Our crew at Atlas Prediction Control spends a significant amount of time inside plants, helping teams recalibrate advanced process control that has quietly drifted out of alignment over years of operation. That work puts us in a lot of control rooms, and often the same question is raised: Where does AI actually fit here?

Rather than guessing, we recently posed that question to the engineers, process managers, and control room staff who live with it daily through four live polls.  

The first finding: AI has arrived (71% of respondents reported using AI for productivity tasks), but not where it matters most. Attendees reported heavy use for administrative work and for their own learning, far ahead of anything touching daily operations. Nobody in the room reported using machine learning for applications like predictive maintenance or soft sensors.  

The next two findings explain why. Asked where the knowledge that actually runs the plant resides, the majority (62% of the respondents) answered that it wasn't a system; it was a person. Experienced staff, not documentation, carry the operational truth of the process. When something goes wrong, the instinct isn't to open an operating procedure but to pull up the trend screen and find someone who has seen this before (53% attendees reported this as their first step).  

Every experienced system integrator (SI) who has walked a plant floor already knows this pattern. The poll data confirmed that the real knowledge base in most plants is distributed across tribal memory and historian trends, not the documents an audit would point you to.  

An AI agent trained in generic manuals and public documentation is training the wrong way. The ones that matter are proprietary, plant-specific, and often never written down at all. 

The fourth finding is the most practical one for SIs. The barriers users named weren't about doubting AI's value. They were about data-security concerns and not knowing where to start (tied at the top at 53% of the response rate), with cost and trust close behind. That's a deployment problem, not a conviction problem. It points toward a specific answer: keeping the model on hardware the plant already owns inside the OT perimeter, so proprietary data never has to leave the building to produce value.  

Cloud-hosted tools remain a hard "no" on most operational-control networks, and no amount of capability closes that objection by itself. On-premises AI agents purpose-built for a narrow set of plant-specific questions are what actually close the gap. 

The architecture of this on-prem AI agent designed for control rooms looks different from a general chatbot. Instead of one model reasoning over a pile of loosely related documents, a better design routes a question to a handful of narrow, targeted lookups—one against the P&ID, one against the OEM manual, and one against the historian—each returning a specific fact rather than a stack of text to guess from.  

When asked "Column 4 overhead pressure has been climbing all shift for unknown reasons; what should I check first?" a well-built system should answer with the fouled steam trap, state which other causes were ruled out, and show exactly where each part of that answer came from. This is in contrast with the generic information we get from the cloud-based language models, but it is, indeed, the level of cognitive thinking operational teams need.  

Two rules follow directly from what these plants actually need:  

• A system operating anywhere near a control loop CANNOT hallucinate. It has to be right, or it has to say plainly what it doesn't know.  
• It has to run on-premises or air-gapped, because that is the only architecture that resolves the cybersecurity objection without asking the plant to give up the reasoning power an AI agent provides. 

If you're evaluating this category of tool for your own plant, consider the following questions: Can it show its work, and does it have to leave your network to do it?

In the end, closing knowledge gaps leveraging AI agents begins with understanding where information on plant operations resides. This is the exact problem that Atlas is working on, and it is indeed possible with careful thought of our current landscape and our users' needs.