There's a fork at the start of every plant AI project, and it decides almost everything that follows: does the model write to the control loop, or sit beside it. Closed-loop control is where the dream of full optimization lives, and it's also where the safety case becomes enormous and the timeline stretches into years. Advisory intelligence, sitting on top of the control system and recommending to a person, delivers most of the value far sooner and clears review far more easily. For most plants, advisory first is the right build.
Advisory beats closed-loop as a starting point
A closed-loop system that adjusts setpoints automatically has to prove it's safe under every condition, because no human is between it and the process. That's a heavy safety case for good reason. An advisory system surfaces a recommendation and a person decides. The value is similar, the risk is far lower, and Management of Change is far simpler because a human stays in the loop. You can always graduate specific, proven recommendations toward automation later. Starting there rarely makes sense.
Reading the systems of record without writing to them
The build pattern is to read widely and write narrowly. Pull from the DCS, the historian, and the advanced process control layer to understand what's happening, and surface intelligence back to the engineer or operator, without writing to the control loop. The systems of record stay exactly where they are and keep doing their job. The plant gains a layer that turns all that telemetry into a timely recommendation, and nothing in the safety-critical path changes.
Keeping the safety boundary intact
The discipline that makes this work is a hard line between analysis and action. The AI is allowed to read everything and recommend anything; it is not allowed to act on the process. That boundary is what keeps the deployment inside the safety envelope and makes the MOC conversation manageable. It also keeps accountability clean: the person who acts on the recommendation owns the decision, exactly as they do today.
What a working example looks like
Kaysee, Cortland's reliability platform, follows this pattern on critical rotating equipment. It reads compressor, pump, and turbine telemetry from the systems already in place, surfaces developing problems to the engineer who owns the call, and never touches the control loop. The historian stays the historian; the DCS stays the DCS. The plant gains an advisory layer that makes them answerable.
The fastest, safest way into plant AI isn't to take over the control room. It's to make the data the control room already has answer the questions the engineers are already asking.