How Vision AI Scales Across a Manufacturing Network - with Jeff Witt

Summary

Podcast Summary

Podcast: The AI in Business Podcast
Host: Daniel Faggella
Guest: Jeff Witt, Digital Transformation Leader at a Fortune 500 Building Materials Firm
Episode Title: How Vision AI Scales Across a Manufacturing Network
Date: May 28, 2026

---

Episode Overview

In this episode, Daniel Faggella speaks with Jeff Witt about the real-world challenges and successes of scaling computer vision AI in a global manufacturing environment. The discussion moves beyond technical hurdles, focusing instead on organizational, architectural, and operational bottlenecks that determine whether AI pilots ever move to production. Jeff shares candid stories and practical lessons learned from rolling out vision AI solutions across 100+ facilities, and his perspective provides invaluable frameworks for executive leaders seeking to unlock value from vision data at scale.

---

Key Discussion Points & Insights

1. Why Most Computer Vision Pilots Stall in Manufacturing

• Statistic Confirmed: 77% of AI vision initiatives never progress past the pilot stage ([02:00], Jeff Witt).
• Root Cause: It's rarely the technology itself; rather, organizational misalignment and disconnected teams slow down or halt deployment.
• Quote:

  "The majority of our use cases are in pilot or POV status. The ones that make it out are the end to end solutions... it's more about the people and the process than it really is the technology."
  – Jeff Witt [02:52]

2. The Importance of Architectural Planning & Data Integration

• System Disconnection: Existing manufacturing camera systems are often segregated on separate networks, limiting their ability to provide integrated data and insights.
• Overcoming Bottlenecks: Early focus on architectural design—ensuring vision data connects seamlessly to enterprise systems—lays the groundwork for repeatable, scalable solutions.
• Quote:

  "When we want to take that manufacturing level data from the vision systems and move it and combine it with other process data... that's a barrier that needs to be planned and handled early on in the architecture."
  – Jeff Witt [03:52]

3. Iterative Approach: Start Small, Expand as Value Emerges

• Evolution from Augmentation to Infrastructure: Projects often begin by layering computer vision over existing infrastructure. As success becomes apparent, demand for more cameras and capabilities leads to broader infrastructure upgrades.
• Lessons Learned: Start by meeting plants where they are; expand as real operational value is demonstrated.
• Quote:

  "It very quickly evolved from adding value to existing infrastructure to now it’s an infrastructure project in itself... expanding and upgrading these existing camera systems."
  – Jeff Witt [05:18]

4. Real-World Complexity: Maintenance and Sensor Challenges

• Environmental Realities: Industrial floors are messy—cameras get bumped, dirty, or foggy; hardware needs ongoing maintenance and monitoring.
• Preventative Strategies: Life cycle management, preventive maintenance, and programmatic alerting for sensor misalignment are crucial.
• Quote:

  "As with any sensor on the shop floor... sensors get bumped, they get wet, they get dirty... preventive maintenance needs to be part of the program."
  – Jeff Witt [08:30]

5. Criticality Tiers: Not All Applications Are Equal

• Use Case Hierarchies: Some applications (e.g., safety proximity alerts, chemical identification) carry higher risk and require stricter controls.
• Human-in-the-Loop: For high-stakes scenarios, humans currently validate AI alerts before any automated action is taken.
• Quote:

  "There's different levels of criticality... if you're just monitoring for defects... that's lower criticality. But then there are other ones [like safety]..."
  – Jeff Witt [09:46]

6. Data Integration as the Real Enabler

• Unified Modeling: Creating reusable, integratable data models for vision outputs allows broad analytics and action across the manufacturing ecosystem.
• Stack Synergy: Timing the rollout of computer vision with broader data infrastructure upgrades multiplies impact.
• Quote:

  "I modeled the vision data so that it's reusable models that can be applied to virtually anything... That allows us to easily send it... to the data lake and then use our BI stack to do the analytics."
  – Jeff Witt [11:33]

7. "Good Enough" Threshold – Avoid Perfection Paralysis

• Reality Check: Immediate value can be achieved; waiting for perfect, clean data or models delays ROI.
• Rapid Prototyping: AI models can frequently deliver useful insights within hours or days, not months.
• Quote:

  "Our experience is we almost immediately receive value... of course it can be perfected over time, but very quickly they’re able to turn this stuff around for us."
  – Jeff Witt [13:12]

8. The Power of Human Ownership & Change Management

• Business-led Deployment: Fast expansion occurs when operations teams, not IT, own use cases, model training, and deployments.
• Visual Storytelling: Computer vision’s inherent transparency (video evidence) helps drive internal advocacy and change management.
• Quote:

  "We found a lot of acceleration, a lot of excitement... as soon as we moved it from an IT centric product to this is business led now... that ownership has really allowed our plants to accelerate and expand use cases."
  – Jeff Witt [18:40]

---

Memorable Quotes

• On Perfection vs. Progress:

  "If you wait for completely clean data, that's a recipe for never deploying."
  – Daniel Faggella [14:29]

• On Organizational Buy-in:

  "It's a platform, not a point solution. We picked our tools specifically for that so that we can pretty much build any vision system or application we want."
  – Jeff Witt [20:38]

• On the Power of Vision Data:

  "It’s a visual medium and so it’s very easy for us to tell and show this story of the value of the solutions... I can pull the video out of the system and share with people and tell stories."
  – Jeff Witt [16:34]

---

Timestamps for Key Segments

• Why Vision Pilots Stall – Organizational vs. Technical Challenges: [02:00–03:40]
• Architectural Bottlenecks and Data Integration: [03:52–04:57]
• From Proof-of-Value to Infrastructure Upgrade: [05:18–06:49]
• Sensor and Maintenance Realities: [08:30–09:29]
• Criticality Tiers in Applications: [09:46–10:42]
• Unified Data Models and Integration: [11:33–12:49]
• Defining "Good Enough" and Skipping Perfection: [13:12–14:51]
• Human-in-the-Loop & Change Management Strategies: [14:51–16:34]
• Transferring Ownership from IT to Business Units: [18:40–19:53]
• Recap and Major Takeaways: [21:32–End]

---

Actionable Takeaways

• Early architectural alignment between vision data and existing enterprise systems is vital for scale and repeatability.
• Start with what you have, demonstrate clear value early, and invest in infrastructure upgrades as justified by initial wins.
• Empower business users (not just IT) to define, deploy, and expand use cases—this accelerates adoption.
• Perfection isn’t required for value; focus on achieving “good enough” and iteratively improve.
• Make use of the visual nature of computer vision AI for internal storytelling and buy-in.
• Treat vision AI as a platform, not a one-off project, ensuring flexibility for future use cases and organizational needs.

---

Essential Advice for Leaders

"If we have a leader listening and they have a vision AI pilot that has been stalled for the last six months, the first thing they should do Monday morning is engage as directly as you can and put the tools in the hands of the people in the plants...ownership of IT has really allowed our plants to accelerate and expand use cases."
– Jeff Witt [18:40]

---

Summary brought to you by The AI in Business Podcast team – for insights where AI meets real-world business impact.

Transcript

[A]

Welcome everyone to the Emerge AI in Business podcast. Today's guest is Jeff Witt, a digital transformation leader at a Fortune 500 global leader in building materials and fiberglass composites. Jeff discusses why the majority of computer vision pilots in manufacturing stall before reaching production and why the bottleneck is consist organizational rather than technical. The conversation addresses the architectural decisions that allow vision data to integrate with existing manufacturing and enterprise systems, the role of business led ownerships in accelerating deployment across sites, and the practical threshold at which a model is production ready without being perfect. Today's episode is sponsored by roboflow. Just a quick note for our audience that the views expressed by Jeff Witt are his own. According to Edison Research, 79% of Americans age 12 and older listen to online audio monthly. That's an estimated 228 million people. For busy executive audiences, podcasts offer a rare opportunity to capture 20 or more minutes of the attention with VP and senior ranking leaders in America's largest enterprises. Emerge reaches 1 million listeners every year. To learn more about how recipes drive pipeline for other AI brands, download our media kit at emerge.com add one that's E M E R- dot com ad number one now the conversation with Jeff Wood. Jeff, welcome to Emerge's AI in Business podcast. Today.

[B]

It's my pleasure to be here.

[A]

I always start off saying I'm excited to have this conversation, but this is one that I'm really excited to have because computer vision is not not something that we get to talk about every day. I think there's still a lot of questions and a lot of confusion around it, but I want to start with a failure rate or the failure rate because it's quite striking and not in a good way. There's a widely cited figure circulating that around 77% of AI vision implementations in manufacturing specifically never go behind the pilot phase. And this is mostly not because the technology didn't work. It's because they were treated as standard software projects instead of operational transformations. Been inside large scale manufacturing transformation for long enough to basically have seen this pattern play out over and over again. What's actually happening when a computer vision project stalls?

[B]

I can validate that statistic. I would say we have the majority of our use cases are in pilot or POV status. The ones that make it out are the end to end solutions. Like you talked about, it's more about the people and the process than it really is. The technology, the AI, the computer vision, picking up the signals, generating the alerts, all of that is rather simple work. It's integrating it into day to day operations. So it's relying on some of our other pillars that we work with, workforce enablement, our TPM training teams to really ingrain the new processes into operations, into daily management of how they're doing the work on the shop floor.

[A]

It sounds like we do struggle with disconnected systems, disconnected teams, disconnected departments in this idea. At what point in a project does that disconnect usually surface? Is it during discovery deployment or somewhere after we go live?

[B]

Disconnect is partly an architecture design problem in that the camera systems themselves that we have installed in our install base are diverse. They're usually on the manufacturing IT network and they're separated from our other data pipelines and BI systems. So when we want to take that manufacturing level data from the vision systems and move it and combine it with other process data, that's more in our enterprise systems, that's a barrier that needs to be planned and evaluated and handled early on in the architecture. So that's a big barrier that we faced was getting that end to end integration. Now we have it now it's repeatable, it's scalable, something that we can deploy remotely. We don't have to go on site to facilitate. So as we continue to scale this across 100 sites, it's not something that needs to be so unique to every individual plant's vision systems.

[A]

Well, that's interesting. Once you have that recipe that works, it's integratable everywhere. And saying that you've now reached that point where you can go beyond pilot and it's actually working and the transformation has happened. What is a question that you didn't ask at the beginning of this, that you now think, oh, if I had to go back, I'll start with asking these questions.

[B]

So our approach in the beginning was to meet the facilities where they were. So if most of our facilities had some sort of process camera system in place, they were in various levels of maturity with those systems, but most of them had infrastructure already in place. My goal was to take the computer vision layer that on top of the existing infrastructure to make it more valuable to unlock more value out of that existing infrastructure. So one of the questions that immediately comes up as soon as we start deploying this and people see and get excited about the possibilities is great, now I need another camera. Now I need another location. Now I want to monitor this. So it very quickly evolved from adding value to existing infrastructure to now it's an infrastructure project in itself because now we're expanding and upgrading these existing camera systems to make them usable in other use cases in other areas of the facilities.

[A]

That sounds like a good problem to have to say, okay, well this is working so well that we should be doing this and building onto it and adding to it. What is the problem that you run into though? Is there a sensor placement things that we need to take into consideration, environmental things that we need to take? Or what are there certain problems that you ran into? And you're like, okay, so this is something that we now need to take into consideration and very cautiously think about before we continue this build.

[B]

Yeah, it quickly becomes very specialized use cases, so we can implement using the existing process cameras. But as soon as we want to go and start looking deep into the process or deep into quality defects, or we want to start evaluating really closely safety, proximity alerting, for example, that requires a high, higher level of infrastructure, more bandwidth, higher resolution cameras, more capacity on the manufacturing switches to pipe all this video around the plant. So that adds to the complexity of the projects. We try to always start with the lowest level, right. And just meet the plants where they're at and get something up and running. And then once it's proven they're receiving the value, then we can use that as a basis to kind of cost justify the further improvement, investment in infrastructure.

[A]

That makes sense. I like how you said start where you're at. And I think the biggest key there is just actually starting, not trying to plan this out outside of the plant and try to perfect it outside and then try to go and take this and implement this whole idea that should be working perfectly. That's not going to start with what you have and just start. I know that with Vision AI specifically, we have a lot of listeners, executive listeners, thinking that, you know what, a year or two ago we looked into this and yes, it worked beautifully in a controlled lab, but manufacturing floors are messy, lighting changes, the product variants will multiply. So is this still a continuous problem? Has this been solved?

[B]

As with any sensor on the shop floor, there's risk and sensors get bumped, they get wet, they get dirty, camera lenses become foggy over time. All of that just needs to be built into the entire scope of, of the project. Right. Once you put in this new infrastructure or use their existing infrastructure, there's a life cycle management, there's a preventive maintenance component to the infrastructure that you just put in. So it needs to be part of the program. It can't be an afterthought. Cameras get bumped all the time. And the machine learning, we can monitor for that and say, hey, camera's out. Of position. Now we can trigger an alert to maintenance to go redirect that camera back where it's supposed to be. But we can take care of some of that programmatically in the system. But other things are just good old fashioned boots on the ground.

[A]

Makes me think of buying a car. Knowing that when you buy a vehicle that it's going to have to have its annual services done. And then knowing that an engine light can come on or you might bump it on something at the side of the road and the sensor can. It's not a weird or terrible idea, it's just. It's actually a good thing.

[B]

There's different levels of criticality on our applications, right. So that has to be taken into account too. If we have a safety algorithm running, that's pretty important to us, right? Safety's number one, always has been. And we have other use cases where we're monitoring chemical bin labels, right? We're checking does this chemical that you just loaded into this bin, does the label match what the bill of materials says we're supposed to be running right now? That also could be a safety issue. It could be a process issue. If we load the wrong chemical into the process, we could potentially dump an entire day's worth of product because that chemical polluted the system. So there's different levels of criticality, right. If you're just monitoring for defects or you're evaluating a stream for production metrics or something like that, that's lower criticality. But then there's other ones. So that has to be taken into account as well.

[A]

This is blowing my mind because if you go to things that specific, saying that product labels, this all has to be based on data. And I think one of the greatest points of friction in manufacturing is the transition from getting something from the lab to commercial production. And it always comes down that friction is, I would say, almost always caused by fragmented data and fragmented systems. And we hear a lot of manufacturers say that their stack is fragmented and their data is messy. It sounds like things like that would be really crucial for this to really work. And it almost feels like blaming it on data is more of a description and not necessarily a diagnosis. So for you, having actually built out asset health and computer vision programs, what's the specific thing that was missing that made deployment hard and kept that pilot in pilot mode?

[B]

Data. You're right, data is largely the limiting factor. Right. So we don't want to introduce another application or another data silo. We would rather try to use our existing infrastructure. So we have a relatively New manufacturing it ot data stack that goes all the way from the shop floor all the way to enterprise level data. And one of the good things about the CBML data is because it was being introduced at the same time as this new manufacturing data stack was coming online. I modeled the vision data so that it's reusable models that can be applied to virtually anything that we deploy. So that allows us to easily send it along with the rest of the manufacturing data to the data lake and then use our BI stack to do the analytics or trigger the actions that are required for the reporting and the advanced analytics. So taking the vision data, combining it with the quality data, the manufacturing data, the ERP data, all of that in the data lake and then exposing that to the analysts that need that level of detail, it's been really effective. And we were just lucky, timing wise to be able to get our data built into that new stack.

[A]

Lucky or I'm sure that it was a lot of trial and error built in there as well. And thinking about that, if someone's listening to this and like, okay, we're not there yet, but what Jeff is saying is telling me what direction we should be going in. What does good enough look like for a first deployment? At what point do you stop optimizing the model and start getting into production?

[B]

Almost immediately. Our forward engineers from the vendors, they're always coaching us that it's three to six months, months to refine and perfect the model. That's not our experience at all. Our experience is we almost immediately receive value, right? Some level of value. Of course it can be perfected over time. But we will bring a forward engineer on site, they will download some video. By the time I'm back to my office, they will have a video sample of how the AI might look at that particular stream. So very quickly they're able to turn this stuff around for us. And that gives us the visibility that we're looking for. Right. We have the process cameras. The process cameras are already pointing at important parts of the process that we want to monitor. It's just five years ago, I had to have a person, an operator sitting in a control room monitoring those screens. Now I can have the AI do it, and it just pops up when an anomaly happens. So that's the change. And so having the AI look for those anomalies in real time and then alert us in near real time, that's the value that we get out of it. It doesn't have to be perfect. None of our models are perfect.

[A]

I love that you said that because I think a lot of the pushback that we're getting on this is skeptics saying that we have to wait for completely clean data, clean data environment to get to perfection before we even start deploying AI. And I feel that that's going to be a recipe for never deploying. At some point you have to just, just ship with the data you have. Has that been your experience?

[B]

I agree with you. The human in the loop concept is something that we really believe in, right? We are not going to take an AI signal from the computer vision system and control the line with that directly. We could, we can in the future. We probably will. Right now we have a human in the loop, right? So there's whether it's an alarm on the floor or it's an email to a user or it shows up on a report somewhere, a human will evaluate that before taking action. Eventually, like the, like I mentioned the chemical bin label issue, if it did determine that that's the wrong chemical for the process that's being run at that time, we could stop that pump. We could allow that pump to not be turned on through a control loop. So that can potentially save a million dollars right there. Without a human in the loop loop, we're not quite there yet. People need to have confidence in the systems, they need to have confidence in the AI before we take that step. I think eventually we'll get there, but right now we're not there.

[A]

It's great to just kind of see the future through your eyes. We're heading towards the closing part of our conversation and I think this is something that is evident in almost every AI conversation we have. Regardless of what specific AI we're focusing on, is that a lot of successful deployment is not based on the technology alone, but rather successful change management, management, internal training, a well planned integration, dedicated implementation teams, even these things are all more valuable than the technology's capabilities. In your opinion, what's the first structural move that actually determines whether anything gets done?

[B]

So I think in this particular space there's a built in benefit to the technology in that it's visual. You can see the alerts, you can see the AI processing the images, you can see it's making decisions. An operator is too close to a fork truck as it travels down the aisle of materials. It's a visual medium and so it's very easy for us to tell and show this story of the value of the solutions. So I think that's, that's a trick that I used early on was to use that visual medium. And I embed videos into the PowerPoints and the project updates. And if we had an event that was particularly interesting in a plant, I can capture that video and share that. So that's a benefit that these types of vision projects have, is that it's very easy for people to understand. It's also very easy for them to see the technology working. So while a lot of tech might be black box servers racks, nobody, you know, a handful of people log in to see it. This is something that I can pull the video out of the system and share with people and tell stories. And it's really powerful and it's very effective. And if you use it as part of your change management process, you can really get a lot of adoption with it.

[A]

That's very. I haven't really thought about it this way, but it makes sense. It's really in your face. You cannot push back on it. Well, you can. I'm sure you can, but it's going to be a bit more tricky because like you said, it's not that black box. You don't have to figure out or try and wonder where the information came from. You gave it the context and you can see what it's doing with it, which is a very unique point of view to have. Jeff, I want you to leave our listeners with one last idea. If we have a leader listening to this right now and they have a vision AI pilot that has been stalled for the last six months, what is the first thing that they should do Monday morning when they get to the office?

[B]

I would engage as directly as you can and put the tools in the hands of the people in the plants. In doing this from an outside in, probably a little difficult, a little rough, but we found a lot of acceleration, we found a lot of excitement and a lot of speed in deployments. As soon as we moved it from an IT centric product to this is business led now, and they can define their own use cases, they can deploy the technology, they can build the models, they can do it all themselves with the help of our vendors, but it doesn't necessarily have to be involved. And so that ownership of IT has really allowed our plants to accelerate and expand use cases. If it's working in this plant and I have a similar line in 14 other locations, I can now take this what is working well in New York, and I can take that and I can scale it across the country to 14 other locations relatively easily. So taking it out of the IT hands and just letting the business run with the platform now, that it's relatively mature.

[A]

I like that. I think from our conversation it's very evident that the technology is not the bottleneck. Firstly, our disconnected systems. That's going to be our first stop is if we're already in a state where we've been hanging onto a pilot for way too long, that's where we should be looking first is the disconnected systems. The disconnected teams that are not working together know that Vision AI works. I. You've made it very practical and very true that it's in your face, you can see it work and you can also see when it doesn't work. And I'm sure that it's also easy to then pinpoint exactly where the problem is coming in. Use the data that you have, use the infrastructure that you have and build on it as you see the requirements come in or as you see the need build. Is that that the essence of our conversation today?

[B]

Agree with that. It's a platform, not a point solution. Right. And that's we picked our tools specifically for that so that we can pretty much build any vision system or application we want. It's not just safety, it's not just quality, it's not just looking at one particular product. We can train it to do just about anything. So it's a platform, it's a tool, it's in our arsenal. It's not right for every solution, but it does solve a lot of problems.

[A]

That's great. And it's great to see how this is evolving and it's important to keep ourselves in the loop of that. Jay, thank you so much for sharing your insight today. It's been loving having this conversation with you.

[B]

Great. Happy to do this. Thank you.

[A]

Foreign. Let's look at our three key takeaways from the conversation with Jeff. First, the technology is rarely the bottleneck in a stalled computer vision program. The failure almost always lives in how vision data is connected to existing operational systems and whether the right ownership structure is in place before deployment begins. Second, building a reusable data model for vision data from the start, one that integrates with the broader manufacturing data stack rather than creating another silo. That is what is separating programs that scale from those that remain point solutions. And finally, shifting ownership from the vision platform from IT to the business units closest to the floor is what generates deployment speed. Once plants can define their own use cases and train their own models, adoption accelerates without requiring IT involvement at every site. According to Edison Research, 79% of Americans aged 12 and older listen to Online Audio Monthly. That is an estimated 228 million people. For the busy executive audiences, podcasts offer a rare opportunity to capture 20 plus minutes of attention with VP and higher ranking leaders in America's largest enterprises. Emerge reaches 1 million listeners. To learn more about how recipes drive pipeline for other AI brands, download our media kit at emerge.com addone that's emerj.com ad1 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. It.