Engineers, process managers, and control room staff joined Atlas for a session on where AI is actually showing up in control rooms and on the plant floor, and what has to be true before it can be trusted with an operational decision. Rather than presenting AI as a finished capability, the Atlas team used four live polls to ask attendees directly how they use AI today, where operational knowledge really resides, how they troubleshoot, and what is holding adoption back. The answers, summarized below, describe plants where the technology is arriving faster than the trust and infrastructure needed to put it to work on the floor.
The recording of the lunch and learn can be viewed here!
Poll Results
How Members Are Using AI Today
| Application | Share |
| --- | --- |
| Administrative tasks | 71% |
| Learning and upskilling | 57% |
| Development tasks | 52% |
| Operational support | 24% |
| Traditional ML (preventative maintenance, soft sensors) | 0% |
Where Operational Knowledge Lives
| Source | Share |
| --- | --- |
| Experienced personnel | 62% |
| A mix of sources | 43% |
| Formal documentation | 29% |
| Vendor documentation and support | 19% |
| Disconnected systems | 19% |
Where People Look First When Troubleshooting
| First stop | Share |
| --- | --- |
| Historical trends | 53% |
| Subject matter experts | 41% |
| Alarms | 35% |
| SOPs | 18% |
Barriers to AI Adoption
| Barrier | Share |
| --- | --- |
| Not knowing where to start | 53% |
| Data security concerns | 53% |
| Cost | 47% |
| Trust in the technology | 41% |
> Across all four polls, the same thread kept showing up. The knowledge that actually runs these plants lives in people and historical trends, not in documentation or connected systems. Before AI can help on the floor, it has to be built on top of that reality, not around it.
What It Takes to Get This Right
An engineering-grade AI assistant built for this pattern looks different from a general chatbot. It has to operate inside the control environment, connecting APC, DCS, historian, and engineering knowledge in real time, rather than asking an engineer to correlate all of it by hand across a set of standalone screens.
Where a single similarity search hands an agent a stack of loosely related material (a P&ID, an OEM manual paragraph, a shift log) and lets it answer with whatever looks closest, a more careful design plans targeted subagents around the actual question being asked. Each subagent reads one source and reports back a fact, not a chunk, so the main agent only ever sees what is relevant.
> Operator asks: “Column 4 overhead pressure has been climbing all shift. What do I check first?”
>
> A P&ID subagent confirms the overhead valve line-up is normal. An OEM manual subagent surfaces the threshold at which reboiler pressure drop indicates a fouled trap. A historian subagent finds the log entry that shows the trap fouled two days ago. The system reports the fouled steam trap as the first thing to check, and states what it already ruled out, the way an experienced engineer would.
That architecture answers the adoption barriers attendees raised directly. Cost and data security concerns point toward tools that run on hardware the plant already owns, inside the OT perimeter, with inference on site and proprietary data never leaving the facility. Not knowing where to start and trust in the technology point toward tools where every answer is sourced back to the P&ID, log, or manual it came from, rather than delivered as an unexplained recommendation.
DCS / APC Integration · CLPM & Diagnostics · Cross-System Context · OT-Native · On-Premises · Multilingual
Takeaways
1. Language models can extend institutional knowledge and shorten the distance between a problem and a decision, but only when they are grounded in the plant's own data.
2. An AI system operating in a control room cannot hallucinate. It has to be right, or it has to say what it does not know.
3. Cloud-hosted solutions remain a nonstarter on the operational process control network under most plants' current IT philosophy.
4. An on-premise or air-gapped approach resolves the cybersecurity objection while still putting the reasoning power of a language model to work.
Keep the Conversation Going
If you're thinking through how to bring language models safely into a control room, we're glad to keep the conversation going. Reach us at atlasprx.com.
Presented By
Madhur Bedre is President and Founder of Atlas Prediction Control.
Alden Olyniec is Technical Lead of AI Engineering at Atlas Prediction Control.
