Can AI run your plant?
Short answer: yes, if you believe the marketing. We’ve all seen the videos: lights-out factories, humanoid robots pulling off backflips. If a machine can do gymnastics and prove theorems, surely it can keep an eye on your reactor too?
Long answer (this article and our new video 😀): most of the sites we visit aren’t there yet, and getting there might be genuinely hard. That doesn’t mean giving up. There’s a lot of useful AI work available even if your plant looks nothing like a Silicon Valley demo reel.
Quick interruption before we get into our new video and article:
📖 We’re publishing this a month before our book, the IT/OT Handbook, comes out, and we need your help beating the Amazon algorithm overlords. If you pre-order (or leave a review once it lands), that genuinely is the only lever we have for visibility, and pre-ordering also gets you into our virtual lounge event.
Pre-order at itotbook.com →🙋 And really really the last call for our next IT/OT Academy which kicks off September 18 with an updated program.
Join the ITOT.Academy (new cohort in September) →
The dream, from a technology angle
Start with what people actually picture when they say “AI running the plant”: a system that sees everything, understands everything, and can take the physical action needed on the shop floor. That’s not complete fiction. Lights-out warehouses have existed for a while now, taking an order from click to picked, packed and loaded onto the right truck with barely a person in sight.
What makes that possible isn’t just AI, though. It’s a structured, often repetitive environment with automation designed in from day one, paired with a genuinely good digital twin the system can simulate and act on. Car plants (or parts of them) work the same way. And yes, the line between “AI” and “just very clever automation” is getting blurry. Vision systems, how robots get trained, cameras sorting potatoes by grade: this stuff has genuinely changed in the last five years, not just been rebranded.
However… most industrial processes don’t have that luxury.
No greenfield, no perfect digital twin, decades of brownfield reality instead. If that’s you, the good news is that current AI tools happen to be good at something else entirely: knowledge.
Where the real opportunity is: knowledge, not control
We’ve spent years turning signals into data: building historians, contextualising tags, optimising how we handle time series. What we mostly left untouched is everything that doesn’t come from a sensor: manuals, procedures, shift reports, maintenance notes, incident reports, P&IDs, SAP work orders, emails, and the things an experienced operator just knows (low pressure plus high temperature plus high rotation speed means watch out for cavitation). Before generative AI, none of that was something you could automate. You needed a person to hold it all in their head.
Take a maintenance engineer trying to work out why Pump P101 has been losing performance for three weeks. Traditionally, that means:
finding the pump
pulling up the P&ID
querying the historian for anomalies in process conditions over the past weeks
checking OEE and downtime events
digging through SAP maintenance history
finding the manufacturer’s manual
looking at the P&ID again
reading the shift reports
mentally correlating everything
Building software to automate that used to be a Herculean integration project. Now, an agent can (in theory):
read the shift reports
query the OEE database
retrieve the SAP records
look at the historian trends
find the pump manual
understand the P&ID
combine the evidence into a first-pass analysis, or help the engineer brainstorm
Will it replace the engineer? Almost certainly not, at least not beyond the obvious cases; even with all the data in hand, a proper investigation can still take a dedicated team days or weeks. But it can get them started a lot faster. And that, we think, is where industrial AI’s first real win will come from. Not the autonomous-control dream everyone’s been sold, but autonomous investigation.
From copilot to autonomous investigation
If an agent can do all of that on request, the next question is obvious: what’s stopping it from doing it continuously, on its own? Watching your data around the clock, surfacing an anomaly, drafting a maintenance notification, summarising a shift, finding similar past incidents, recommending a procedure, preparing a work order, or simply routing the right information to the right person.
That’s a far easier problem than closing the loop with autonomous control, and it doesn’t require redesigning your brownfield plant to get there.
What it takes
On the technology side: contextualisation, asset models, naming conventions, DataOps discipline, and increasingly the ontologies and knowledge graphs that let tacit engineering knowledge sit alongside your data rather than only in someone’s head. We’ve published plenty of content about that, for example our Industrial Data Platform Capability Map.
On the organisational side, it’s simpler to state and harder to do: this isn’t something IT or OT can solve alone. If both sides keep working in silos, none of this happens. We’ve written about that at length in our “What is IT/OT convergence?” explainer.
Don’t forget: Starting today is always the better option
Picture two factories, ten years from now. One still has its historian, SAP, manuals, shift reports and engineering documents sitting in their own silos. The other has agents continuously working through those same sources, surfacing anomalies, opportunities and maintenance insights as a matter of course.
You probably don’t want to be the first factory.
If you want to get past the buzzwords and actually dig into digitalising manufacturing, head to itotinsider.com for more free content, itot.academy if you want to work through this together with us, or grab the book at itotbook.com or your favourite bookseller.
Willem & David
The distinction you draw between autonomous investigation and autonomous control is exactly right, and I wish more vendors led with that framing instead of the lights-out factory reel. I have spent the last few years embedded inside food processing plants specifically trying to close the gap between what AI can genuinely do in a production environment and what the pitch decks promise. The messiness of real plant data, shift handoffs, and operator trust are problems that rarely show up in demos. Curious what you are seeing as the biggest sticking point when manufacturers try to move from a promising pilot to something operators actually rely on day to day?
The brownfield point really stands out. In most plants, the challenge isn’t a lack of data it’s that the useful context is spread across historians, manuals, reports, SAP, and operator experience. Making that knowledge usable by AI may be a much more practical first step than trying to automate control.