“I’ve Been Fooled by This Before”, What It Really Takes to Scale Industrial AI
Leaders from Bruce Power, Hydro One, and InSource on why operational intelligence (AI) pilots stall, and what it really takes to scale them.
It’s not every day you hear someone who leads an operational intelligence practice admit he’s been fooled by AI. But that’s how Devon Keane opened our panel at InSource Innovate. Before joining InSource, he saw demos he was sure would transform the building materials and specialty chemicals companies he worked for, and learned the hard way that a great demo doesn’t guarantee a solved problem, especially when it comes to industrial AI.
That honesty set the tone for the rest of the conversation. Moderated by Zubin Devar of AVEVA, the panel brought together leaders from Bruce Power, Hydro One, and InSource, along with systems integrator Dan Purcell, to talk about what it really takes to move operational intelligence (AI) from experiments to results. Across very different industries, the same lessons kept coming up.
Start with a “hair on fire” problem
Devon admitted that some of the use cases he once thought were well planned turned out to be his own agenda, or his group’s, rather than a problem the whole organization needed solved. That matters because the hardest part of any deployment is change. Operators and frontline teams have often done things the same way for years, sometimes decades.
So he starts with two questions. What problem are we actually trying to solve? And is it a “hair on fire” problem, urgent enough to earn buy-in across the organization and the resources to see the change through? Only then does he ask about technology, whether it turns out to be AI or not.
To prioritize, Devon recommends a simple value-versus-effort matrix. Quick wins build momentum, and success stories can spread from line to line and site to site. He also urged teams to look past lagging results like downtime reduction and track leading indicators, such as how well maintenance is planned and how long repairs take. Those show you’re on the right path long before the final numbers arrive.
Look for what’s different
Sometimes the best use case hides in a pattern no one can explain. One manufacturer’s long, continuous line broke far more often on weekends than during the week. With analytics running on AVEVA Historian and AVEVA MES data, the team narrowed the cause to four or five setup parameters and had operators keep them within set limits. Progress was slow at first, but they kept refining until the plant had its first weekend ever without a line break.
For many organizations, the first steps toward AI are cleanup. Years of silos leave the same asset with different names in different systems, and governance and security have to come first. As Aaron Cole of Hydro One put it, “Building on a shaky foundation will give you shaky results.” Experienced experts spot bad data quickly, and once they lose trust, the project stalls.
The good news is that you don’t need to digitize everything at once. Start with the problem and build a roadmap from wherever your data lives today, even a spreadsheet. Open standards like MCP (Model Context Protocol) are also making it easier for AI tools to reach historians, MES, and maintenance systems instead of staying stuck in silos.
Test pilots in the real world
AI shines at recognizing context-dependent patterns and taking on manual analysis that eats up hours or days. Tasks that follow fixed rules are cheaper to solve with rules. Many pilots never scale because they run on clean, narrow data with no clear ROI target. The ones that make it are tested on messy, real-world data, with success criteria set from day one.
Strong governance may slow adoption, but it protects the data that gives a company its competitive edge. A simple test helps decide how much autonomy AI should have: what would a wrong decision cost, how serious would the consequences be, and can it be undone?
No matter how capable AI becomes, people stay accountable for what it produces. Think of a doorman: the job looks easy to automate until you see everything he really does. Much of industrial work is the same, and as AI takes on more of it, human judgment matters more, not less.
Advice for getting started
To close, Zubin asked each panelist for one piece of advice for someone starting to bring AI into live operations.
- Dan Purcell: Demystify it. Nobody ever worried that an MES would take over the world, but AI gets lumped in with the fears. Show people where it helps, because the humans aren’t going away.
- Devon Keane: Stay intellectually curious and get comfortable being uncomfortable. Walk the plant floor and listen to the pain points of people in other parts of the organization.
- Tyler Cowie: Don’t use AI as a hammer looking for a nail. Understand the problem and what it’s costing you, so you can build the business case from day one. And remember that AI isn’t free to run. Usage costs add up.
- Aaron Cole: Learn how to surf. The AI waves keep coming, from machine learning to prompt engineering to whatever is next. You can see them as dangerous, or you can line up and ride them with excitement.
- Jeff Miller: You don’t need to build the entire data foundation everywhere before you start. Begin with a use case, decide where it’s going, and let your use cases fill in a consistent roadmap over time.
The common thread will sound familiar if you heard our recent podcast with Rob Bansek and Devon. The technology is often the easier part. The real work is choosing the right problem, building data people can trust, and bringing the people who run the operation along from the start.
Don’t Get Fooled Again
The panel’s biggest lesson was that technology is often the easier part. Our latest podcast episode picks up where the panel left off, with Rob Bansek and Devon Keane on how manufacturers can get their people, process, and data ready for operational intelligence (AI).