Responsible AI Made for the Asset: Takeaways from Future Oil & Gas 2026

Future Oil & Gas 2026 landed on a theme that felt less like a slogan and more like an industry finally naming its problem: “Encouraging Responsible AI.”

Over two days in Aberdeen, speaker after speaker arrived at the same conclusion from different directions. The value of AI in oil and gas not only depends on which model you buy, but also on whether your organization trusts the data underneath it, has governed the way data gets used, and can get the right information to the right person at the right moment. That last phrase could have been lifted from any of a dozen sessions, and it happened to be the closing line of ours.

Why pilots stall before they ever reach the field

CNOOC's Stuart Houston opened a thread that ran through the whole event: AI proofs of concept rarely die because the model was wrong. They stall on business alignment, traceability, governance, and the absence of subject-matter-expert validation.

We've watched this play out from the operator side for years, and it's the reason our own session opened with the observation that AI has already transformed the back office— reviewing contracts, drafting emails, and building slide decks. Ask a room full of people whether they've used AI this week and nearly every hand goes up. Ask what's changed out on the asset, and the room goes quiet. Crews are already lean. The equipment still needs hands on it. The work hasn't gotten any less skilled.

The reason isn't a lack of ambition. Deloitte's 2025 Smart Manufacturing Survey of 600 executives found that 92% expect smart technology to be the single biggest driver of competitiveness over the next three years—yet the average productivity gain manufacturers report today sits around 20%. That gap between belief and result is exactly what Houston was describing. Something is being missed in the deployment, the usability, or the fit with how the work actually happens.

Data foundations, not more data

Shell's Menno van Vechel set the tone in the opening keynote: responsible AI depends on trusted data foundations, redesigned processes, and technology-agnostic architecture—not on layering more AI on top of the same fragmented systems.

This is the same wall we see generic AI hit every time it meets a real plant. It's trained on clean, corporate text—documents, spreadsheets, email—and none of that prepares it for a P&ID, an engineering data sheet, or a historian tag that has never been connected to the work order it belongs to. It can't reach into the ERP, the EHS system, or the inspection record where the actual work lives, so it can't combine them into an answer anyone can act on. And critically, it doesn't leave room for the human in the loop—on the asset, AI must augment the expert's judgment, not replace it.

Operators told us this directly, and their words framed our whole session:

• Planners lose up to 80% of their time to admin—copying data between screens and checking status codes instead of planning.
• Historian tags and maintenance data live in separate worlds. Bridging them, one customer told us, is "the holy grail."
• Scheduling work on a confined-space asset can mean manually checking five systems to confirm permits and isolations. Miss one, and someone gets hurt.
• A single well going down cascades through the compressor station, pipeline capacity, and truck schedule. Nobody has a system that sees the whole chain.
• Turnaround planners spend their time on admin when they should be asking how to take a job from ten hours to five.

None of these are data problems in the abstract. They're context problems—exactly what van Vechel was pointing at.

The move toward agentic

Our session, "Grounded, Governed, Embedded: Responsible Agentic AI for Oil & Gas Operations," sat squarely in the conference's most forward-looking thread, alongside other speakers weighing how far organizations can push past human-in-the-loop AI toward more autonomous operations.

We took a more practical view: Agentic AI's job—today and in the foreseeable future—is to connect field data, ERP systems, work orders, and asset context, so the right person gets the right information at the right moment.

That's the distinction we draw between horizontal AI and vertical AI. Horizontal, generic AI suggests and summarizes. Vertical AI—built for the asset, not adapted to it—interprets, decides, and executes inside the operational workflow itself. It works because it's already integrated with the ERP, whether that's SAP, Oracle, or IBM Maximo, with no exports, no parallel systems, and no migration. What's in the ERP is what the AI works with, and what it produces flows straight back.

Where we fit

We walked the room through practical applications of agentic AI today:

• An AI-assisted planning and scheduling engine that turns days of manual schedule-building into minutes—including work-order assistant that assembles and enriches orders instead of someone rebuilding them by hand.
• AI-enabled scheduling optimization for upstream operations, unifying well delivery, asset development planning, and production operations into a single schedule — with "what if" scenario modeling that shows the cost and production impact of a decision before it's made.
• AI purpose-built for shutdowns and turnarounds—the largest controllable event on a facility's calendar, and the one most exposed to overrun—that accelerates work-package creation from historical norms and surfaces risk before the freeze date.
• An asset performance management solution that monitors sensor data across critical and balance-of-plant equipment to catch and prioritize reliability and efficiency issues before they become failures—moving teams from reactive to predictive maintenance.
• A process-safety layer that keeps barrier management, permits, and isolations visible and validated before work proceeds, so safety is built into the workflow rather than bolted-on after.

The rest of the room, in brief

The breadth of practical work on display was worth noting even outside our own focus.

Galp showed how OCR, generative reasoning, and human oversight can strengthen upstream audits. NEO NEXT +ENERGY applied deep learning to 3D bathymetry data for subsea pipeline integrity. Roderick Perez explored reinforcement learning against seismic data for drilling strategy.

On the people side, Infosys's Ian Gaylard and Avinash Darisa, along with MOL Group's Achilles Georgiu, made the case that none of this works without workforce readiness, clear problem definition, and “strong human judgment guiding the data,” as Georgiu put it. ITI Group's Steve Taylor framed the platform question the industry is quietly having: standardize on existing systems or embrace disruption?

And on the transition side, Floating Power Plant's Anders Køhler and ExxonMobil's Michel Teughels looked at what open digital ecosystems and integrated renewable generation mean for powering the assets AI is meant to serve.

Different problems, same undercurrent: none of it scales without trust, governance, and a foundation that respects how the work actually gets done.

Where this leaves us

We already have the ERP integration. We've already resolved the siloed data problem.

That's not a claim of having finished the job. It's what lets us spend our time on the operation instead of the “plumbing”, and it's why we keep building around what our customers tell us adds value—not around what a generic model seems to do well.

If you were in the room in Aberdeen, thank you for the conversation. If you weren't, we'd still like to have it. Reach out to your Prometheus Group contact, or request a live demo, and let's talk about where AI actually fits your operation.