Two Languages, One Discipline

There is a number that should stop any chemicals or energy executive mid-scroll: of every industry McKinsey has measured for generative AI adoption, chemicals and energy sit dead last, with 14% exposure to the technology against a cross-industry average of 23%.¹

If you run a plant, that number probably doesn’t feel like a gap. It feels like discipline. Heavy industry doesn’t adopt things because they’re exciting; it adopts things once they’ve been proven safe at scale, under real conditions, with a rollback plan. That instinct is correct, and it turns out AI has exactly the same instinct, even if the two industries have never compared notes on it.

MIT’s Project NANDA looked at 300 enterprise AI deployments across every sector in 2025 and found that 95% of them showed no measurable financial return.² The interesting finding wasn’t that AI underperforms. The failures were concentrated in organisations that skipped straight from pilot to full deployment, with no staged validation in between. The 5% that worked did something specific: they treated putting an AI system into production the way a process engineer treats commissioning a new unit. Prove it in a controlled environment first, then let it touch something real, under supervision, with an operator who can pull it back.

That is not a coincidence. It is the same engineering culture, wearing two different vocabularies.

The same discipline, two dialects

In October 2025, Saudi Aramco and Yokogawa commissioned multiple autonomous AI agents to directly control acid gas removal operations at the Fadhili Gas Plant. It was one of the first times an AI system had been given real-time authority over a live industrial process, rather than just a recommendation to hand to an operator.³ The results are concrete: a 10 to 15% reduction in amine and steam consumption, roughly 5% lower power use, and measurably fewer manual interventions.³

What makes this worth studying isn’t the result. It’s the sequence that got there. The agents were trained and tested first in a full plant simulator. Only once they performed reliably there were they progressively integrated into Yokogawa’s live control system, one section at a time, with the plant’s existing safety interlocks left fully intact underneath.

Autonomy was earned in stages, never assumed on day one.

If you’ve spent a career commissioning industrial assets, you already have a name for that: phased commissioning, proven first in simulation, then brought online section by section, with safety systems that can always override the new equipment.

If you’ve spent a career shipping AI into production systems that touch real money or real customers, which is where I’ve spent mine, building fraud detection and transaction monitoring systems and real-time logistics systems for delivery platforms, you also have a name for it, just a different one: shadow mode, then canary release, with a human-in-the-loop override that can always take back control.

Neither industry invented this out of caution for caution’s sake. Both learned it the same way: by watching what happens when you skip it. The 95% of AI projects that never pay off, and the industrial incidents that get written up in safety case studies, tend to share one root cause.

Something was trusted with production before it had earned it.

Why the risk conversation is about to feel familiar, too

There is a second place where these two worlds are about to converge, and it’s less comfortable: data.

On 3 June 2026, the European Commission proposed the Cloud and AI Development Act (CADA), a framework built explicitly around reducing Europe’s dependence on a small number of non-EU cloud and AI providers, with graduated “sovereignty assurance levels” tied to the kind of data a system can be trusted to handle.⁴ Separately, as of 2 August 2026, the transparency duties in Article 50 of the EU AI Act are legally binding.⁵

It’s worth being precise here. The AI Act’s heaviest obligations, those covering high-risk systems under Annex III, were pushed back to December 2027 under the EU’s Digital Omnibus package.⁵ Nothing about AI regulation in Europe is finished, and headlines claiming otherwise are getting ahead of the text. But the direction is unambiguous, and it was set in law twice in the same two months: where your AI infrastructure physically sits, and who else can see what runs through it, is moving from a preference to a policy question.

For a plant operator, this should sound familiar, because it’s the same category of risk this industry has always managed, just with a new kind of asset. A proprietary feedstock formulation, a plant’s real-time process data, its cost structure relative to competitors: these have always been treated as assets worth protecting, the same instinct behind trade secret controls and careful JV governance.

Running that same data through a general-purpose AI model hosted by a third party, outside the company’s control, is a new instance of an old problem, not a new problem. The organisations already thinking about where their AI runs, not just whether to use it, are applying an instinct this industry has had for decades. They’re just applying it to a new kind of pipeline.

The point of writing this down

None of this argues that heavy industry should move faster on AI, or that AI should move faster into heavy industry. It argues something narrower: the two disciplines already speak the same language underneath the terminology. The organisations that will get real value out of AI, in chemicals, in energy, in any capital-intensive, safety-first industry, are the ones that recognise that and act on it, rather than treating AI adoption as a separate, unfamiliar risk category that needs its own rulebook from scratch.

It doesn’t.

It needs the rulebook this industry already has, applied to software instead of steel.

Sources

1. McKinsey Global Institute, generative AI adoption by sector: energy and materials (including chemicals) at 14% exposure against a 23% cross-industry average. mckinsey.com ↩
2. MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025” (July 2025). Forbes coverage, 26 August 2025. ↩
3. Aramco and Yokogawa, autonomous control AI agents at the Fadhili Gas Plant, announced October 2025. Yokogawa press release. ↩
4. European Commission, proposal for the Cloud and AI Development Act (CADA), 3 June 2026. digital-strategy.ec.europa.eu ↩
5. EU AI Act, Article 50 binding from 2 August 2026; Annex III high-risk regime deferred to 2 December 2027 (Digital Omnibus). ComplianceHub.Wiki. ↩