4 Ways AI Can Actually Speed Up Problem-Solving on the Manufacturing Floor

Transcript

[00:00:18] Intro: Welcome to the Shop Floor Top Floor Talk Show, where we have casual conversations with manufacturing pros. Each episode digs into the challenges and opportunities for improvement that steer both frontline execution and big picture progress. No fluff, just real world practical perspectives from the people driving the industry forward, from the shop floor to the top floor. Let’s get to the show.

[00:00:55] Josh Santo: Rich, thanks for being here again.

[00:00:59] Rich Nave: Excited to be back, Josh. It’s fun to get a chance to talk with you every month.

[00:01:03] Josh Santo: Right back at ya, and I wish it were more. I will say we did get some great feedback on the first, so we are looking to make this a very consistent series here. So if you like Rich, make sure to reach out and let him know. Rich is here to regale us of some insights that he’s learned through the years, and even more recently.

[00:01:22] Josh Santo: We’re talking specifically today about AI and how you can leverage AI for problem-solving and root cause analysis. So Rich, I’m gonna kick us off here. With AI, what and how can it help manufacturers navigate and perform their various problem-solving approaches?

[00:01:46] Rich Nave: Actually, Josh, maybe I’ll start with the negative. AI isn’t a singular beast that’s just gonna fix everything. And so if you have this illusion that you’re gonna just take a customer complaint, throw it at some AI — whether it’s Claude or ChatGPT or whatever — and all of a sudden it’s gonna spit out an 8D for you, you’re kidding yourself, okay?

[00:02:10] Rich Nave: AI isn’t gonna replace engineering judgment, and it’s not gonna replace some hard work. So I just wanna make sure that there’s no illusion going into this.

[00:02:19] Josh Santo: Let’s be clear. If you tell AI to take this information and give me an 8D, it will give you an 8D. It doesn’t mean it’s gonna help you solve your problem. To your point, it’s not gonna replace years of experience, the expertise that’s been built up those years of experience, et cetera.

[00:02:39] Josh Santo: So I think you’re absolutely right. There’s that gap. So you can put it in, you can get information out, but that doesn’t mean you can trust?

[00:02:46] Rich Nave: It can. But it can almost make you look stupid to your customer. And so you really wanna be a little bit careful about that. You know, I sometimes think that we have this almost overarching belief that it’s just gonna fix the problem on its own. So I wanted to start out with that — there’s still gonna be a need for engineering judgment. And you know, Josh, one of my favorite topics is always that it’s gonna need the commitment of leadership to deploy the resources to problem-solving. That’s not gonna go away. You’re not gonna take problem-solving and cut it by 90%. You’re still gonna have to deploy resources and still gonna have to work on this. However, having said that, I don’t wanna come across as the old guy sitting on my porch yelling at the kids to get off my grass, okay? There are lots of benefits, and there are lots of ways that as you integrate AI into more and more of your business, it can be used more and more in problem-solving. Because AI already knows where your scrap data is, or AI already knows where your machine uptime is, and it can go look at that and data mine that for you. So the first real area that we’ve used AI in is to improve the problem definition, okay? We have some things that we’re gathering. We’ve got the customer complaint.

[00:04:17] Rich Nave: We know generally what the customer’s complaining about, so we can bounce that off of some of our scrap data or other information in that area, and AI can give us a problem definition. Now, oftentimes we wanna still massage that a little bit, but it gets us out of sometimes that stuck spot that teams get into of “well, the problem is bad parts.” That’s not a problem definition, okay? And AI can help us do that, and can help us do it more quickly. And once again, as I said, if AI already knows the folder where your scrap data is or where your production reports are, it’s really easy for it to go into that folder and look at — well, the customer’s complaining about paint defects. Are we seeing scrap from paint defects and things like that, and at what percent are we seeing them? So that’s the first area that we look at.

[00:05:08] Josh Santo: Starting with that idea that you can’t just assume AI is going to fix things. You’re gonna have to make sure the AI has the context that is needed. And one of the things you called out is the data. Which systems house your data? Is it accessible to the solution that you wanna use?

[00:05:24] Josh Santo: If not, you’re either gonna have to find a way to make it accessible or take that manual approach. And when doing that, plus providing the other context that’s needed — and this depends on which tool you’re using, because that’s a whole other topic — so much depends on what AI tool we’re actually talking about.

[00:05:41] Josh Santo: But depending on the AI tool, it has to have the right, for lack of a better word, knowledge set, personality, expertise that you’ve set up with it so it has a frame of reference of the character that it is playing and the discipline that it needs to enact. And then you’re talking about the task that needs to be performed.

[00:06:03] Josh Santo: In this case, we’re talking about the task of problem-solving. Your job is not necessarily to define the problem for us, but to make it faster and easier and more efficient for us to identify what that problem is, and as a team agree and ratify — this is the problem that we’re tackling.

[00:06:22] Rich Nave: Yeah. And Josh, it’s really easy with AI to not just ask for a problem definition, but ask it to give you five versions of the problem definition. Then you can look at what were some of the different angles that the AI sort of looked at and came up with, and maybe pick numbers two and three and merge those together to create an effective problem definition.

[00:06:44] Josh Santo: That’s actually something that could lead us to another point to talk about, and maybe this is worth saving until we get to the end of the different ways in which AI can help with problem-solving. But when you go through that process of iteration with the AI, one of the things that’s very beneficial and powerful to do — again, depending on which tool you’re using — is to have it understand the differences that you asked it to come up with, and understand why you went with the decision, like you said, to combine two and three, and what it came up with, so that it improves over time. So the next time you’re solving a problem, it already has a better understanding of these are the types of problem statements I should be putting out there.

[00:07:29] Rich Nave: We’ll talk more about that as we go along, but the idea is that the first time you do this, it might only be 40%,

[00:07:37] Rich Nave: The second time it’s gonna be 41%. So if you do this 30 times, it’s gonna be 80% type of thing. So we wanna recognize that it’s an iterative process that builds upon itself, and that’s part of the value.

[00:07:50] Josh Santo: Absolutely.

[00:07:51] Rich Nave: So moving to the second area — and this is probably the one that I’ll say is most obvious or intuitive to people — is data analysis. We worked with a company that was making axles for a vehicle, and this was back, I wanna say maybe 2015, 2016. And at that time, they were already collecting over 1,000 data points on their manufacturing line. So you get to an end-of-line test, and you get a test failure, or it gets past your end-of-line testing, and you get a customer complaint. Now you have to write this 8D, but you’ve got 1,000 data points. How do you know which one of those data points is varying that’s related to your problem? Now, one of the things that I like to point out to people — or remind people — is that AI is really good at finding correlation. Correlation is not causation, and causation is where the engineering judgment comes in. And so just because ice cream sales and shark attacks go up in the summer, it

[00:08:57] Rich Nave: doesn’t mean that there’s causation. So it’s great for data analysis. It can give you a subset of the data to really look into deeply. But remember, just because it shows two things are moving together, it doesn’t mean that those patterns are a fact of causation.

[00:09:17] Josh Santo: Now, the AI might help you recognize that the flavor of ice cream was chum flavored,

[00:09:23] Rich Nave: Yes. There might’ve been some causation.

[00:09:26] Rich Nave: Exactly — in that case.

[00:09:29] Rich Nave: So I remind people that when you’re doing pattern recognition with AI, you’re talking about correlation, and it’s up to you to figure out what the causal relationship is.

[00:09:30] Rich Nave: Yeah, fair call-out.

[00:09:40] Josh Santo: AI can help you spot things that are both happening, but can’t really draw that line between because this thing happened, this thing was the effect, and that’s where it really takes that individual who has that experience and expertise to tie those two things together. But to your point, the ability to source, here’s all the observations, list them, and then the person can just grill them.

[00:10:06] Josh Santo: show me this, show me that — certainly helps speed up how do you get to that conclusion.

[00:10:11] Rich Nave: Absolutely. And I would guess too, Josh, keep in mind, if 10 years ago they were collecting 1,000 data points on every axle that was manufactured, they’re probably collecting 10,000 data points

[00:10:22] Josh Santo: Oh, absolutely.

[00:10:24] Rich Nave: And so it’s even more complex.

[00:10:25] Josh Santo: That’s something we talk with manufacturers about here at EASE — this idea that trends in technology have really led to this rise of big data. Big data is not a new term, but big data for the most part hasn’t been consumable data for a variety of different reasons.

[00:10:43] Josh Santo: There’s so much, it requires a specific expertise to really interact with it. You know, a certain strategy for what data are we getting from where. You talked about that idea of causation and correlation. Well, when you’ve got different data coming from different systems, and they’re labeled as different things within those systems, now you gotta figure out how to unify that data model in order to then even have that conversation about causation versus correlation.

[00:11:09] Josh Santo: A lot of factors working against manufacturers throughout the years for really being able to tap into their data, especially for plant-focused folks who don’t have the same resources that might be available at a corporate level. But AI is becoming that differentiator there, really making it possible to quickly overcome some of those hurdles and let the machine understand what all these different digital data points are saying, which ones are likely to match from other systems, and connect those together.

[00:11:44] Rich Nave: Josh, this was one of the things that EASE got right early on — EASE took LPAs, which generate a lot of data, and allowed people to create information from that data. And I can’t remember what the number is now, but there are like 52 pre-programmed charts in the EASE software, and that’s turning data into information. And a lot of that data analytics I was just talking about is actually just about how do you graph or how do you display that data? Because a page full of numbers is just data. It becomes information when it becomes graphical and visual and people can really relate to it, and AI is really good at helping to do that quickly.

[00:12:34] Josh Santo: Yeah, that’s a fair call-out. A lot of operations are still capturing a significant amount of information on paper and Excel, and it’s not in a format in which they can easily do something with it. I used to work with a plant manager — he was with a food and beverage company — and he would always say, “Data has a shelf life.”

[00:12:55] Josh Santo: You capture a data point, if you’re not doing anything with it, it’s gonna expire more or less, and you’re not really gonna get the value of having captured it to begin with. And that always kinda stuck with me. So the faster that you can capture that data point, make it digital and consumable, which is your point, it has to be consumable, and consumable depends on who needs to consume it.

[00:13:15] Josh Santo: So much of it relies on people to consume it, but now we’re entering this age where systems can start to consume data on behalf of other systems, and either output information to people or even output that information to other systems to then automate even more activities.

[00:13:32] Mid Roll – Review Section: All right, quick pause before we jump back in. Like you here on the Shop Floor Top Floor Talk Show, we believe in continuous improvement. You don’t wait for an annual audit to fix what’s broken. You make small adjustments every day to get better. And this show works the same way, but the only way to improve is if we hear from the people that are actually doing the work. So if this episode has helped in any way, maybe it sparked an idea. Saved you from a mistake or even just gave you a knowing laugh. Please do us a favor right now. Wherever you’re listening, tap the rating button or leave a quick review. It takes less time than a toolbox talk, and it helps other manufacturing leaders find conversations like this. Think of it as leaving a note for the next shift. We read every review and we use them, and we genuinely appreciate you being a part of the Shop Floor Top Floor Talk Show community. Alright, that’s all I had to say. Let’s get back to the conversation.

[00:14:37] Rich Nave: So that actually leads into the next point, Josh, which is what I label generally as access to organizational knowledge. Okay? And so at The Luminous Group, we were asked by an organization recently to write an article for them. So I sat down and I wrote the article — between 500 and 700 words on this topic.

[00:15:01] Rich Nave: And I do that for people all the time. And I kick this thing out, then I send it to my administrative assistant, who’s a really great proofreader and also has great organizational knowledge of The Luminous Group. And she says, “You know, Rich, we wrote this article about six years ago.” And I’m like, “What do you mean?” And she pulls up this article that I wrote back in 2019 — that’s basically the same article. And so one of the things that you can do with AI is put essentially all of your 8Ds somewhere and ask the AI to look and see: have we already solved this problem? Or have we solved a similar problem? And you can get that organizational information really brought to the forefront so that you’re not resolving the same problem over and over again, or trying avenues that have already been tried and failed. And so this allows you to really create an effective lesson learned database or data bank without doing a lot of work. Now, The Luminous Group has one place where our articles are stored. AI has created a summary of each one of the articles, so you can easily see the summary. And if you type in a topic, it gives you the best matches out of our database of articles, so we don’t have to rewrite them or recreate the wheel. And this idea with 8Ds is you can do the same thing and have access to that organizational knowledge.

[00:16:40] Josh Santo: What would be the power there? Here’s why I ask that question. If you’ve hooked up AI to wherever you’re storing your previous 8Ds, and you find that — oh yeah, we’ve gone through this process and we’ve solved this problem previously — doesn’t that indicate that you didn’t solve the problem?

[00:16:59] Rich Nave: It can indicate a couple of things, and that’s one we’ll get to a little bit later — when we kind of get towards the end, because that becomes a seventh discipline type of thing, a little bit about why the knowledge wasn’t spread. You can look at that 8D and see: well, this 8D says that we’re supposed to do this prevention control or this detection. Are we actually implementing it in this area, or did we only implement it in one area of our plant? So you think about a company that has plants in most of the major markets in Asia, North America, Europe, that kind of thing. They might have solved this problem in Europe, and nobody knows about it in North America.

[00:17:37] Rich Nave: And so that’s where we could really start to get that institutional knowledge spread around and used more effectively.

[00:17:44] Josh Santo: And I would imagine there’d be some elements of, “Okay, this is how we tried to solve it before. Looks like the problem returned, so we’re gonna have to explore different avenues to really make sure this problem doesn’t return.” But we just saved ourselves some cycles of going through the same process over and over again trying to achieve a different result.

[00:18:05] Rich Nave: And you know, the other thing that we’ve talked about, Josh, at times is that there’s a huge issue in things like plant maintenance — of people retiring now. And so you’re losing that institutional knowledge. But as your maintenance team is retiring, there’s a lot of people who implemented the solutions or worked on that who are just not gonna be there in the future. And so AI can step in and help preserve that institutional knowledge.

[00:18:33] Josh Santo: Yeah, one that can be then accessed by everyone.

[00:18:37] Rich Nave: Everyone across the globe.

[00:18:39] Josh Santo: Yeah, that’s a great point.

[00:18:40] Josh Santo:

[00:18:41] Rich Nave: The last one that we’ll talk about is that when we’re done doing the 8D, that’s one document, but the knowledge that we gained in doing the 8D needs to be implemented in the PFMEA. It needs to be implemented in the control plan.

[00:18:57] Rich Nave: It needs to be implemented in training. It needs to be implemented in LPAs. It may need to be implemented in standard work instructions. So once again, if the AI has access to these documents — we recently did exactly that. We fed an AI the 8D and the PFMEA for that part, and we said, “Based on this 8D, where does the PFMEA need to be updated?” And it did a great job. We didn’t really have to search through the PFMEA or anything. It was wonderful. So from there we then did a natural expansion to the control plan, okay? And we didn’t quite do it the same way for LPA, but we took the 8D and we said, “Okay, if we wanna reinforce the change in this 8D, what are five LPA questions that we could be asking to make sure that this 8D remains implemented?” And instantly we have five questions. Now, I will be very candid. Two of them were really bad questions. They were poorly structured questions. But my point is, it’s not that all five were great questions — but three of them were very usable, adequate questions, and we had them instantly.

[00:20:18] Rich Nave: We didn’t have to do much with it. We had to know enough to review them and kick out two of the questions, but we didn’t have to do much with it. So what might take a day of work to implement those four or five documents — when I was wrapping this up with this team, we did it in 20 minutes, maybe 30 minutes. And you are taking what might be six or eight hours of work and condensing that down to under an hour.

[00:20:45] Josh Santo: That’s a great call-out, and I love the way that you’re talking about the connections and how it can then be brought into the upstream or downstream with whatever it is that you’re working with. In your example of doing the 8D — okay, we’ve come to this conclusion, we’ve made these changes.

[00:21:01] Josh Santo: What are all the things that need to be updated as a result? FMEA or the PFMEA. Okay. Where? What? How? Do it on my behalf, please, and thank you. Now, same thing with the control plan. Now, same thing with the layered process audit. Oh, by the way, with the layered process audit, the questions need to be formatted like this, that, this, that,

[00:21:21] Josh Santo: ’Cause there’s all these specifics, these details, these intricacies that come with maintaining these resources and implementing them. And to the point that you raised earlier, AI is able to essentially compensate for gaps in institutional knowledge by being something that can collect that knowledge as well as collect certain aspects of expertise that may take training courses, research, years of experience to develop, and then simplify all of it so that you can get 60% of the way there in a fraction of the time.

[00:21:57] Rich Nave: Now, I will tell you, Josh, you’re a little more confident than we were. When we had it do the update to the FMEA, we also asked the AI to highlight all of the rows that it updated in yellow. Then we were able to go into the FMEA and very quickly see which rows had been updated, and did we agree with those updates and cross-check them.

[00:22:23] Rich Nave: So I would strongly encourage you… I like your prompt, but make sure that there’s a cross-check to that prompt by highlighting the changes.

[00:22:32] Josh Santo: AI is certainly still in the phase of needing to build trust because its accuracy is not 100%. So finding those ways of keeping visibility and double-checking its work. And in my experience, some of the things I’ve done is either have the AI double-check its work, or I’ve even had other AI double-check the work as well.

[00:22:56] Josh Santo: There’s all sorts of ways that you can come up with to really make sure — okay, what are you doing? ’Cause you may not be the right person to do that, or the right entity, I should say.

[00:23:07] Rich Nave: Like we started this out with, Josh — one of the things we wanna be really careful of is we don’t want to look foolish in front of our customers, okay? The most famous statement that I share with people is when I reviewed an FMEA and it said that they were going to visually inspect for microscopic cracks. That’s the kind of thing that AI might actually say. It knows that you use visual inspection for cracks, but it doesn’t necessarily put together that these are microscopic cracks in welds,

[00:23:41] Rich Nave: And by the very nature, if they’re microscopic, you can’t visually inspect for them.

[00:23:45] Rich Nave: But truly, that was on an actual FMEA — and not done by AI, by the way. That was done by humans, about eight years ago. But nonetheless, you don’t wanna look foolish in front of your customers, so make sure that you’re double-checking this work. And as I said, whether it’s by highlighting it or whatever, make sure you’ve got a human approval in the process.

[00:24:11] Josh Santo: So some great lessons about using AI for problem-solving from Rich here. First and foremost, you cannot fully trust AI. You cannot fully expect to replace an individual’s expertise, the talents, the disciplines of their roles, as well as their job functions. But you can use it to speed up and help people come to the right decision faster because it provides a way of accessing data in a variety of different formats — whether that’s a Word document stored somewhere on a SharePoint drive or a data point stored in a system of record like an eQMS as an example.

[00:24:49] Josh Santo: Gives you that ability to quickly tap into a vast amount of data and start to bring some structure and some sense to it so that you can interrogate the data, you can apply the appropriate methodologies, you can get the guidance that you need as you need, and really come to the right decisions faster than ever before.

[00:25:07] Josh Santo: Rich, thanks so much for stopping by.

[00:25:09] Rich Nave: Glad I could be here. And Josh, I just wanna add one thing to your summary.

[00:25:14] Rich Nave: We always talk about using the right tool. They’re called large language models for a reason, because they have a large base of languages and knowledge. And so as you just said, find all that data — that’s the strength of the large language model.

[00:25:29] Rich Nave: Use that strength.

[00:25:31] Josh Santo: Use it and use your strengths. All right. Thanks for listening. Rich, thanks for joining.

[00:25:38] Rich Nave: Thanks. Have a great day, Josh.

[00:25:39] Josh Santo: You too.