Gregor Stuhler: 0:03

A human is usually only able to take 3, 4, 5 different dimensions into our account, taking a decision. And we usually do that by clicking around getting to an Excel sheet, then do a Dun and Bradstreet record. Then we go to the website of the supplier for ISO9001, and bit by bit we like try to inform our decision. A large language model has multiple hundreds of dimensions that you actually can think of and can actually drive much better decisions if you do so. So, here I definitely see a massive use case going forward.

Tom Raftery: 0:33

Good morning, good afternoon, or good evening, wherever you are in the world. This is the Digital Supply Chain podcast, the number one podcast focusing on the digitization of supply chain, and I'm your host, Tom Raftery. Hi everyone, and welcome to episode 328 of the Digital Supply Chain podcast. My name is Tom Raftery and I am excited to be here with you today sharing the latest insights and trends in supply chain. Before we kick off today's show, I want to take a moment to express my gratitude to all of this show's amazing supporters. Your supporters have been instrumental in keeping the podcast going, and I am really grateful for each and every one of you. If you're not already a supporter, I'd like to encourage you to consider joining our community of like-minded individuals who are passionate about supply chain. Supporting the podcast is easy and affordable with options starting as low as just three euros or dollars a month. That's less than the cost of a cup of coffee, and your support will make a huge difference in keeping this show going strong. To become a supporter, simply click on the support link in the show notes of this or any episode, or visit tiny url.com/dsc pod. Now. Without further ado, I'd like to introduce my special guest today, Gregor. Gregor, welcome to the podcast. Would you like to introduce yourself?

Gregor Stuhler: 1:53

Hi, Tom. Yes, thanks for having me. Gregor, founder of Scooutbee and CEO.

Tom Raftery: 1:59

Okay, Gregor. And what, for people who might be unaware, what is Scoutbee?

Gregor Stuhler: 2:06

Scoutbee drives better business outcomes and we are giving global enterprises actionable insights to perfect their supply base and help them to advance their strategic initiatives such as supply chain resilience, profitability, ESG, or innovation.

Tom Raftery: 2:22

Okay. Well let's dig into that in a little more detail cuz that's very, very high level. So. Who might be a typical customer of Scoutbee and what is it you're doing for them?

Gregor Stuhler: 2:37

Yeah, so a typical of customer, of us is a company above $1 billion in revenue that, have different data systems, that want to connect to external data points, and today are driving decisions based on, that data and based on incomplete data. And what, what we are helping companies do is aggregating data from various different sources such as E RP, SRMs but also the worldwide web and also third party data providers such as Dunn and Bradstreet, EcoVadis and so on. And we aggregate all of this data into one singular graph. That's a new way of storing data, can be seen as a kinda a social network representation of your supply base. That gives a very holistic view of your supply base and gives you a contextualized picture of it. And contextualized picture you can imagine that if you have invested in a risk solution and you have invested in a financial data provider those are usually siloed solutions and they're not contextualized. If you now want to however, understand what companies are at risk and how are they actually related to my internal revenue and what products do they provide, then you have to contextualize this data and we are providing this whole data platform for this. So that you have finally full visibility where you stand and where you are with your strategic initiatives. That's the first phase of it. After that, We give also the opportunity to improve your supply base with our supply discovery solution. So we help you identify weak spots in your supply base and help you to find better alternatives to constantly seek supply based perfection based on your own strategic measures.

Tom Raftery: 4:29

Okay. Okay. Fascinating. And what kind of outcomes are your customers typically looking for?

Gregor Stuhler: 4:40

So the final outcome is to, to reach your procurement strategy, right? So the final outcome is that your, that your supply base becomes more resilient, greener, more profitable, and that is related always to observability. So how green or resilient am I actually in a in a certain category, and then also actionability to say, Hey, this supplier is not performing according to your standards. Here are three alternatives to, to help you actually reach your categories target regarding ESG or resilient.

Tom Raftery: 5:15

Okay. And you've talked recently, and I, I've seen it on your website, about Generative AI. Now that's all the new hotness since, since ChatGPT went public last November 30th, if I remember correctly. What is it that you are doing with generative AI that can help people?

Gregor Stuhler: 5:39

Yes. So just to, to get the terminologies right for the audience. So, there's Generative AI and in the field of generative ai there are different areas and one is large language models and ChatGPT is, a public version of one large language model that is GPT 4 or GPT 3.5. By a company called OpenAI and what what we do or what we have been doing actually over the last three, four years already, we have been using large language models constantly. So large language models are quite new. However, they are not that new. They haven't been created 20 22 or 23. They actually came out around 2015, 16. And we have been using those large language models, a actually in various different ways. The one is for extracting information from company websites contextualizing data, merging different data records from different suppliers into one singular one, and yeah, with the, with the new advancement though, we have also expanded actually a lot on what, what we're doing with with generative ai in particular with large language models.

Tom Raftery: 6:56

Okay. And in terms of functionality for your customers, how is it helping them?

Gregor Stuhler: 7:06

Yes. So, in, in general those large language models have two areas, if you like. The one is direct prompting and the other one is synthesizing data. When we actually interact with, with ChatGPT with a direct prompt, you have a direct question and you will receive the information out of that. The information might be right, might be not in the end large language models tend to hallucinate and make up stuff. Actually, they're lying. With, with a straight, with high confidence. And however they are typical use case for that, where you say, Hey, I want to inform my category strategy a little bit. Or I wanna understand like how does the steel market in China look like, et cetera. Like, so informed decisions, right? What What we see, however, as the, as the much more powerful approach is actually using large language models to synthesize data from different sources like internal ERPs, SRMs, et cetera, representing that in a in a way so the large language model understands it and let the large language model help you to actually drive the right decision. However, That is the easy part, right? The, the, the more difficult part is that you have to come to a state where you have to be excellent in two things. One is excellent in aggregating data from different sources, and second, you have to be excellent in representing those data streams in a dynamic manner. And only then actually you can harvest the, the value that large language models actually bring to the table. Without that, large language models are knowledge bases that have been trained on public web. Those are, those are helpful. This connected to the data however, they, they cannot be fully utilized.

Tom Raftery: 9:06

Okay. Walk me through it a little bit more, Gregor, because I'm not quite sure what you mean when you say synthesizing data. What's, what's the, what's the, the value of that? What's the, the function of doing that? What does that do for you?

Gregor Stuhler: 9:23

Yes. To simplify that and maybe even drop the term synthesizing. At the moment, you question the large language model, like ChatGPT, on, on one data foundation, and that is the, the, the worldwide web data that was, that it was trained on and books, et cetera. What you as a purchaser, however, want to have is you want to have the model actually access your existing data. Right, right. You want to make sure that if you ask for example, things like, Hey, who is my most innovative supplier or my, my greenest supplier in Category X, it should not come up with the creative answer, right? You do not want to have creativity here, right? That's and you avoid creativity in hard wiring, the large language model two specific data points internally and tell it literally, hey, Do not make stuff up. If you don't know, tell me you don't know. But there has to be a hard fact actually in this database that you are querying. And if there's no data points, simply say you don't know. And this is the, that this is the major difference where you actually come to to real value.

Tom Raftery: 10:32

Okay. Gotcha. So what, what you're, what you're saying is rather than using the Open AI ChatGPT solution, you're using a separate large language model that you have yourself, and then you seed that with your own company data. Tell it not to make stuff up and suddenly you're, you're you, you have a ChatGPT analog, which is fed all of your company data and therefore you can apply the power of ChatGPT type large language models to your company data.

Gregor Stuhler: 11:06

I think that's an excellent description. Yes, and as a, as a next step, then that's the where synthesizing kicks in, right? That is when you have all of the data actually available. You can actually start training your large language model if it's private. So OpenAI doesn't help here. If it's private, you can start training it and make it understand actually your decision making. So it it can start understanding what does a stable supplier look like? It can under start understanding what does a supplier look like that that might be shaky in the future. And also understand if you ask a question like, Hey, my Supplier A is going bankrupt potentially. Can you please distribute the products of Supplier A across my supply base and optimize for resilience and diversity? Right? Those are things that you can actually start, and for that you have to synthesize all the different data into the large language model. And this is then the final step. And then you have a very, very powerful AI foundation that you can keep on training. And then you have 24 7, always hardworking, know it all, procurement assistant for your whole procurement team. That can not only support decision making now, but actually can take decisions as as soon this system is fully trained.

Tom Raftery: 12:26

Okay, so let me just to clarify again I, I'm getting a little bit hung up and I apologize for this. I'm getting a little bit of hung up on the term synthesize because maybe it, maybe it's, maybe it's a technical utilization of the term, but the, to my understanding, when you're synthesizing something, you're making something up. And I don't think that's what you're saying. I think you're saying you're, you're actually go on.

Gregor Stuhler: 12:53

The Synthesizing AI, or Syn AI, was a term that was actually coined by Andreesen Horowitz for, for this very specific task. Andresen Horowitz is a quite animated venture capitalist. Sure, sure. And yes, I understand what you mean. I think what, what we are, what we want to say is that you have one phase where you give the large language model actually access to the data. Mm-hmm. And the next stage is actually, to merge this data into the large language model. And by doing so, by doing so, actually make your data part of the large language model and fine tune it on your data and help make it or enable it to take decisions based on your data.

Tom Raftery: 13:35

Okay. Okay. Yeah, no, cuz like I said, it. My, my understanding of the word synthesize is creating something, making something up. But what you're saying is you're seeding it with actual real data. Okay. That makes, that makes a lot more sense now. Thank you. And apologies for the, for the confusion. So let's, let's talk a little bit about that. Well, let's, let's. Let, let's take a step back because some people might be unaware. I mean many people are coming to this very new, and the only exposure they've had to large language models is ChatGPT. But maybe just a set bit of context. Give a little bit of a backstory about how there are other large language models and what they look like, where they come from, you know, open source, closed source, that kind of thing.

Gregor Stuhler: 14:28

Yes. Well, the market moves at a crazy speed here. So I assume that as soon this podcast is probably being released, there's yet another large language model coming out. There was one massive movement just after ChatGPT actually the another large language model Lama from, from Facebook. The weights what is pretty much the large language models were leaked. Based on that, there was a rapid in rapid innovation happening in the open source space. Where most of the model that cannot be used actually commercially but are equally good actually or can be at least be compared with ChatGPT, to a point now where there's almost no defensible moat anymore for, OpenAI, also not for Google or Microsoft. The open source base has taken over the large language model innovation. And we'll see more and more models come out. However the, the model itself will not be a defensive moat also, for you as a procurement organization going forward. It doesn't really matter which one you use, in fact, right? Those things are already so good now that even if you take a mediocre performant one, it will still feel like magic. And it will actually keep on improving over time. So at the moment what we're using, but I'm not gonna tell you what we're using. But we are using a version of also another large language model that was also, that's commercially available. And we are actually we have been starting to train it on procurement and supply chain knowhow. So it can be completely deployed in customer hands without any any interference anymore from our side also, right? So this is then your AI, essentially. We help you to train it, et cetera. But like, no, no data is actually leaving this thing. It's your AI, it's on your machine, it's your, your brain. And coming back to your original question so my, my core message is don't worry too much around what are the large language models coming out. There are so many right now out there that are already excellent and more than good enough for your use case in particular. If you talk about generating novels or pictures or videos, that's a different sure. Different kind of breed. But in, in the space where we are in the large language model capabilities are already more than we need.

Tom Raftery: 16:43

Super. Super. And if we think about procurement and supply chain where do you see large language models having the, the biggest impact?

Gregor Stuhler: 16:54

Yeah. I would probably split this in short term and long term. Long term I totally believe, and I'm talking here 1, 2, 3 years after deploying a large language models like ours in the decision making of the system, I firmly believe that if you give and feed this data that you have actually in a, in a, in a meaningful way into the large language model helping the large language model to drive the right decisions will incredible impact. In particular because a human is usually only able to take 3, 4, 5 different dimensions into our account, taking a decision. And we usually do that by clicking around getting to an Excel sheet, then do a Dun and Bradstreet record. Then we go to the website of the supplier for ISO9001, and bit by bit we like try to inform our decision. A large language model has multiple hundreds of dimensions that you actually can think of and can actually drive much better decisions if you do so. So, here I definitely see a massive use case going forward. Short term it's actually accessing data. So short term is like we, how we also start with our customers is, Hey you can, in a, in a conversational way, access your supply base, your existing supply base. And have a conversation with that and say, Hey, what, what suppliers do we have that can provide X, Y, Z? And then bit by bit, actually start adding data point by data point according to your strategy. And maybe at the moment your strategy is to become more resilient, then you want to have some risk data points in it. Maybe you want to have a more sustainable supply base that you should start actually adding sustainable metrics into the large language models and bit by bit build it out here. And the biggest benefits, so of large language models strategically is start now. Right? Get something, get something rolling, get something using, start learning because else you might drift into a sustainable AI disadvantage. If you start two years later than you competition you probably will nevrt pick it up again.

Tom Raftery: 20:04

Okay? There's in, in the tech space, there's the idea of the, the bleeding edge organization. Mm-hmm. So-called bleeding edge because you're so far ahead of everyone else, you, you, you know, it almost causes you to bleed cuz you're spending a lot of money and then there's kind of the fast followers and then there's everybody else. And if I'm understanding you, what you're saying is you wanna either be kind of close to bleeding edge or fast follower in this space, you don't wanna be left behind or because the space is moving so fast, if you don't move, you're gonna be left way, way behind. Is that, is that a fair assessment?

Gregor Stuhler: 20:49

Yes and I firmly believe actually that there will be like a sudden death of companies that have not actually adopted AI soon enough. And the impact of starting late can be actually seen right now in the market in the most brutal form. So the most innovative company in the field that have invested so much in data and AI, Google cannot actually compete with OpenAI. They have lost this AI race. Right? And OpenAI has now constant usage of the system. And even a company like Google with almost infinite resources and data cannot actually chase OpenAI anymore in a meaningful way. It's crazy. Yeah. And that, that means for you as a company in whatever space you are, if, if your two, three main competitors are picking this up, it'll be incredibly difficult for you to come up with a equal good solution no matter what investment you actually put in.

Tom Raftery: 21:45

Yeah, yeah, yeah. No, I, I know what you mean about Google, because since late last year, early this year, I haven't used Google. I've literally stopped using it. You know, I, I, I use, I used it all the time. I use it every day, multiple times a day for my searches. But as soon as ChatGPT came out, and then you had Bing come out with chat built into it, I've stopped using Google because Bing while slightly slower for a result. When you use Bing with chat, the result is significantly better. In Google you get, a page with 10 results. The first three of them are ads, and then you have to look down to see which one may or may not be applicable and you might have to go onto a second page cuz the first ones are, you know, whatever. Whereas with the Bing search, you get one result. It's generally the right result and you've got a series of links underneath it to source material. So it's saying where it's getting the result from. It's just orders of magnitude better. And yeah, to your point, Google has been left sitting in the dust. They're gonna have to do something really quick to, you know, I'm sure they must have lost a massive amount of search traffic as a result.

Gregor Stuhler: 22:59

For sure. And Like Bing and not being able potentially in the traditional search being able to compete with Google anymore. Bing has now actually the the, traction and the usage of users searching through conversational chat and with actually Bing having the higher usage, it will be equally hard now for, for Google to chase that and to come up with a, with a equal good offering, almost impossible.

Tom Raftery: 23:27

Yeah, yeah, yeah, yeah, yeah. Just coming back to supply chain, right. We're, we're geeking out on all this stuff, but coming back to supply chain, how challenging will it be for organizations who have large legacy systems that they've invested lots of money into transition over to using this new way of doing things? I mean, there's a lot of change management issues to go through there. There's a lot of potential trust issues and, and privacy issues as well. When you're dealing with data, obviously. You know, a a again, we've said that it needs to be the ones who can move quickly, need to move quickly because they'll be left behind. So walk me through that process. How do you think that's gonna work out?

Gregor Stuhler: 24:20

Yeah, so, coming back to a statement earlier in the podcast you have to be good in, in two things. Aggregating data at scale from different systems. And second, representing this data. And this is hard. It's actually very hard. In particular if you have dozens of different disconnected systems. Merging those records, connecting those records, and merging it with the external Dun and Bradstreet record, reflecting this into one single data node, et cetera. All of that is hard, right? So it's And there are companies out there that do that. That's us, for example, right? So we, we do nothing else. We, we our whole DNA is aggregating and representing data. This is our dna and this is where you have to start investing And connecting this then in the end to a large language model is actually quite easy. That's not the hard part. You have to, of course, ideally have a, so, first, that's easy. What you need though, you need a private, large language model. You, you want to have this running on your instance. Second, you want to have a large language model that is trained on your space, right? So, and trained actually into, in your expertise. And you want to have a large language model actually that is connected to your data and actually can digest it, right? So those things also if you go with a completely open source model that is not trained on your space, et cetera. It'll be simply inferior. It won't be bad, it'll be just not good, not, not better, right? So that's, that's about it. So, we have a very clear approach how we start, and that is we actually start with, with small use cases together with the customer. So customers give us not a lot of data or, or not us, but actually feed actually data into the system. Not a lot of data. We usually only start actually with the company names, right? And this is already enough for, for our system to aggregate all the web data, including addresses and and so on and all the web content. And then connect this as a first solution actually to a conversation AI so that the customer can already have meaningful conversations with their supply base. And then bit by bit build it out. And that is incredibly easy. And then your, your, your first AI solution may be actually live within a couple of weeks. We're not even talking months.

Tom Raftery: 26:38

Okay. And is this hosted on your infrastructure or your customer's infrastructure or is it up to your customer to, to decide

Gregor Stuhler: 26:48

Up to the customer to decide. We don't do on premise, but it has to be on the cloud. But in the end it's up to the customer where it's deployed, and it can be also in a, in a, in a private shielded area on the, in the cloud.

Tom Raftery: 26:59

Okay. And I mean, you've, you've mentioned sustainability and ESG data as well at times in, in this podcast so far. How do you see these large language models helping organizations better achieve their ESG and sustainability goals?

Gregor Stuhler: 27:21

Only when the data is connected. Else it's, it's just a better Wikipedia that might come up with wrong answers. So here, it, it won't help you a lot. It'll help. However, if you have connected to your data and you have merged your internal data actually into the large language model. Then the model will be able to reason actually, and also help you to to optimize your supply base by, for example saying, Hey supplier A is performing very well according to our performance standard from the supplier scorecard. The total supplier revenue is 100 million. We are only spending 2 million with this supplier representing only 3% of the spend in this category. We can grow with this supplier and achieve our ESG goal. And this sounds very fancy, but in fact actually it's just connecting three data points and making the the large language model reason that the rest. The power comes, of course, that the large language model can do this 24 7 and can start scenario planning in the background while you are sleeping or on vacation. So it's constantly helping of course to, to drive the right decisions then

Tom Raftery: 28:29

Fascinating, fascinating, and where to next? I mean, we've seen these large language models kind of explode on the scene, come from kind of zero, almost to 100 in a very short space of time. We're on in the ChatGPT space we're on GPT four now, you know, where, what's, what's gonna come next? What kind of features and functionality can we expect to come with subsequent versions of the large language models? Or, or, or, or is this it?

Gregor Stuhler: 29:03

I think this is it. So I, I do not see that there would be a, a radically better version coming out anytime soon. OpenAI did excellent job. And I mean, we, it already feels like magic to us, right? And, and for, for certain things, though, large language models are simply not are not good at, right. When it comes to math, for example, when it comes to complex reasoning from different domains, et cetera. This is where large language models are not good at, and that's where we see, however, a massive innovation right now happening is with technologies and plugins. For example, ChatGPT coming out and saying, Hey, How can I actually level or remove the weaknesses of the large language model by doing chain of thoughts reasoning with certain technology or by connecting Wolfram Alpha from a from a, from mathematics, and maybe also from a, from a space knowledge point of view.

Tom Raftery: 29:59

Fascinating. Fascinating. We're coming towards the end of the podcast now, Gregor. Is there anything I haven't asked that you wish I had or any aspect of this we haven't touched on that you think it's important for people to think about?

Gregor Stuhler: 30:14

I think we actually were quite comprehensive. The one piece that I, I would like to add is that don't hold back and simply start using it. Right? So the, the, the first simple solution that we are deploying is actually quite powerful already, right? So it actually helps you, for example, in, in times of, of risks, right? If a supplier goes bankrupt, you want to know, do I have a supplier that fits as an alternative in my existing supply base? And you cannot rely on the data that is sitting in your ERP or SRM. And missing out on that is actually terrible. So like having actually a conversational AI simply on just a very small data set is actually already quite powerful. So don't aim too high, right? Aim, aim very high when it comes for your vision and your North Star, but then start simple. Start lean, get your stakeholders involved to get buy-in from the organization, and then bit by bit, build it out according to your internal agenda.

Tom Raftery: 31:11

Okay. Fascinating. Gregor, if people who would like to know more about yourself or any of the things we discussed in the podcast today, where would you have me direct them?

Gregor Stuhler: 31:22

Our website, scoutbee.com. There's also a contact form. We also actually just released our white paper on AI and large language models, so feel free to download that to get even more, know, more knowledge. And yeah, looking forward to hear from you.

Tom Raftery: 31:36

Fantastic. Fantastic. I'll put those links in the show notes, Gregor, so everyone has access to them. Gregor that That's been great. Thanks million for coming on the podcast today.

Gregor Stuhler: 31:46

Thanks for having me.

Tom Raftery: 31:48

Okay, we've come to the end of the show. Thanks everyone for listening. If you'd like to know more about digital supply chains, simply drop me an email to TomRaftery@outlook.com If you like the show, please don't forget to click Follow on it in your podcast application of choice to be sure to get new episodes as soon as they're published Also, please don't forget to rate and review the podcast. It really does help new people to find a show. Thanks, catch you all next time.