Posted inThe Lede

AI could plan an oil field in hours, but approving and financing it still takes time

Adnoc wants a single AI system to plan its oil fields, a job now handled by separate teams working one after another. Subsurface engineers hand the plan to facilities engineers, who pass it on to the economists. Adnoc uses the final version to decide which wells to drill, which processing plants and pipelines to build, how big they should be, and when. Once that work takes hours instead of months, “the bottleneck shifts towards validation, alignment and decision-making,” RIMS Digital Solutions founder and CEO Ruslan Suleymanov tells EnterpriseAM.

IN CONTEXT- Under a multi-year partnership with OpenAI and G42, Adnoc will use OpenAI’s Frontier platform to turn field master planning from a sequence of separate subsurface, facilities, and economic studies into a single, integrated decision, and it expects plans that took months to take hours, according to the company. The system builds on ENERGYai, the subsurface AI system Adnoc developed with AIQ, and is meant to combine decades of data from thousands of wells, multiple reservoirs, and hundreds of surface facilities with investment costs and commercial considerations.

Why it matters: Even if Adnoc only manages to speed up part of the way, more pressure is bound to fall on the next stage: the period between developing the project and making the final investment decision (FID). That’s “when the decisions that shape a project’s economics and eventual delivery are made,” Wytham Group co-founder Nicholaus Rohleder tells us. Many construction problems start with incomplete engineering or poor coordination between design and execution, and AI’s biggest opportunity is to improve design before substantial capital is committed, he says. The risk is that faster planning comes at the expense of design quality.

Developers can afford to explore more options before spending, while lenders get more time to pick apart the contracts. AI could help developers “identify critical dependencies earlier and secure the commitments necessary to make a project bankable,” Rohleder says. If they can show the engineering is actually better, lenders can turn more of their attention to weak contracts and unclear lines of responsibility.

Hurry up and wait

Good news or bad news first? The good news is that AI can draw up the plan in hours. The not-so-good news is that the holdup moves to getting someone to sign off on it. “The real question becomes not how quickly we can generate a plan, but how confidently we can approve and take ownership of it,” Suleymanov says.

It’s not all rainbows and sunshine at the drawing board either, and that’s where AI could do the most good. “Many of the difficulties encountered during construction originate much earlier, often in incomplete engineering, inadequate project definition or poor coordination between design and execution,” Rohleder tells us. “The greatest opportunity for AI is to improve the quality and thoroughness of design before substantial capital is committed.”

You’ve probably been told not to run before you can walk. It applies here, too. “Faster planning comes at the expense of design quality, particularly where assumptions are insufficiently validated or the transition into construction is poorly managed,” Rohleder argues. “The measure of success should be more reliable project delivery, not simply shorter planning timelines.”

Come fast, come slow, lenders will still ask the same questions: “The fundamentals of project finance remain unchanged, particularly the need for sound engineering, enforceable commercial agreements, creditworthy counterparties and clearly allocated construction risks,” Rohleder tells us. If the gains can be shown, lenders could spend more time on contractual weaknesses, gaps in responsibility, and critical dependencies, he adds, provided “investors understand precisely what the technology has improved, which assumptions have been independently verified and where uncertainty remains.”

In AI we trust (mostly)

So the tech may be ready. Is everyone else? “The main challenge is not necessarily the AI technology itself, but data quality, integration with existing systems, trust in AI-generated outputs, and people’s readiness to adopt new ways of working,” Suleymanov tells us.

Someone still has to put their name to it: That should be the people who know the asset and answer for how it performs, Suleymanov says. “Sign-off should come from the relevant technical and asset leadership. They need to challenge the assumptions, validate the data, assess technical and commercial risks, and make sure the plan is realistic and aligned with business objectives.”

AI might get a seat at the table, but it doesn’t get the final say. Adnoc says human experts will keep oversight of consequential decisions, a principle Suleymanov backs: “AI should strengthen human decision-making, not replace it.”

But who holds the keys?

Adnoc may own the house, but not everything inside it: Its data stays at home, while much of the technology running the system stays in US hands. “Hosting the models on domestic compute gives Adnoc data residency and physical custody, which matters for reservoir data that is among its most sensitive commercial assets,” June Park, political economist and senior resident at the Frontier AI Security Residency at ERA Cambridge, tells us. “But the model weights, update cycle, safety policies and licensing terms stay with a US vendor, and the chips underneath are US-designed and export-licensed.”

The more Adnoc relies on AI, the more it has to trust decisions made outside its control. “The UAE controls the location of the stack while the US retains control of its most consequential layers: the model, the silicon and the permission to operate,” Park argues. “The more field planning depends on these systems, the more Adnoc’s decisions are shaped by choices made elsewhere, such as which model version is available and how it behaves.”

And there's a deadline attached to all those chips: G42’s license-free access to advanced US chips runs until 6 April 2027. In July, the US moved the UAE into Country Group A:5, its most-trusted export-control tier, and named G42 and Core42 approved recipients of advanced computing items. That authorization lapses on 6 April unless the two become US companies or win further approval, as we reported in our Lede on the Gulf’s AI buildout. Adnoc can keep running OpenAI models on its own data only if both keep that status, Park tells us.

Washington is getting stingier with its chips, and that could spell trouble come April. Chip-export hawks now have the upper hand, making the renewal less certain. Nvidia CEO Jensen Huang’s abrupt downfall has hardened the US line on keeping advanced AI chips out of China, ending the administration’s mostly hands-off approach to chip exports. Commerce Secretary Howard Lutnick, whose department approved the UAE and Saudi chip deals, has been called to testify on them before the Senate Banking Committee next month. Expected incoming committee chair Elizabeth Warren has previously called the UAE agreement corrupt GO DEEPER: our piece on the Huang fallout for the Gulf.

The UAE has a way to keep its access, but nobody has explained exactly how. The deal floats the idea of G42 and Core42 becoming US companies to retain access to advanced chips, but doesn't spell out what either would have to do to qualify. If Commerce refuses to renew the authorization or attaches heavy conditions, Gulf AI would find it much harder to work near the cutting edge. That would also put pressure on Adnoc's planning system, which depends on that compute.

So far, so good for the Gulf: The US left the GCC off a new list of countries where data centers face extra licensing over chip-diversion risk. Indonesia, Malaysia, Singapore, Thailand, and Brazil made the cut. G42 and Humain “are government-backed entities with close relationships with many US firms,” Khalij Economics Director Justin Alexander told us. Chips already delivered also buy the region some breathing room: “It’s hard to take back what is already there,” former US Commerce Department export-control official William Reinsch told us.

Keeping your data at home doesn’t necessarily mean keeping control of it. The announcements describe sovereign hosting, but leave out the technical and contractual guarantees governing who can actually see Adnoc's data. “That detail, rather than the server location, determines how much control Adnoc actually has,” Park says. She expects tightening at the margins rather than a cutoff: “The realistic scenario is incremental: tighter conditions on hardware, auditing or usage, rather than an abrupt shutdown.”

Aramco went with “if you can’t beat them, join them” and bought into its AI supplier. It completed its purchase of 49% of Saudi AI company Humain for SAR 9.9 bn (USD 2.6 bn), leaving the Public Investment Fund (PIF) with 51%, and plans to use Humain’s services across its own operations.

You wouldn’t build an oil field around a supplier you could lose, so why plan one around a model you could? Producers that plan on someone else’s model should plan for losing it. “AI should inform field development plans, not own them,” Park argues. Operators should contract for continuity, keep the ability to run planning on an alternative model, retain ownership of their data, and test how planning would proceed if the system went down. That means “treating it as critical infrastructure with the same resilience planning as a pipeline or grid,” she adds.

THE BOTTOM LINE- A plan produced in hours can only be approved as fast as it can be checked, and those checks (sign-off, lender diligence, contracts that can withstand policy changes, and a fallback model) are now part of the infrastructure.