Most AI conversations in a refinery go well until the AI gets near anything that matters. The moment a model touches a safety-critical operation, a different set of rules applies. Process Safety Management governs the unit. Any change triggers Management of Change. A Process Hazard Analysis has to account for it. Operators need training on it. For a generic AI vendor, this is where the pitch quietly ends.
For an industrial operator, it's where the real requirements begin, and it's exactly the environment we design for.
Safety-critical changes the AI question
Outside the fence line, an AI tool can be evaluated on accuracy and convenience. Inside it, the questions are different. Can a reliability engineer explain to a regulator why the AI recommended what it did. Does a human stay in the loop on consequential decisions. Does the deployment survive MOC review. Does it fit the audit trail the operation already maintains for SEC disclosure, EPA reporting, and PHMSA integrity obligations on any pipeline assets.
These aren't reasons to avoid AI in operations. They're the specification it has to be built against. A vendor who treats them as friction will lose at the first PHA review. A vendor who treats them as the design brief ships something that lasts.
Human-in-the-loop as a design principle
The reliable pattern in safety-critical work is straightforward: AI surfaces the analysis, a qualified person makes the call. That isn't a limitation bolted on to satisfy a committee. It's the design that lets the system be useful and accountable at the same time.
Built this way, the AI does what it's genuinely good at, reading across the historian, the CMMS, and years of inspection history faster than any person could, and hands a clear recommendation to the engineer who owns the decision. The human stays accountable. The work gets faster and better documented. The regulator gets an explainable trail.
MOC-compatible from the start
The difference between an AI that ships in a refinery and one that stalls is usually whether change management was considered on day one. A system designed to be MOC-compatible carries its documentation with it, makes its reasoning inspectable, and fits the operator's existing approval and training processes rather than asking them to invent new ones.
This is the part coastal AI firms tend to miss, because they've never sat through a PHA. Cortland speaks the language of the unit, which means the governance work happens up front, not as a surprise at the gate.
The evidence
This isn't theoretical for us. Kaysee, our reliability platform, runs in operator environments with human-in-the-loop preserved and governance documented. It's the proof point for any conversation about AI in safety-critical operations: production AI that respects the controls instead of fighting them.
The path into it follows the same ladder as the rest of our work. A Walk engagement maps where AI can and can't responsibly go in your operation and what governance each use case requires. For the ops and HSE leaders who want to ground that conversation in a working understanding of the stack first, the Practical AI Warmup is the entry point.
Governance isn't the thing standing between you and AI in operations. Done right, it's the thing that lets the AI stay.