Lauren McKinley:
Hello, everyone. Thank you for joining our session today. AI, machine learning, and risk. Six things supply chain leaders need to know. Presented by EverStream Analytics. All attendee lines are on mute, and all attendees will receive a copy of the session recording after the conclusion of the session. You can add any questions in the Q& A box, and we will get to as many as time allows.
Now let me introduce our speakers today.
Jim Hayden is EverStream’s Chief Data Scientist with over 25 years experience producing award winning solutions through the practical application of knowledge discovery and machine learning. At EverStream Analytics, Jim leads the data science teams in building scalable, predictive, and prescriptive supply chain risk solutions.
David Schillingford is the co founder of EverStream Analytics and has led the development of several other industry wide data and analytics platforms, both as an entrepreneur and as SVP supply chain at Verisk Analytics. We are excited today to be joined by Daniel Stanton. Daniel Stanton, widely recognized as Mr.
Supply Chain, is a distinguished author, educator, and speaker, specializing in supply chain and project management. He authored the bestselling book, Supply Chain Management for Dummies, which has been translated into multiple languages. Stanton’s expertise is further showcased through his popular online courses on LinkedIn Learning, which have been viewed by over 2.
5 million learners globally. And now I will turn it over to our speakers and Daniel Stanton, Mr. Supply Chain.
Daniel Stanton:
Thank you everybody for joining and thank you so much for having me here. A chance to talk about one of my absolute favorite topics, actually two of my favorite topics. Supply chain risk management is one that, that we’ve been talking about for a while now. AI is one that we’ve been talking about in sort of abstract terms for quite a few years.
Um, but in, in much more, uh, down to earth practice. The last couple of years. And so it seems like a perfect time to have this talk with a couple of real experts in the space and thought leaders about how AI is really, you know, without overstating it really transforming the way that we approach supply chain risk management.
David and Jim, thank you guys both so much for having me as a part of the conversation today. Okay, so one of the things that I find is a challenge is when we talk about, AI and also to some extent supply chain risk management. It tends to be kind of an abstract conversation, right?
And supply chain people, our people are very pragmatic, right? And so one of the things I’m really hoping can come out of this conversation are. You know, specific framework, specific sorts of examples that people can relate to about how AI is being used in supply chain risk management.
Um, and and what they need to be doing in order to make sure that they’re keeping pace with the capabilities that are there and taking full advantage of the things that we can do today that frankly weren’t possible just a couple of years ago.
David, do you want to take another crack at the how and why we’re using AI in supply chain risk management today?
David Shillingford: Yeah, I think this is, it’s a great topic. At the intersection of AI and supply chain risk management, because on one hand, supply chain risk management should touch essentially every function in supply chain. So we get a chance to talk broadly about supply chain and supply chain data, but it’s also to be successful requires an enormous amount of data, both on the risk side.
And on the supply chain network side, we’ll talk a bit more about that. So it’s really, it’s an excellent topic. But before I go any further, I’ll say that calling it supply chain risk management is in some ways problematic. What people associate with that is The idea of reducing risk, and I’m not saying that’s bad.
That’s a good thing. That’s a very good thing. But it risks taking the conversation in the wrong direction because what we should really be talking about is how we optimize the supply chain risk. And it’s all about balance. You gotta know the risk. You gotta understand the risk, you gotta analyze the risk, but then you’ve gotta make that part of the broader supply chain decision making process.
So I, I prefer to call it risk adjusted or risk optimized supply chain rather than supply chain risk management.
Daniel Stanton: Risk optimized supply chain. I like that because like you say. Not only can you not eliminate risks, but you probably don’t want to, because, you know, with risk can come opportunity, right?
That’s part of how we manage the business, but it’s about that balance. David, one of the things that you and I have talked about for many years is how to figure out what your supply chain actually looks like, right? We use the term supply chain and we all, I think, agree. It’s a terrible name.
We’re talking about the flow of value through a complex network, but okay, supply chain. David, what thoughts can you share on supply chain mapping? And the role that plays in what did we call it? Risk optimized supply chains, right?
David Shillingford: Yeah, well, it’s another, it’s an example of why today’s topic is a broad one, because almost every supply chain function is going to require some level of visibility to the supply chain [00:06:00] and with supply chain risk management, because it touches every function and because risk Operates at the network level as much as at any single point in place or time, you really need to be able to see the entire network and people use the phrase and when they use the phrase visibility and so when we are talking about end to end visibility as relates to supply chain mismanagement, we need to be talking about The raw material, the extraction, the refinement of raw materials all the way through the final delivery and whether the company that we’re talking with, or to, or about is upstream or downstream in that entire process, they need to be taking into account the flow of value upstream from them and downstream from them.
And that. I’ll throw out another misnomer where people talk about supply chain risk management, or they say that actually what they’re talking about is supplier risk management, which is an enormous category. It’s a critical part of supply chain risk management, but it’s only part of that because you’ve got to look upstream of suppliers.
In other words, supplier risk, and you’ve got to look downstream and you’ve got to be able to look at the logistics risk at the network level at lane level and at the shipment level so to say everything is not very helpful but as you say to be specific is it’s critical when we’re talking about what where when and how
Daniel Stanton: Yeah i couldn’t agree more that supply chain risk management is much bigger than supplier risk management and even as you say with supplier risk management It isn’t just your tier one suppliers, right?
You have to understand tier two, tier three. We’ll often use the term tier N or N tier suppliers. You’ve got to go the whole way back, but then you also have to think about your own operations, the risks that that you introduced to the supply chain, honestly as well as your customers and the risks that they introduced, which is something that we illustrate a lot.
using tools like the beer gate, right? To show how Customer behaviors can cause all kinds of chaos. So, okay. So we need to map the supply network. Jim, how does AI help us do that?
Jim Hayden: AI helps in a lot of places. One of the hardest problems in this business of who, what, and where are my potential risks. That’s where we’re really getting at is what companies. What locations and what products and materials are at risk.
There’s lots of companies and products and materials out there around the globe. And understanding that entire ecosystem is where AI can help. It can look at literally tens of billions of trading records and derive who’s trading with who and what they’re trading and almost where they’re trading. But not exactly.
The biggest problem there, and this is important for anyone trying to do this themselves, is called entity resolution. And that’s the process of identifying and linking these records from all these multiple data sources that refer to the same entity and inside your own operation, you probably have this problem as well across multiple ERP systems.
You’ve got suppliers, they’re spelled differently, different languages, abbreviations, and you don’t treat them as the same entity. Well, that’s a big problem. If you’re in the business of supply chain risk management. So that’s a difficult problem, and that’s exactly where some AI can help you. It can do way more than things like fuzzy name matching to see if this is the same company because lots of companies out there have the same name.
There’s, you know, Delta had a fire. Well, is it Delta Airlines, Delta Dental, Delta Faucets? You need to be able to discriminate between those entities and where. Hey, I can help is adding more context. Well, is it the Delta that’s in this industry? Is it the Delta that trades with these tier one suppliers of mine?
Is it the Delta that’s in this geo location? Okay. Now it’s more likely to be a risk that’s impacting me. So ATD resolution is key and a must do from a starting point here.
Daniel Stanton: I love the point about entity resolution and I’ll say anybody that’s gone through a Supply chain mapping exercise has seen exactly what you’re talking about, right?
You’ve got a supplier called Acme But they’re Acme Inc. They’re in one, you know, maybe in the The contracts are written with Acme Incorporated, but in the ERP, the orders are sent to Acme Inc. And in the supplier relationship management system, it’s just Acme. And so recognizing that those are all the same company is what you call entity resolution is one of the big challenges.
The other one, Jim, I was going to ask about is identifying locations. You know, a situation that I’ve seen is, you know, You’re trying to map the supply chain. You’ve, you know that you’re working with Acme and you have an address for Acme, but it turns out that’s just the office where you send the orders and the checks to, that’s not the warehouse.
That’s not the factory. So having that location doesn’t really give you an understanding, of what your supply chain network looks like. Any thoughts on AI helping to improve resolution about different facilities for a supplier.
Jim Hayden: Yeah. So that’s where you need multiple data sources and that’s what AI is really good at.
Right. So import export data. That’s great at telling you who’s trading through and what they’re trading to some degree, but. Maybe it’s just the ports that they’re using, or maybe it is just the billing addresses that they’re using on that information. So you need alternative sources. And some very good ones out there are directories.
Sometimes governments put out directories of here’s legitimate suppliers in my industry, and here’s their facilities. Sometimes it’s on companies web pages. Hey, here are my facilities all around the world. And here’s what I produce at these facilities. So understanding those alternative data sources that you can then fuse with this import export data to get more specific on those locations is key, especially in our business where a lot of the risk is geocentric, it’s weather based it’s, there’s a fire at a factory.
Is it that factory or the factory next door? That’s key to risk management.
Daniel Stanton: Okay. So we’ve talked about. mapping the network, mapping the supply chain so that we identify in particular, you’re talking about geo location, right? Knowing where, not just what’s out there, but where it is.
But then we need to marry that up with, okay, so what’s happening there. Right. That would matter that I would care about. And so I, we often talk about that in, as event monitoring. Um, David. Thoughts about some of the challenges involved in monitoring events when you’re running a global supply chain?
David Shillingford: I think the first challenge you hit on that in the question, and that is global, even companies that don’t think of themselves as global companies are more likely than not going to have global exposures in their supply chain, either because their suppliers are in Asia or because the supplies of their suppliers are in another country.
To be able to capture data across the globe. Is it’s an enormous challenge and it’s an enormous challenge when you redefine the data as very often local events that are reported locally. If if you’re just looking at traditional new sources, you’re going to miss most of what is important within the world of supply chain disruption and industrial fire that your supplier might be very big news for you.
But on the global stage, that’s less likely to be the case. It’s likely to be reported in local media. That’s likely to be hard to find. It’s likely to be in a foreign language. And even then, the difference between a news report, even if local and what is actually happening on the ground can be enormous.
So having essentially boots on the ground People who are working in supply chains as partners to provide information as to the situation on the ground is critical. And so to be able to marry up a global view, a real time dynamic or as real time as possible dynamic view of what is happening right now in the world at a local level through boots on the ground across the entire globe is critical.
It’s an enormous undertaking. I’m yet to meet a company who’s able to do that for themselves effectively.
Daniel Stanton: David, when you describe that, it sounds very human to me. Very human dependent. You know, you need to have people there that [00:15:00] understand the language, understand, maybe the industry.
understand how that impacts supply chains. Right. Um, we’re here to talk about AI stuff. So, let me ask Jim. Any thoughts about how A. I. Plays into and can help with that sort of event monitoring?
Jim Hayden: Sure. Sure. Yeah. So A. I. Especially in event monitoring. Let’s say you’re monitoring news posts from around the world, and there are multiple languages.
So you need models that are trained in multiple languages. Eventually you get to languages. Say you do the top 20 languages eventually get to languages where you don’t have a lot of examples in those languages outside the top 20 languages. Okay. So then what do you do? Well, you translate that to English and it’s not quite as good, but it’s pretty good.
And so in event monitoring, it’s an example of stringing together AI models to get to an answer. It’s what’s this post about? Is it about a person, place or [00:16:00] thing? Okay. It’s about a company. Great. Is that a company I care about? Anyone in our customers care about? Okay, great. Now, what’s the story about?
Let’s classify that another algorithm. This is about financial risk. Okay, great. Is it? Is it happening now? Is it real time? It’s a negative earnings report. Well, that could be a precursor to risk. Okay, let’s let’s note that. And now let’s come up with some sort of warning for our customers. Now let’s send that to a human.
That’s how you want that human in the loop there. Humans have context. They can add based on experience. Humans are really good at handling previously unseen events. That’s where I has some shortcomings. It’s trained on data of things it’s seen. If it hasn’t seen something before, it has a very hard time responding to that, where humans are really good at that.
And the best part about having a human in the loop is they can say, No, this wasn’t a valid incident. And that feedback can go back into your machine learning models. It helped them become better and better.
Daniel Stanton: Yeah, it’s funny how the training or how the feedback to the responses from an AI model then help to train it, right?
Yes. And it improves over time. Without going too far into the weeds, Jim when you talk about, um, those different steps, or you described it as different AIs that you kind of need to string together. When we talk about AI, it’s not all large language models, right? Can you share some thoughts about what sorts of AI tools are appropriate for each of those kinds of decisions?
Jim Hayden: Yeah. This is the job of the data scientist. So they understand all of the algorithms out there. They understand the business problem. They understand the data available to solve the problem. So it’s their job to pick the right algorithm that will work best on the data. So I’ll give an example of where a couple algorithms could work.
Let’s say I’m trying to predict arrival times of cargo. I’ve got tens of millions of historical examples of planned arrival time versus actual arrival time. I can understand if it was late or not. I could use a regression model. On a now a shipment I get, and it can predict whether it’s going to be late or not by this amount of time.
Let’s say I’m getting 80 percent accuracy on that. Maybe a better model chooses a classification model that only tells me it’s going to be on time or late. If I get 95 percent accuracy on that, well then the better model, the better algorithm chooses is a classification model. There’s all types of models out there.
In general, they do prediction. They do classification. They do clustering. Hey, these things are all alike. They do outlier detection. Hey, this thing’s different. And then, of course, the latest generation is Gen AI, where they can actually generate content in the form of language or images or video.
Daniel Stanton: Excellent. Thank you for that. That’s exactly what I was looking for there. So, okay, so we’ve talked about mapping the supply chain. And we’ve talked about doing event monitoring, and in both cases, AI can open up some real capabilities. Um, Once you’ve identified through this, these algorithms that something is a risk, then how do we decide whether we should do anything about it?
And if so, what we should do? David, philosophically, just how do you approach risk scoring
David Shillingford: if you’re asking me to be philosophical, Daniel I’ll start by, by saying that on my bucket list is to be the smartest person on a panel and I just realized that that’s going to have to stay on my bucket list for today.
But I, I think the philosophical side of this in, in terms of, you know, how do you bring all this together? Is you really got to start with impact and what do I care about? And what are the levers that I can actually pull? And if you work backwards from that, you’ve got some decisions that are somewhat binary.
If you think about temperature risk, should I use reefer or dry van? That is, there are a limited number of dimensions in that decision. Whereas. At the other end of the spectrum, if I’m thinking about something that, like, do I move from one supplier to another, there are multiple dimension to be taken into account when you’re making that decision, risk is obviously one of them.
And it’s also a decision that is going to take an enormous amount of time to execute. So when you think about the, the impact and the actions that this type of analytic is driving, you really need to start with. What are the actions that I can actually take? How long will they take? What is the cost of that and how complex is it?
Is this something that I can automate today? Is this something that might be able to be automated in the future? Is this something that, uh, I’m going to need a human in the loop for a long time?
Daniel Stanton: Yeah. You know, I know we’ve had this talk and I wrote about it in supply chain management for dummies that a lot of companies that I’ve seen when they’re doing risk scoring and trying to determine impact, right?
Or significance they come up with you know, kind of a matrix to say, okay, well, you know, once we’ve identified the risks, what’s the probability That each one might occur. And what is the impact that it would have on our business? And then when you can multiply those two values together, the probability times the impact and come up with a significance score, right?
Something that helps you filter through what are the risks that you need to do something about? What are the ones that maybe can wait? And then just at a high level, the other thing I’ll throw in is Most of the time I for folks that work in this space, I think we acknowledge once you’ve identified a risk, you’ve only got four options, right?
You can ignore it. You can change something so you avoid it. You can transfer that risk to somebody else by buying an insurance policy or something like that, or you can mitigate, right? You can do something to make the risk less severe and back to that idea of probability and impact. Mitigation is either make it less probable or make our own supply chain more resilient so that it would have less of an impact on us.
Do you think that applies or would you add anything to that or take anything from it?
David Shillingford: The thing that I would add is around severity versus, uh, likelihood. Because I think that’s the sort of the traditional matrix that you described and I think it’s. It’s important to add a couple of dimensions to that so that people just to clarify because severity can mean different things to different people.
And so I think the number 1 is how likely is this to happen? Number 2 is how severe is the event itself? That’s an important difference between the [00:23:00] impact. How is it? It’s a big storm. Okay. That’s a severe event. Number 3 is relevance. As Jim said, is it this Delta or that Delta? Is it my warehouse or is it, you know, the head office is this actually relevant to me and then you need to think about impact.
So impact is what is the ultimate impact? This is going to have on my business. You might add another, this one’s maybe a little more nuanced, but how predictable is this event? And that’s when you think about risk in general is some things are easier to predict than others. Some things when you predict them, you can predict more accurately.
And that’s just it’s worth having in the back of your mind. So I think all of those dimensions that I listed a different and they are important to stack on top of each other and to make part of the models. So that ultimately you can get to actual operational and business impact.
Daniel Stanton: Okay, so I think that’s the perfect segue and, you know, as soon as you use the word model, then it’s time for Jim to come back on stage, right?
So, Jim, thoughts about how we model or use models for risk scoring?
Jim Hayden: Yeah, so you can model probability. The best example there is weather, probably, right? That’s what your forecasters do. There’s this probability you’re going to get this storm. And then you can model severity, too, based on history.
Last time this storm hit this area, it was this severe, right? And so that’s exactly how you get to that risk. But that’s the risk of the event. And that is no matter who you are, that event’s going to happen with this likelihood and the severity. And then in our world, the impact is customer centric.
So impact is about you. Impact is. Based on what product families are impacted. What’s your value at risk? You know, these days, what’s your reputation at risk? And those things are a little tougher to model without some a lot more information about your specific business. So this is where you can combine more of a generalized solution with information about your specific business to get to a true impact score.
And then eventually, what should I do about it?
Daniel Stanton: Yeah, I like that. You know, I often describe what you were talking about. It’s, you know, we look at the past, which is statistics, and then we translate that into the future, which is probabilities, right? Um, and that works as long as the future looks like the past.
We do find situations where the world changes around us, and all of a sudden, Back to the conversation about the role of humans in the loop, it takes some human judgment to be able to look at it and say, yes, if I didn’t know what was happening right now, that’s what I would expect would happen in the future.
But things have changed. And so we have to kind of take a different approach. Gentlemen, we are getting sort of close on time. Lauren, I would like to open it up. If we have any questions coming through.
Lauren McKnley: Yes. Thank you. Great conversation. So a couple of questions. What do supply chain leaders get wrong about AI?
That’s an interesting one.
Daniel Stanton: Let’s take that one. And I’m going to let David and Jim each take a crack at that. David, do you want to go first?
David Shillingford: Sure. I think they either over index or under index. Actually, probably both. Whatever use case you’re talking about, there is either an expectation of outcomes that is unrealistically high, or sometimes there is a unrealistically high.
A level of cynicism that prevents experimentation and progress. I think it’s very difficult to find that right balance in the middle based upon That’s the type and variety of information that business leaders are being fed these days
Daniel Stanton: All right over index or under index jim your thoughts about what business leaders get wrong Supply chain leaders get wrong about ai.
Jim Hayden: Yeah, it’s not magic And it requires work to be good at it And so, uh, often they underestimate the work and this is, I’ve been doing this for 25 years and this has been true forever. The more energy you spend on understanding your data before you even try to model, the more success you’ll have.
You need to understand your domain, the problem domain. One of the big problems I’ve seen throughout my career is not really understanding the problem you’re trying to solve. So I would spend energy saying what probably trying to solve and keep asking that question and good data scientists do this, keep asking that question until you get to what metric are you trying to improve on?
And when you can state that with clarity, then you can start building your model.
Daniel Stanton: I love that. And I will share, um, you know, I’ve got a lot of concerns about chat, GPT and LLMs and how they’re changing learning as as an educator. One of the things that, that. I have noticed myself as an active user is it’s forced me to get much better about asking questions.
You ask your question, you get an answer. You go, yeah, that isn’t really what I needed. But the reason is because I didn’t ask the right question in the first place. And it sounds like Jim, you’re on the same page that if we want to get good results, we’ve got to really be clear about what we’re asking.
Jim Hayden: Okay. Yep. Like I said, good data scientist. You will learn to hate because they keep asking that question over and over again until they get a succinct answer.
Daniel Stanton: Great. Hey Lauren, did we have another question that we can bring in?
Lauren McKinley: Yes, great. It looks like we have time for one more. So the question is how to get started with AI in supply chain?
How do you apply AI to your supply chain to reap these benefits that were discussed earlier? Where do you get started?
Daniel Stanton: All right, let’s do it in that same order. David, I’m going to let you take the first crack at that.
David Shillingford: Yeah, I think building on what Jim said in the last answer, I think you, you got to start with what are most pressing problems and work back from The problem that you’re trying to solve, not assuming it at that point that I is the solution or even part of the solution.
But as you map backwards from each of those different problems, you’ll get to a point where you’re asking, well, what insights do I need to solve this problem? What data is going to drive those insights? And it’s at that point that you’ll start to form a picture around the role that I might play in that particular outcome.
And. I think within that, at the risk of stealing any of Jim’s thunder quite often what you’ll end up doing to solve the problem is not actually data science. It’s something else. But that, you know, don’t start with AI, start with the problem and work backwards until you find something that genuinely can benefit the public.
From using a certain type of AI.
Daniel Stanton: Okay. So I’m just going to be clear on the answer. So the question was, how do you get started with AI? And David’s answer is don’t start with AI, start with the problem. And then when you convince yourself that you really need AI, then go. Look for the right tool. Jim, do you have any thunder left or did David steal?
Jim Hayden: Oh, no. Sure. Sure. Yeah. So if you want to do this yourself and you’ve already decided you want to use AI firstly to do is get some good data scientists. And that’s not, that’s easier said than done. The one trait you need to look forward to data scientists is curiosity. You find somebody that’s curious, you’ll have a good day to scientists where you should spend your energy.
First, we talked about already entity resolution. Get your data in order. Understand the who, what and where of your business before you try to start modeling behavior of those who like
Daniel Stanton: that. All right. We’re getting really close. If I were to [00:31:00] ask each of you very succinctly, one last piece of advice, one take away for our audience.
very much. David, one thing that people ought to keep in mind about AI and supply chain risk management.
David Shillingford: I’ll finish where I started and that’s to emphasize that what we’re looking for is a risk optimized supply chain. That doesn’t mean reducing the risk. It may mean reducing the risk and it’s going to mean different things in different parts of the supply chain, depending upon risk appetite and the actual risk.
So start with the end in mind, work back from there. And think about optimizing risk within the broader context of all supply chain decisions that are being made.
Daniel Stanton: I like it. Jim, one piece of advice, one final takeaway for our audience.
Jim Hayden: Yeah, I’d keep that human in the loop. So AI is probability based.
The essence of most every algorithm is probability, distribution of probability. And even if your model is 99 percent accurate, it’s going to get something wrong one out, one out of those 100 times. And you need the human in the loop to make sure you’re not taking action you shouldn’t be taking. I
Daniel Stanton: love it.
I love it. Gentlemen, thank you both for sharing. This was a fun conversation. Lauren, I will pass it back over to you.
Lauren McKinley: Great. Thank you, Daniel and our presenters for a great session today. We will make sure that we get the answers to the first questions with a few audio issues and make sure that is packaged up for all of you.
If you have any questions about the content discussed on the session, you can reach out to us at info at ever stream dot a I, or visit our website to learn more from David Jim and the rest of the team here on how we apply a I to supply chain risk. Daniel, thank you so much for joining us, Mr. Supply Chain.
If you’re interested in learning more about Daniel, you can find him on LinkedIn. We also have a few kind of handout materials to help explain a little bit more in depth around some of the concepts discussed here on the call that we will share and send out with the recording. So with that, we will wrap the session.
If you have any questions, reach out to us, but again, thank you so much for joining and have a great day. Thanks everyone. Thank you. Bye