Most of the conversation about AI in the energy sector right now is happening at the wrong altitude. The press wants to talk about demand forecasting and grid optimization. The industry wants to talk about chatbots in the control room. The actual leverage is somewhere else entirely: it is in the digital substrate underneath a brand-new plant.
The Legacy DCS Problem
The average mid-sized US refinery is running a distributed control system that was designed before the iPhone. The vendor lock is real, the alarm management is reactive, and the procedural automation is limited to whatever the original commissioning team was able to configure inside the project budget. Replacing that DCS in place is a five- to seven-year program that nobody approves, because the cash return on a like-for-like upgrade is too thin.
A Clean-Sheet Build Is the Cheap Window
The economics flip the moment you are commissioning a new plant. The DCS, the historian, the alarm system, the procedural automation layer, and the PSM document set are being purchased at the same time anyway. The marginal cost of buying them as one integrated AI-native stack instead of four separate vendor-locked silos is small. The operating cost difference over twenty years is not.
What “AI-Native” Actually Means at the Plant Level
An AI-native control stack collapses three things that have always been separate: live process data, the PSM document set, and the operator’s daily workflow. Every alarm is rationalized against the current operating procedure. Every operating procedure is traceable to the standard edition that approved it. Every MOC is auto-drafted from a one-line description and routed to the right reviewers. Every PHA recommendation is tracked from open through closed across the full five-year cycle. None of that is science fiction. The math is straightforward; the integration is the work.
The Audit Story Changes
The compliance auditor’s question is no longer answered by pulling a binder. It is answered by querying a system. Asked to produce the RAGAGEP citation that supports a piping wall thickness calc done in 2024, an AI-native plant returns the exact API 570 edition number, the calculation worksheet, the inspector who signed it, and the next inspection due date in under a second. Asked the same question, a legacy plant produces a paper printout from someone’s filing cabinet two days later — if at all.
The Operator Headcount Math
A ten-thousand BPD micro-refinery built on a legacy stack would need eight to ten board operators across three shifts, plus an engineering team to keep the document set current. The same plant built on an AI-native stack runs comfortably with four to five board operators across three shifts and an engineering team that spends its time on optimization rather than paperwork. That is the OPEX delta that pays for the entire control system before year three.
Why Day One Matters
Retrofitting an AI layer onto a commissioned plant is harder than designing one in. The cable trays were not sized for it. The naming conventions across the historian, the DCS tags, the equipment list, and the P&ID drawings are not unified. The PSM document set was written in Microsoft Word and lives in three SharePoint folders. None of those problems exist if you specify the integrated stack as the project basis of design from day one. That is the window we are inside today, and it closes the moment the EPC contract is awarded.
The Bottom Line
If you are sponsoring a new modular refinery in 2026, the most expensive line item on the project is not the hydrocracker, the column, or the EPC contract. It is the cost of getting the digital substrate wrong. The plant will run for thirty years. The control and compliance stack you choose at FEED will define the operating cost profile for every one of those years. Choose carefully — and choose once.
Timothy Porritt is the founder of Porritt Inc., a Salt Lake City–based AI-powered engineering software company building the NEXUS suite for refining, process safety, and modular plant design.