like a black box in an industry built on traceability and accountability. What has worked far better are hybrid and physics-informed models, where AI speeds up a process the operator already understands and a human remains in the loop to verify outputs against operational reality.
AI is both a tool energy companies can deploy and a major new source of demand for their product. The IEA reports that data-center electricity demand grew 17% in 2025, with AI-focused data centers surging roughly 50% in a single year, and projects data-center consumption roughly doubling to around 945 terawatt-hours by 2030. In the United States, the Department of Energy projects AI could push data centers to as much as 12% of national electricity consumption by 2028. With grid connections constrained, developers are increasingly building on-site power, much of it natural gas, which places the energy industry directly inside the AI buildout.
Start with diagnosis, not deployment. The dinner’s clear consensus was that the highest-leverage work happens before any tool is chosen: understanding the real business problem precisely enough to know whether AI is the right instrument. Leaders should establish a shared, honest baseline at the executive level before committing budget, anchor any model in first principles and domain expertise, and ensure systems are integrated enough for the AI to reason across them.
The June 2, 2026 closed-door dinner in Houston, hosted in partnership with Plug and Play Texas, brought together more than a dozen leaders across energy and oil and gas: founders building in hydrogen, carbon capture, produced-water optimization and fiber-optic sensing; operators and advisors with decades inside the supermajors; a consulting and engineering procurement leader; two venture investors focused on climate and energy; and an entrepreneurship professor with deep technology-commercialization experience. Power, grid and data-center energy were discussed as adjacent pressures. Utilities, renewables developers and downstream chemicals were not directly represented, and the findings should be validated independently before applying them to those sectors.
This was one conversation, and we intend to host more. If you were at the table, thank you for the candor. If you recognized your own organization in any of the patterns above, we would like to continue the conversation. A follow-up gathering is already on the calendar, and the people who were in the room have an open invitation to the next one.
Until then: diagnose before you decide, anchor the model in the physics, and keep the leash on.
Methodology: This post synthesizes a closed-door dinner of energy, oil and gas, climate-tech and venture leaders held in Houston on June 2, 2026, in partnership with Plug and Play Texas. The conversation was recorded and transcribed with consent and analyzed across thematic patterns. Direct attributions to individuals and specific employers have been withheld to preserve the candor of the exchange. Figures stated by participants are presented as their own claims, not as verified benchmarks. External statistics are drawn from independent sources, named and linked throughout.
About Zallpy: Zallpy is a technology consulting and software engineering firm that helps enterprises turn AI, data and engineering into measurable business results. We host these conversations because we believe the distance between AI ambition and operational reality is the defining business problem of the decade, and that closing it requires the kind of honest exchange that rarely happens in a conference room.
Event partner: This convening was held in collaboration with Plug and Play Texas.
Citations and sources:
This post was published in June 2026 and reflects the state of energy AI adoption as observed at that time.
Verified AuthorA strategic technology consultant with experience supporting CTOs, CIOs, and engineering leaders in identifying and resolving execution constraints that impact delivery speed, stability, and organizational agility. At Zallpy, he works within complex enterprise environments diagnosing bottlenecks across architecture, integrations, data flows, and delivery processes, helping teams uncover root causes and implement practical, sustainable solutions. He collaborates closely with technology organizations to improve system reliability, strengthen integration resilience, and increase delivery visibility, while also focusing on knowledge transfer and capability building so teams can sustain improvements independently and scale with confidence.
A strategic technology consultant with experience supporting CTOs, CIOs, and engineering leaders in identifying and resolving execution constraints that impact delivery speed, stability, and organizational agility. At Zallpy, he works within complex enterprise environments diagnosing bottlenecks across architecture, integrations, data flows, and delivery processes, helping teams uncover root causes and implement practical, sustainable solutions. He collaborates closely with technology organizations to improve system reliability, strengthen integration resilience, and increase delivery visibility, while also focusing on knowledge transfer and capability building so teams can sustain improvements independently and scale with confidence.
Verified Author
Verified Author
Verified Author