産業AIを形作る現場の条件――GE VernovaのJeff Wittが語る

Welcome, everyone, to the Emerge AI in Business podcast. Today's guest is Jeff Witt's Global Process Digital Authority at GE Vernova. Jeff discusses why industrial AI can't be measured the same way enterprise AI is, since the margin for error on a factory floor involves physical safety, not just redone work. He explains how context determines whether frontline workers actually trust and adopt an AI tool and lays out a practical framework for proving a process is ready to automate before scaling it across multiple sites. Today's episode is sponsored by Poker. A quick note for our audience, but the views and opinions expressed by Jeff on today's program are his own and do not respect those of GE Renova or its leadership. In this episode, we cover what separates industrial AI from enterprise AI. To go deeper on this topic and learn how to evaluate AI vendors by assessing leadership expertise

and why funding benchmarks can signal product maturity and stability, download our free PDF report, Five Ways to Select the Right AI Vendor at emerge.com slash AIV1. That's emerj.com slash AIV1 to download your copy. Now the conversation with Jeff. Jeff, welcome back to the Emerge AI in Business Studio. Thank you for having me. I'm glad to be here. Great conversation waiting for us. So if our listeners remember correctly, the last time we had you on the show, we spoke about computer vision, and this time a very interesting conversation about how we should be thinking about AI in the industrial environment. So when I have conversations about AI outside of this environment, it feels as though most people think that AI is AI,

regardless of which direction you're pointing it. So that's where I want to start today, is what is genuinely different about applying AI in an industrial environment versus an office one? I think there's two components to this that make it significantly different. So first would be consequences. So if I'm using generative AI or an agent to analyze some sales figures, the worst possible case is that I have to redo it. I have to correct the data. I have to redo some work. If I'm using AI on the shop floor to make recommendations, it potentially could tear down a piece of equipment. If I do something incorrectly, worse than that, I could maybe put a technician right next to an operating piece of equipment, a spindle that's still running. So there's a safety component. There's a equipment safety impact component that just doesn't exist in the enterprise AI environment. So it is a little bit different. It changes what good looks like, right? So for enterprise AI, you might expect

80% accuracy and be perfectly happy with that. In the OT space, I can't have a fork truck not hits a pedestrian 80% of the time. That's not good enough, right? So the way it's judged is completely different. So that just adds a whole different layer of rigor to an AI project on the shop floor. Absolutely. So what's the one thing that a software team building AI capabilities or AI programs for offices always gets wrong when they try to apply to a factory? Is there a general assumption? I think there's constraints on the architecture that we have in the OT space that are definitely not present in the SaaS subscription space. So I can securely access data in my cloud environments. All I need is the proper SSO login credentials, and I can get access to that data to run a model or to analyze the data. There's a rather complex architecture on

the shop floor that some of our levels of equipment are air-gapped intentionally, right, so that people can't maliciously get into our equipment. So running models in the cloud against systems on the shop floor that might be a control loop that are actually controlling the equipment or controlling a process, we are not going to round trip something to the cloud to analyze or run a model against that sort of data. So just the architecture itself has to be different in the OT space than it is in the enterprise space. Bringing the model down to the edge and running it locally is usually the solve for that. So instead of round tripping to the cloud we run edge compute locally to bring that model down a little bit closer to the edge where we have more control over the security That makes a lot of sense

And I think it's basically the same as building a house for domestic use versus building a big commercial building. It's the same. Both AI, but you have to treat it differently. Both houses or both buildings, but the safety requirements will also be completely different. So this is the part where I try to get more practical, and I'm going to read from my notes here because a very interesting concept that jumped out to me was that the PwC and Manufacturing Institute study from the spring found that only about a quarter of frontline workers come into AI feeling excited about it. Most of them are very skeptical about it. And over half of leaders aren't confident that their frontline teams are ready to lead through this change and adoption. So this isn't really a technology question anymore. It's a, what does this tool need to do for the person on the floor for them to trust it and not question it anymore? And given the work that you've done all over in asset health and predictive maintenance, what AI capabilities would meaningfully support the unique workflow environments in industrial settings?

I agree with that statistic. What I've observed is that when a frontline leader is making a decision today, they have a lot of context. They know the machine. They know the work cell. They know the people that are working on that machine. They know the history. They know what's happened over the last couple of weeks. And they know what's planned, what's upcoming for that machine. So there's just this huge ecosystem of context that a frontline leader knows inherently just by working in the environment. That is the expectation, right? So if the AI comes in and it gives a really generic recommendation, it goes out and searches the web and it recommends, hey, when I see this signal, I suggest you generically perform this function. you have one shot with the AI recommendations to make that first impression.

And I think what the mistake people make is they start with this general feedback, general chatbot type feedback loop, and it's not going to work. You have to start with that very specific operating context of that pump, that machine, those operators, this product that we're running on that machine. And we have all that data. But if you keep it superficial, they're not going to see the value in the tool. And so my personal opinion, and I've seen it, is you have one shot at this. And if you put an AI recommendation engine or a coach or something out there, and it gives a super generic Google-type response to your first question, you've lost credibility, right? It has to be very, very specific. I can see that. But now also at the same time, the problem of data, so you just say we do have the data. There's no reason for it to have to go and give us generic answers. But how do we capture that knowledge? How do we capture the knowledge of someone who's run a line for over 20 years without turning it into a major documentation project?

There's two things, right? So for a lot of our equipment, we might have history going back 10 years. So there is a lot of context. Some of it's structured data. A lot of it is unstructured, notes, documentation, things like that, but that the AI can access that quickly. But the important thing to know about putting a new workflow on the shop floor that's AI enabled is that we have this built-in feedback loop. So right now, if the AI makes a recommendation, a human in the loop is going to evaluate that recommendation. They're going to plan the work. They're going to schedule the work. They're going to resource it. They're going to make sure we have the right tools, the right parts, the right documentation. It's a safe environment. They're going to do everything they normally do, but the recommendation came from the AI. But all of that execution part is now a feedback loop to the model. So it can now take that feedback and make itself even better.

That's a component of feedback to the AI model, the training of that model, that is very, very valuable in improving it over time. So if we can go into it and it can give a lot of context and it knows the history of the asset, and then we augment it with this feedback loop to improve it over time, now you've got something. So basically we're using existing data to start everything up, give it enough context to start, and then continuously just develop it with up-to-date data as we go, and it becomes this super capability-driven machine. Is that the right way to think about it? Yeah it as simple as a thumbs or a thumbs interface on the recommendation I really like that interface because I give the AI recommendation a thumbs up I liked this This was good This was valuable Show me more recommendations like this or thumbs down. This was not helpful to me. Please don't show me more like this. This is just noise in the system, right? But having all of the context

and then having this built-in feedback loop to the model so that it can improve over time, very valuable, especially as you start to scale them out. Yeah, I like that. Is there a capability that everyone assumes is essential that you would think it ranks pretty low and how essential it is? Something that you think, oh, you know, we can actually live without this. It's nice to have, but it's not worth the hype. So where AI started gaining traction was the chatbot. And honestly, on the shop floor, that interface does not work. Do not have time to stop, go to a terminal, pull up our phone, start asking questions, right? It needs to know already that I'm standing in this work cell in front of this machine and give me some context and provide those recommendations to facilitate my work to make me more efficient, more safe, more productive, and use that context as an

interface rather than just an empty box waiting for a question. That interface does not work on the OT environment. Yeah, that's interesting. I think you're probably the third person this month to say that, to say, you know, we really passed the era of chatbots. And to be quite frank, it never was that useful. And it sounds like you're saying exactly the same. So definitely out of that era and not using that anymore. So that's interesting. I like to think about this environment as something, and AI in general this year specifically, it's not something that new anymore. Everyone is trying to adopt some kind of AI. And this is where I have a nice little stat again. So what we're seeing is that the number going around right now is that AI is running somewhere on the floor at least at 42% of all manufacturers. But only 12% have gotten it past a single use case. So it's not that the companies are stuck at 0.0, but they're getting stuck at 0.1.

So they've started and then that's it. So for leaders in the industrial environments that are at that threshold of transformation, what practical steps should they consider first? I would agree with that statistic. It is difficult to scale. And I think it's because we are not starting in the right frame of mind. So I've heard the team here say something recently that's kind of stuck with me, and it's we have to earn the right to automate. The way we are embracing that is, has the process been fully documented? Has the process been optimized? Have we applied lean principles to this and taken all of the waste out of this process already? If yes, then great. This is a good candidate of a use case that we could probably augment with AI, automate some portion of it, and then potentially scale. Right. So that's what we're scouting for is those processes that are mature, well-documented, efficient processes that are used in all 35 sites.

And so that would be an ideal candidate that we could then take and apply AI tools to that, elevate that process, automate it, and begin to scale that out to multiple sites. because you're right, we don't want to do a single use case, no matter how successful. The value is not in being successful in one site. It's in eliminating that waste across all 35 sites that are running that same product. And that's exactly something that I also wanted to ask you today. If we're thinking about good documentation, optimization, everything that you basically mentioned, what does that actually look like? I think it's something that we hear the phrases or we have to make sure that we have good data foundations. But what does good data foundations really mean on a factory floor in plain terms? How do we know it's good enough to start with? We all know we have lots of data. What we generally are lacking is the context and the information model that lays on top of the data.

So we do believe that that is very important. We're investing in it. It doesn't really exist today. It hasn't been something that we've needed. But to facilitate and accelerate AI projects, we do see the value in building out our context layer, adding context to all of our data sources, building a unified namespace so that all data flows through this one template And building an asset hierarchy that makes sense to even the newest of employees So if they walk in and they get access to our data set, they can follow the structure. Okay, this division, this plant, this line, this work cell, this pump, it makes sense and they can follow it. And then once they get down to that layer of that pump, it has a pump unified namespace template laid on top of it.

So no matter what pump you pull up in the organization, you're going to have the same data structured the same way. And that's an investment, right? It's an investment to do that and build that out, especially for the first couple of use cases. You might look at it and say, well, we don't have the time to do all of that. But that basic structure and context work pays dividends when you look at the third use case, the 50th use case, the 100th use case, because it's a build it once and then just maintain it, improve it, and continue to use it forever. So every AI use case after you begin this organization of the data just is accelerated and much quicker to implement. Yeah, I like that because it is a scary concept thinking about what makes the data ready and what makes for a good foundation. But if you break it down to that, it becomes more practical and becomes more doable. It takes the scary part out of it. It's been a great conversation.

I'm looking at my notes and I don't even know where to start. I think for me, the biggest takeaway is just starting with getting the difference between industrial AI and enterprise AI. And something that I think will really stick with the audience as well is that you said the biggest difference is consequence. And with industrial AI, it's kind of a life or death situation, which is not necessarily more important or more serious, but the human is involved and the human safety is involved. It changes the equation a little bit. And when we're thinking about the AI capabilities that will be meaningful in this area, context from the worker is important and not relying on generic information. Because like we mentioned, generic information is not going to build trust. In fact, it's going to just – I think that's where we see AI pilot stall or just something being stuck in pilot mode because no one is really using it and it's not going to scale because no one trusts it because it's just something that they can also get on Google, nothing special on them. Starting in the right state of mind, I like that. I like thinking about it and more than just, oh, this sounds like a good plan.

Let's make sure that our plan has a plan. That's where we ended off today. Anything else that you would like to leave the audience with? I, I, I'm a big believer in standards and followings, existing standards. And I look internally first to see if we have that standard. If we don't have it internally, then I look to my network or to the industry and see if that problem has already been solved and documented somewhere out there. And then if, and then we take those standards and we build with the people doing the work, right? And that's how we get to success. So don't reinvent the wheel. It's a new tool, but it's not a new process. We're just eliminating waste in the processes. We're just trying to be more efficient. Oh, what a great final thought. Jeff, thank you so much. Thank you for your time. Thank you for coming back. I know that we have more to talk about, so we'll definitely get you back again for a third time, but it's been wonderful having this conversation with you. Thank you so much. Happy to be here.

Wrapping up today's episode, I think there are three key takeaways from our conversation with Jeff. First, the standard for good enough AI changes completely once a recommendation can affect physical safety, not just read-on-work. Second, context, not generic first-up answers, is what earns an AI tool trust from frontline workers, often within the very first interaction. And finally, scaling AI across many sites starts with earning the right to automate, documenting, optimizing, and improving a single process before expanding it further. In this episode, we covered what separates industrial AI from enterprise AI. To go deeper on this topic and learn how to evaluate AI vendors by assessing leadership expertise and why funding benchmarks can signal product maturity and stability, download our free PDF report, Five Ways to Select the Right AI Vendor at Emerge.com.aiv1.

That's emerj.com.aiv1 to download your copy. For further executive level analysis and to join our network of leaders delivering workflow impact with AI, visit Emerge.com. On behalf of the team at Emerge, we'll see you on the next episode.