AI Functional Use Cases: Manufacturing

so that being said, uh let's kind of talk about uh orchestrations. And you know, as as I mentioned before, if any of you attended my previous sessions, which is that orchestrations are are very much the heart of anything to do with AI and manufacturing, AI and JD Edwards in general, because orchestrations are essentially the way that JD Edwards enables and consumes AI services, right? So our AI services for the ones we use are all on the Oracle cloud infrastructure OCI. Um you know that's because we um have a great relationship with Oracle. We we really love the technology they're building in OCI and we really love able the ability to natively authenticate with and consume many of those OCI services that are um that are available right through uh orchestrations, right? And um what OCI enables us to do is really create that artificial intelligence-based tooling inside of their platform in OCI. Things like you know through the digital assistant tooling they have through the um the uh document vision components that they have the um the predictive analytics components that they have. And then what we can do is using orchestrator, we can send data back and forth and then do actions inside of JD Edwards based on the outcomes and the results we get from the AI tools inside of JD Edwards. I will apologize in advance. My cat Onyx will probably jump up on me at some point. Um, she's the real brains behind our AI practice and our our our orchestration practice. I'm just a front. So, just uh an FYI for anybody listening. Um, I don't know where she is, but she's she's here somewhere. I I I see her meandering around me. Um so you know uh as you can see here on the right the example with this is and is that you know with OCI technology we can extract data from documents. We can extract data from images. We can uh have it run very deep analytics uh and then send that information back to JD Edwards via an orchestration and we can do something inside of of JD Edwards via that orchestration. And like in other sessions again you know I always draw the use cases towards an outcome of what will happen but the outcomes are very company use case business industry specific. So don't focus so much on what the outcome is. Focus on more what the insight is that you can gain from AI because that's going to be more applicable to your business right the outcome for some businesses is to update something for other businesses is to inform something. For other businesses this is to just do nothing. the the real value is in what are the insights that you're going to get and how quickly can you get those insights. What you do with it is going to be very unique to your business, right?

So, um, you know, I think the first use case we're going to talk about, again, I go I do these in categories of like vision, agentic, and then predictive, right? Um so and and one of the things you'll notice is that they're they all kind of build on each other right um so you know it's very much like okay what's kind of easier to do what do you build on top of that how do you more do more enablement with it and how do you make this more proactive more future facing right.

So, uh the first use case we're going to talk through today is around um using computer vision to do quality assurance so this is similar to what we talked about with inventory right? Where something's coming off your production line. And you know, again, a lot of my customers, especially manufacturing customers, the way they do this is your um you're doing um you know, you're doing kind of like, you know, random sampling or statistical sampling and then that's how you're determining you're you're taking you're doing QHX every like 10th item off the line or something along those lines, right? uh or instead what you really can do is you can actually use AR um technology right or VR technology or computer vision technology capture images around your production line you can capture images for everything right and start to extract and you know contextualize the items for those images and then um you know even if it's like around packaging of those items and then identify anything that's not compliant, non-conforming, defective and then immediately sort it out right before it kind of gets to uh into your storage facilities or your D distribution centers or anything like that, right? So, and and if you don't then well, you know, you've got really great data around um storing uh things that that are compliant, right? So, you can you can kind of capture both, right? You know, is something within tolerance? Is something um the right color, the right shape, the right size, right?

Um and these systems have actually been around for a long time, you know? Um uh I've had a customer who's a recycling plant and they've been using something like this to separate out recycling for a long time. Um Agra business uh you know has has been using this to to for fruit and and vegetable sorting, right? So this is technology that's been around for a while. Um but it's just again it's more available now because we have all these great OCI services that let us do this, right? You know, we're not talking about millions and millions of dollars to do these things anymore, right?

So, um I think another one of my use cases is and this is one of my favorite ones and and this is actually an example of the one that I did uh back in 2018 that I was I was mentioning earlier is where we can take engineered specs or drawings even CAD drawings or PDFs um use computer vision to extract information about those um and then find um what would be the bill of material or the routing for that item.

Right? Based on historical items, based on past items, based on past estimations or anything you've done and then send those bombs and routings to orchestrations which will create the item, the bomb and the routing, right? So, if you think about that, that's that's pretty revolutionary for a lot of customers who are doing all of this by hand. like you know if you just you have a PDF drawing you you import it uh a AI program essentially extracts that and says hey in this is like a 95% or 99% match to this other item that you did you know like 10 years ago or four years ago or it's an 80% match to another item you did like four or five years ago or or hey it doesn't even know it's a match it just kind of is able to assume that this is this part this is this part this part right um and then it can build out a um uh you know it can build out what the bomb and the routing is and then send that to an orchestration. It can enter that into into JD Edwards.

So you know this can really really expedite your product creation time. This can really expedate if you're a business that does a lot of engineering or estimating uh based on like customer provided specs or drawings. Uh this can really really expedite that time. Right? So if you're if you have a customer or you are a customer that runs a business where you're taking in CAD drawings and your business folks are doing estimations um based on those CAD drawings and your estimate time is like if it's more than like two or three hours um if it's if it's a couple days because a lot of my customers are estimating time is like a couple days right um this can help bring that down drastically very drastically.

So this is essentially the same use case but rather than creating new you can update right and the reason uh this one is important is because you know and and this use case came out of a real customer example of ours where we had a customer who um was was doing some work around MRP and um was finding that their MRP um you know results were a little inaccurate right. So there was a dis there's there's a disagreement between um the outcomes of MRP and what reality was really reflecting and what we uncovered was that there was actually uh quite a discrepancy between the engineered drawings and the actual bills of materials and the routings.

So you know what we came up with was saying, "Hey, can we take your engineered drawings? Can we extract the item information around those? Can we give you essentially a percentage of how far off your drawings are from your bumps?" Right? And I think that once you kind of start taking that to the next step, it's okay, I know how far they are. Can I now get them to be corrected? based on my my analysis of these drawings. So they get for a human being to do this would take hours and hours and days, right? But you know with an AI program, we just dump all the drawings in. You know, we we we spent hours to build the AI program and now it doesn't really need hours to do the analysis. It really just needs seconds or minutes. And then it can send that updates to to JD Edwards. And again, you can update the drawings. You can have this go to a recommendation engine. you can have this be presented to the user and then they can make an acceptance or or rejection. But this is this is also kind of building off that use case of saying, okay, you know, again, if you're a business that does estimating or you're a business that does a lot of um you know, a lot of estimating based on specs or engineered specs or or drawings you may get. um you know this would be a really great use case for that because you can take those drawings you can extrapolate information out of them and you can update JD Edwards with with bomb and routing information for from those right.

So now we get into more of the predictive um capabilities around AI with manufacturing. Right? So you you'll find that in a similar theme with all the other ones we've talked about so far, we're able to use anomaly detection and pattern recognitions to really ident like take past work order and inventory data and start to identify items with outdated or incorrect scrap percentages. Right? Um this is another thing that's hard to do. It takes a lot of studies um for companies to figure out what really is our scrap percentage. Um and then there's data that can help you with that, right? There's data that and I'm not going to be one of those people that says, "Hey, the data is going to just tell you the whole answer, right?" But I think the data gives you a great starting place, you know? It's like it's like you know uh it's like when you know um I'm sure some of you have kids who are writing papers and stuff in school and you know they're they're using chat GPT to write these papers and you know you know the thing that they're teaching kids now is that you know this is not going to give you the final result but it gives you a good starting place right same thing with using code uh coming out of some of these AI tools it's not the final result but it's a great starting place right so same thing with this is you know you have an you could have a predictive al uh or sorry anomaly detection algorithm that goes through and says, "Hey, you know, um the these multiple data points when combined together don't really agree with this outcome, right?" Um you know, you've got a scrap percentage of 0.25% here, right? Or 2.5%. Um but really all your math and everything else is actually saying it seems like it should be more, right? Or it seems like it should be less. So, you know, you can find what those favorable or unfavorable amendments should be. But again, it's it's a recommendation engine. Ultimately with these, I still always draw and recommend that a human being is making the final decision, you know, um uh whereas the AI is kind of providing the recommendations. So, we're kind of taking the laborious task of the analysis out of the human being's hands.

um and and instead putting that that kind of in the hands of an AI agent and the laborious task of updating stuff inside of the system out of the out of the human's hands. So really what we're kind of putting and leaving in the human's hands is their judgment, right? You know, um and that's ultimately what what human beings are um are meant to be good at.

So another example you know or use case I want to talk through is reviewing manufacturing and and inventory data to recommend routing and bomb changes right so same thing here is we can analyze work order data uh especially if you have shop floor or machine data being passed back into JD Edwards where you're capturing actuals around hours or you know operation um you know if you're actually capturing things like uh you know different machinery that you're using you know you can you can mine that data and start to recommend bomber routing changes or alternate bomber routings that you need to create um that that are kind of along this right so same thing you know with a recommendation engine you can you can take that information you can extrapolate different patterns out of it and then you can provide that back and again you still want the human being kind of there as the person who's making the ultimate yes or no decision. Um, but we're taking the laborious task of analysis and the laborious task of updates out of the human being's hands and we're really keeping the the value ad task of of of judgment and decision-making inside of the human being's hands, right? Human beings basically Robocop, you know, it's ultimate judge, right?

So another predictive model that I think is really interesting here is um anomaly detection and pattern recognition to identify work orders that are at risk or of um or high probability of delay or quality risks. Right? So um if you you know I I for example have a customer where they complete a work order and oftentimes they they complete this work order and they'll find that oh man um you know 30% of the parts we made in this work order are out of spec right so first of all we should be talking about that first very first use case we talked about where you're doing QA checks while you're manufacturing. uh if you don't then you know the other thing you do is you kind of come with the d other side of it which is okay can I identify items that have a high probability of risk around quality or high probability of risk around completion because my capacity is constrained I I'm I'm at a you know I'm and again you can get kind of really cool with this where you can get some pretty nifty things in this like you know if you're you can take like seasonality into account you know am I in flu season right am I in um you know am I in back to school season where people are going to take time off right? Am I going to um start to see any sort of delays because of capacity constraints or anything like that where um you know an order is at risk of of delay right uh and then you know the big thing still you know as I'm as I'm always drawing back to these is that you draw to a recommendation and then the human being makes the decision on do we need to change work order dates does that need to be com you know you know planned back does that need to adjust ultimately our master production schedule Right.

Um so the engineer reviews the recommendations you know a production planner reviews the recommendations and then um you know that gets entered back into JD Edwards if necessary or or action in some other way right.

So lastly a lot of our customers use a a you know a work order sales order combined or like a W line type right? So this the same probabilistic uh you know risk identification works on both the sales order side as it does on the work order side. Right? So if you have sales orders that are at risk of delay because Wline type work orders are going to be delayed or because um there are chances of delays happening in in manufacturing overall again due to capacity constraints, due to environmental factors or anything else, right? You can you can get very very deep into the level of of of kind of analysis on this, right? which you know again in kind of this this new um you know very dynamic geopolitical world we're getting into um you know I I foresee the types of data that we use to analyze um risk is is going to drastically get more and complicated and change right um you know it's becoming a very complicated and complex dynamic world out there so you know it's just how we adapt so same thing you know can we identif identify sales orders that are at a high probability of delay or risk of delay. Can we provide recommendations to customer service on changes or or or possible delays that they can communicate back and then amend those sales orders that again and and and you're like why would you want to do that, right? Because you know that data then ultimately will feed into our planning systems. that will ultimately feed into our production scheduling systems and it creates a more comprehensive supply chain and more comprehensive production plan because we're having that that continuous updating of data. Right? If data stays stale, a production planning is stale, right? So, it's really really important to keep that continuously updated where we can and and and kind of um keep it dynamic. Not saying you want to always change your sales order dates, but you know, it's it's one of those things where if you you have to, you know, you want to make sure that all areas of your business are are kind of in sync when you're doing it.

So, that being said, those were all the use cases we have for manufacturing. I very much appreciate your time and your attention. If you have any questions, please feel free to drop them in the chat. I will hang out uh here for a few minutes to ensure that we get all questions answered. Uh, and thank you all for your attendance with AI week. I hope you all enjoyed it uh as much as we enjoyed putting it on for you. Um, and I hope you got some really cool insights around AI or cool insights around AI uh out of this out of these sessions. So, thank you all and and as always, please reach out if you have any ever ever have any questions around AI, want to nerd out about manufacturing um or or want to, you know, talk about my theories on on robotics and manufacturing, right? So um always happy to happy to talk through anything. So thank you all for your time and attention. Appreciate uh you joining and uh please join the customer panel uh later today.