A plant gets pitched plenty of AI. The trap is picking by capability instead of by the two things that actually matter in refining and petchem: does it move margin, and can it be deployed inside the safety envelope. Plenty of impressive models fail the second test, looking great in a simulation and never clearing Management of Change for the real unit. The use case worth doing sits in the overlap of high margin and real deployability.
Margin is the yardstick
Refining and petchem economics are made in the process. Yield, energy intensity, throughput, and equipment availability are the levers. An AI use case that doesn't touch one of those is a science project. One that improves yield a point or trims energy intensity pays for itself fast. So the first screen is simple: which lever does this move, and what is that worth.
The candidates
For most plants, the high-margin AI use cases cluster in a few places:
- Yield and process optimization. The biggest prize and the hardest to deploy, because it touches the control envelope and draws the most MOC scrutiny.
- Energy intensity. Often high-value and lower-risk, because much of it can be advisory rather than closed-loop.
- Equipment reliability and availability. Protects uptime on critical rotating equipment, sensor-rich, and typically advisory, so it clears review more easily.
Each moves margin differently and carries a different MOC burden. The right first move balances both.
The deployability screen
Before committing, run each candidate through three questions:
- Can it pass MOC as designed? If it directly manipulates the control loop, the safety case is heavy and the timeline long. If it's advisory, with a person in the loop, the path is far shorter.
- Does it run on data you already have? Favor use cases that read the historian and DCS you already operate over ones that need new instrumentation.
- What is the margin worth? Put a number on the lever it moves. A modest, deployable improvement beats a large one stuck in PHA review indefinitely.
Start where margin and deployability overlap
The instinct is to chase the biggest prize, full yield optimization, first. Often the smarter first move is the high-value, advisory use case that clears MOC quickly, proves the pattern, and earns the credibility to tackle the control-envelope work later. Start in the overlap, prove it on one unit, and let the margin result fund the harder next step.
The plants that get value from AI aren't the ones that picked the most ambitious use case. They're the ones that picked the one that moved margin and could actually be deployed.