Name: AI Agents in Action: Orchestrate the Modern Supply Chain
Uploaded: 2025-08-25T16:30:49.115Z
Duration: 1 h 1 min 8 s
Description: AI Agents in Action: Orchestrate the Modern Supply Chain

Transcript for "AI Agents in Action: Orchestrate the Modern Supply Chain": Great. Well, hi, everybody. Thanks for joining our session today, this webinar covering AI agents in action, specifically as it relates to supply chain management and orchestration. So let me, start with a few introductions. I am Christine Andrews. I am a senior solution consultant here at Dataiku, specialized in manufacturing and supply chain, and I've been, in manufacturing for over twenty five plus years in various capacities. So I look forward to leading, a little bit of the discussion early on, and then I'll let Romain introduce himself as he will be sharing the demonstration. Yeah. Thank you, Christine. Hello, everyone. So my name is Romain Menini, and I'm senior solution data scientist at Dataiku. I've been building, actually most of the, manufacturing supply chain energy, solution, I will, I will have on the catalog, and I will walk you through this, this demo of, supply chain, tonight. Great. Awesome. Thanks, Vermont. And, look forward to actually getting some questions here as we go through this. So, first, a little bit of, like, basic setting of the stage here, if you will. And for those that don't know Dataiku, let me share a little bit about what we are. Dataiku is, analytics orchestration platform. Analytics can take many different forms. Right? Most common and and reason you're probably here today is around artificial intelligence and machine learning. And we'll talk specifically about agents and also a little bit about how that relates to the more traditional AIML approaches. As a analytics platform and orchestration engine, Dataiku, part of the reason that, our customers leverage our platform is for a couple different things. One is we are cloud and infrastructure agnostic. So you can leverage Dataiku. You can just Snowflake, Azure, GCP, AWS, all of the classic approaches. So it'd be able to blend the multi cloud strategy, in support of analytics. And the other big part of it is that we aren't a bespoke niche, you know, analytics platform provider, which it allows for, you know, your supply chain professionals, your manufacturing professionals, so your leaders across your various domains throughout your business to be able to drive insights, answer questions, leverage AI and machine learning to help them be more productive, and do that in a transparent, very scalable way. So in other words, really, we're about how do you support decentralized analytics pipelines in a consistent manner. So that's a little bit about who we are. Now let's get into talking about AI agents, and then we'll talk specifically about, how they fit into the supply chain. So first of all, what is an AI agent? And it depends the way you kinda get a a myriad, I guess, variety of answers depending where you're going for this kind of information. But behind it all, there are large language models when we talk about agents. So those of us that are familiar with Gemini or ChatGPT or any of the other number of, AI chatbot assistance, that's fundamentally what's behind an agent. However, what agents do beyond just what, traditional LLM approaches is that they allow for a really automated autonomous decision making. You don't have to go to that extent, but they do provide a means of going beyond retrieval to actual, helping with the execution of tasks in a very seamless way as well as prescriptive recommendations on what to do next. And we'll talk about how that applies, and it can apply, by the way, broadly across various different, domains of supply chain. Right? And they also the other thing I'll say about, you know, why people investing in AI agents and why it's getting a lot of focus right now is, I guess, to use an analogy, right, if, LLMs made NLP, you know, natural language processing more usable and valuable, AI agents have made, you know, artificial intelligence in the decision making therein more usable and attainable for a lot of the organizations we work with. And in the end, the reason why that matters is we can talk about agility, trust, and and all of that's very important for supply chain, but, fundamentally, it is about competitiveness and cost. So if you think about why do I care about agents, really, it should be about understanding, are we doing things in the most efficient way possible so that our supply chain professionals can be more productive. Right? And they're not having to hunt for information. We'll talk about how agents can help empower that specifically in supply chain, but that's fundamentally where we see a lot of our customers getting benefits from agents. It's it's also about being able to kind of orchestrate and automate some of these traditionally manual processes. But let right now, I'd like to jump into really quickly a poll question. We talked a little bit about the just just basic understanding of what agents are and why peep why organizations and leaders are investing in them. But I'm kinda curious how is your organization, your enterprise currently using AI agents? And you'll see so let's open up, Stephanie, we can open up that poll question now. It should be open. And yeah. And I already see we have votes coming in, so that's great. And it's okay if you don't know. Right? Or not at all, just getting started. So really kinda curious on the breadth of maturity, across and we'll as, the those results come in, we'll talk about them here in a little bit. Okay. Let me go forward while we're waiting for the poll. So feel free to keep, answering that poll. That's great. Looks like most right now, by the way, are on the just getting started, which is good, which is pretty appropriate on that probably means there's some experimentation. And we do definitely that's usually how customers start. Okay. So let's talk a little bit about the use cases, whoops, for supply chain itself. Here we go. Alright. Challenges. So anybody I don't know the the breakdown of everybody here and in this webinar today. But if you're intimately involved with supply chain, a lot of these terms are not gonna be, you know, very foreign to you. Right? The challenges across supply chain are if I had to say what the biggest kind of challenge overall is really around the, wrangling and understanding of data across myriad systems. So supply chains in and of themselves, right, imply a lot of different departments and necessary silos, by the way. I'm not this is not a conversation about breaking down silos, but it's really about how do you leverage those silos effectively and how do you actually bridge them with analytics. But the challenges are around with that, a lot of manual processes. So there's a lot of ad hoc reporting, a lot of, like, clear to build type initiatives that are just have to be brute force initiated. Still a lot of Excel macros out there and, of course, tremendous amount of BI tools. Add to that fragmented data sources. So a lot of our customers have myriad ERP systems, different types of implementation of ERP, warehouse management, MES systems abound, and it's it's quite diverse what that could consist of. Fragmented supplier information. So usually not everything you need to properly manage your suppliers is in your master data, in ERP, your supplier master data. And then finally, outsourcing. So that can take a lot of different shapes and forms, depending on what industry you're in, but outsourcing from everything, of core manufacturing, of course, to, three p l, you know, third party logistics providers, things along those lines. And that can ease it that can just add to that data, complexity and where you have to go for information. And if you look at some of those statistics out there, you know, only 53% of, organizations report that they have adequate master data quality. So what does that mean? Well, limited visibility. The the quality the data quality hampers some of that visibility. But without a doubt, when I mean, this has been the case for for many years. But when people talk about supply chains, they talk about transparent supply chains. Right? And traceability of products, traceability of suppliers, not just tier one, but down to tier third, two, three, four, etcetera, which all have an impact, right, on, supply chain risk. The data quality issues we just talked about, that's rampant and that's endemic, by the way, throughout organizations. That can be incurred everything from technical debt, outdated ERP configurations, all of those types of, of issues than just human error. And then harmonizing across different users. So a good example of that would be inventory information. So inventory information can be reported different ways in different parts of the organization. All of that can produce opaque analytics, whether those analytics reside within your ERP system or separate supply chain management system. But sometimes that, explainability and transparency of the analytics can be lacking as well. And then finally, I would say one of the most I have a lot of conversations about is adaptability. So everybody talks about supply chain resiliency, but that in and of itself implies you have to be able to adapt to the unexpected. Right? Which may sound like a paradox, but what what we find a lot of customers do today is they need the ability to import and actually input new and ingest new kind of data signals, new data sources that maybe with our fixed ERP configurations that can't easily support today, because there's always a new disruption around the corner and evolving requirements. So with that, how do agents and analytics generally help? So these are examples of agents at use in supply chain. But if you go you think about the manual effort, so there is that process automation aspect to it. And I think you'll, start to see some aspects of that when when we get into the demo as well. But that process automation allows for, greater reuse, more seamless automation of some of these workflows with better analytics involved and doing it in a in a, higher level way. Right? Not just updating things based on, you know, a fixed input, but being able to adapt to kind of new requirements. The worker augmentation, this, by the way, is without a doubt, probably the number one use of AI agents in supply chain today. It's usually less about what I'll say in the intelligent business chains, which is about taking agents and knitting them together for a more seamless kind of a dynamic, right, workflow execution. But worker augmentation was without a doubt the biggest example, and it makes sense. Right? It's a little bit lower risk. You're still getting the people involved. You're improving the productivity and efficiency of your supply chain professionals. And that's really where we see a lot of growth and a lot of current interest in agents today, and particularly in supply chain. Now with all of that, you what I'll say is that those are the ways that you could overcome some of those challenges by automating things, by improving the productivity, by seamlessly wrangling, combining data, and then surfacing up some insights for things like safety stock optimization or better demand planning. But, really, when you think about it, agents can apply. I I get I actually get this comment a lot where people will say, oh, we're not ready for GenAI. We're not ready for agents. And I I would challenge that in most cases because, really, they can apply in any part of your supply chain organization. It's just how you're gonna leverage them may be different. It's not gonna be for probably full, you know, autonomous, right, execution, and updating of master data, for example, necessarily. You're not gonna start there, but you absolutely have the ability to surface up, and that's what we see our customers doing today, and there's some examples of it. Things like supplier risk or, where should I opt you know, how can I optimize sourcing, for example? And then how do I adjust how should I adjust my demand plan if I want to or if I'm gonna change some discounting structures? In other words, how does my price influence demand? All of those things can be identified, if you will. Same with routing optimization for better dispatching, and better load planning. And when you bring all of those together, that's really obviously, it's the whole point of an integrated supply chain, is to really set the stage for that kind of seamless intelligent business automation and orchestration. So what are some of the, use cases on and and the journey is often a question I get. And it is a journey. Right? Any kind of analytics. And when you're talking about agents, that's also going to be a a bit of a journey as well. So what I will tell you, these are some examples. They're not exhaustive, but there's some examples of, how we see customers progressing oftentimes on this journey and implementing things like GenAI agents. And they might start with, for example, a clear to build, which is more descriptive analytics based, but then moving on to things like safety stock optimization, different ways to facilitate that, but basically tying that to demand plan and looking at variances and then updating safety stock on a quarterly or even monthly basis. Materials forecastings, that's more your, dependent demand where you're looking not just and we have customers that are doing that today as well. We're doing a better job, essentially, of, that material plan and that material planning component. And then finally, when we when you start to have that basis, obviously, you have to understand what your material plan is and have some handle on that. That's really when we start to see, a little more interest interest in the the GenAI agents for things like inventory optimization. So that would include it'd be a little bit more encompassing, more than just safety stock. Right? Things like the minimum order quantities, reorder points, and, those kinds of metrics. And, really, that's about process optimization, being able to opt, to dynamically update, some of that information. Like I've said, minimum order quantity, lot sizing, and things like that. One of the areas, though, I'm gonna I'm gonna, spend probably a little bit more time here because, this is one where we're gonna show a demo here a little bit. But the other part is that this is probably the area that is getting the most attention. I see the most traction today among enterprises, and that is around supplier management. Spend analysis, that's your typical descriptive starting point. You're probably doing that. Your organization are probably doing that in some capacity today, but oftentimes, it needs to be enriched. There's and as you wanna go through, for example, some optimized sourcing use cases, and supplier risk management, you'll oftentimes we have organizations that discover their spend analytics aren't what they need it to be to do the proper kind of categorization of suppliers. And then it's supplier scoring and, you know, we certainly have customers with hundreds or hundreds of parameters that they're using, believe it or not, for things like supplier scoring. So that's why it lends itself to something like machine learning. But where we where I am seeing and we have been seeing at Dataiku, a lot of traction is around a couple of things when it comes to supplier management. One, contract. So contract management and contract insights, and then also, supplier risk management and kind of corollary to that, this optimized sourcing. That's really about using agents to do a couple things. One, to surface up and that's what Ramon's just sharing a little bit to surface up some of those risks, operational performance risks among supply you know, vast supplier networks, being able to, you know then the the agent what the agent could do is not just surface up those risks, but then provide recommendations on what you can do to mitigate those risks. And then similarly, on the optimized sourcing, being able to then retrieve what is who should I, when it talks about optimized sourcing, who's the lowest cost supplier, based on a certain lot quantity or quantity that that you, delineate? And then, also, what about, optimize from, say, a sustainability angle or or ESG metrics angle? So, really, that's what another area that agents are very, very good at helping, to define. Because, again, being powered by those large language models, they can kind of scours publicly available information and also help determine, you know, who's the lowest cost, but also who's most compliant. So that's really and and, again, I think also because of supplier management being an area where there is a lot of documentation, it just naturally lends itself to kind of an LLM based approach. But, again, that's not the only area we see it. And then finally, on logistics. So this is, another area where there's a lot of opportunity for more automation, and that's around, that what what we see typically is, here around fleet management. And there is actually some opportunity for agents and, in some cases, we do we're starting to see a little bit more of that agents in the fleet management side, but you could think of that as more maintenance planning. So recommended maintenance planning based on, routes, based on utilization, and, also the the health of the fleet or particular vehicle. Then, delivery scheduling is a lot of times the next, kind of use case that a lot of our customers, you know, pursue, and that would be around understanding, right, lead times, understanding delivery times, and when those are gonna when, so that you could better kinda coordinate with the the the logistics providers, transportation providers, but also keep the customer apprised of what's happening as well. But really, when we talk about, a couple different areas for worker augmentation and actually here in process automation, it's really around route optimization on the automation side, but, warehouse capacity optimization. So that's really about surfacing up. So when you think about, there's a lot of analytics. That's the other thing I'll I'll point out. When we talk about some of these agents behind them, at least the the projects that we work on, a lot of times, there are some underlying AI machine learning models outside of the LLMs that are used to help define, right, those kind of prescriptive recommendations. The agents do a great job though of being able to, go beyond just the model to explaining it in an intuitive way And also then going to step b, you know, two, three, and four in terms of, okay, what does that mean for the longer term and what could you do about it? So in this case, it's really about arming that, warehouse management team, so your supply chain leaders with the information necessary to then predict what the warehouse capacity is going to be, where they should allocate. Right? And that gets into a kind of a an optimization discussion. And then where should they allocate inventory, based on the demand plan? So those are the kinds of fees that go in, to something like warehouse capacity improvements. So I'm gonna pause here for another poll question and ask everyone where in supply chain see, we might have some questions coming into, but where in supply chain do you see agents as most impactful to your organization? So we can go ahead and open up that poll. We'll start to see, I believe it's that one. Yeah. And it definitely on by the way, to review the previous poll, it was just getting started. Majority of the people that answered just getting started. It looks like, okay. Yeah. Demand forecasting and supplier management are the two areas that I'm seeing get the highest, votes for impact on your yeah. Then okay. Supplier. And then inventory optimization's coming up as well. But definitely supplier management, demand forecasting, and, looks like inventory are gonna be the top thing. Doesn't surprise me given given the, the climate and the challenges in most supply chains today. So while you while everybody is continuing, to answer that poll question, and I'll tell you that in a a little bit, but keep feel free to keep on answering that. Let's talk a little bit about, where we're going on this demo. So, what we've got is a demonstration. Wanted to show you an example of agents at use for supply chain management orchestration. And the one that you're gonna see is, our supplier risk management assistant. What it does is a couple things. There's different components to it. But one of the things it does analyze is we have a lot of customers that are impacted by tariffs. I mean, everybody's always had some tariff component, right, to a trade they all have trade you all have trade compliance, departments within your organizations. But the volatility and the extent of the tariffs is what's really changed and put a, a spotlight on a lot of these supply chain challenges, and, you know, supplier networks, honestly, challenges. So some of the questions are that I know a lot of you and your organizations may be asking and may have some answers today is what is the impact, of changing tariffs on, you know, current orders on outstanding purchase orders for sure. And then, of course, what's what's the impact in the future, depending and being able to simulate and and analyze that. Also, though, what are you gonna do about it? I want you to know what the impact is. Right? That's always the question on any of these analytics discussions is, well, how do I address this? So one one way to address it is, right, by diversify your suppliers. So I in order to do that, you need to understand which suppliers are gonna be most impacted. So who do I need to prioritize if that's in fact going to be one of the tactics that you employ? What could be adjusted to mitigate those tariffs impacts could be everything from changing sourcing for components. It could also be there is obviously also a price component in being able to look at, you know, increasing prices as a result of the raw material impacts. How do you pass that on? How does that impact your demand? And being able to understand those kinds of, issues and impacts consequences. And then finally, which items are most impacted. So kinda goes along with the so the supplier's understanding who is you know, what I need to prioritize from both, an an item perspective or component perspective or material perspective as well as a vendor perspective. So what, I'm gonna stop sharing here in a bit and then we can go through a demonstration. And it is gonna be here of this supplier management assistant. And really, this is a workforce augmentation type approach where if you're a procurement manager, purchasing analyst, purchasing specialists, you would leverage the supplier management assistant to help you surface up risk from tariffs, also risks from, performance delays among your suppliers, And use that information, based on then the agent will tell you, give you some recommendations of what you could do next. And then you what you do with that information, right, as a as a subject matter expert, as a procurement analyst will be up to you. But it really does help, I think you'll see, streamline the entire process of retrieving information, servicing up supplier risks, and then understanding how to mitigate those risks. So I'm gonna stop sharing now and hand it over to Paulain. Yes. Thank you very much, Christine. I'm checking your second poll, actually, and I see that most of the people are actually interested in the supplier management. So I'm happy because, that's the topic of my demo. So I'm saved, I think. And I will, again, show my screen as well. Let me find it here. Alright. Nope. Here. And this is the one. So what I'm sharing with you right now, it's a Dataiku project in which we have, what we call it, agent connect. It's a, web application, but I will, show you in detail how it's made. And you will see that there is no code involved here. It's, a component of that echo that we are shipping for the one who has this Adam mesh. And I will start with, showing you just few example of, what could be with, what could be done with the superior management assistant. Christine show you already the, architecture, global architecture of it, but I will, go a bit more into the detail of the project itself. One thing we could, imagine on that is to have estimation or analysis on the cost impact, of, tariff. So let's take this example. The user is asking what is the cost impact of tariff on the current orders. And then we have this phase of, reasoning in which the agent is searching into the historical orders to get the tariff rate, by suppliers. So it's doing the operation by itself. And then we got these results, But it's also checking the total transaction cost by suppliers. And then from that, we have kind of a, analysis of the results, after this reason in part. So we have, this result where SpeedLine and GeoPro are the highest, total, transaction cost. And they have high tariff rates as well, and then we can check that, indeed, the high tariff rates compared to others, especially auto supply here. And then the next step would be, what is the cost impact of switching from speed line? I took speed line as an example because it was one of the two, mentioned, to an advantage for the buyer in terms of, of course, tariff. And then same thing here. What we have is basically some, reasoning steps, and then we have some results, which is an analysis of it. So we have, kind of a reminder of what is the duty rate of, this supplier, what kind of parts are supplied, and then, oh, maybe we could switch to a gear pro here. And for example, we have, this example of the headlight here, in which we see that we could do some saving and especially thanks to the, tariff, reduction. And we see that here we have the import duty rate advantage, in in which the gear pro is lower than, speed dialing. And so, we have also, as usual, some recommendation, but still, the agent is very cautious saying that, before any making any switching decision, you should maybe consider, supply reliability, quality consistency, Gear Pro, and other stuff. But still, that's the first step. The first analysis that is done only by checking, the current data and also that is done autonomously, and that is very powerful. Then the user, is also asking if there's any component that SpeedLine supplies that Gear Pro cannot cover. So here, this one is a bit tricky because we are not connected to any, catalog, but it's still checking in the historical data of orders. And for example, we have found, some products that are not available from Gear Pro apparently, especially the water pumps. And so, we got kind of impact analysis. So, yeah, we have, I think, some recommendation that if we wanted to switch from, SpeedLine to Gear Pro, we would need to check for a strategy for sourcing these kind of components and especially the water pumps, to make sure that we can cover it and also, maybe find another, a supplier for these parts. And so that's why here I'm asking, are there other suppliers beside your pro that can provide these water pumps? Because it says here that it's a very one of the biggest problem we have for for this. So same thing, reasoning step, and then we found some, suppliers. So at least in the historical data, we have few suppliers, and here is found somewhere in the bank. I would explain you what the bank use a bit later, but it's found in the bank that Dentsu could be a potential alternative of it. And that's interesting because it's not coming from the historical, orders because, actually, maybe some other, of our suppliers can supply these water pumps, but maybe we have never ordered to them for some reason, like pricing or other. But still, he said, hey. We have the annual report of Tensor, and maybe, they could, supply this kind of water pump. And so here are the key findings and the recommendation we have. It definitely, definitely makes sense, actually, because it says that, okay, Gear Pro does not provide everything. So SpeedLine seems still to be the the most, relevant, supplier, but, also, maybe you should check with them so that he's already one of your, supplier that maybe they could have it could be, or become an alternative supplier. So that's it, for this first scenario. The idea is ready to to show that it's really, an assistant, and it's also performing steps by its own thanks to this, these re reasoning steps. I have another one, which I think is shorter. Here, I'm checking the the lead time the delivery lead times. And so I'm asking, what are the average delays per supplier? So we have some things. I don't think it's, real data, by the way, because, this is huge. I I hope it's not real data for sure. So same thing. It's highlighting few, suppliers that have high, delays in day. Also now I'm asking, the on time delivery rate for the most important supplier. And I said here, it's important, the top 10 suppliers. And then the output of it is a bit strange because we have, you you see some some suppliers have a 100% on time delivery, but, actually, it's not the top 10 players. When I said the top 10 player, I was, in my mind, I had this, thing. I was thinking of the most, important players in terms of, total merchant cost, for example. But here, it did not make it. So I'm reasking the question, on the high spend, suppliers, that have poor on time delivery performance. And now it's including this total spend and the on time delivery rate. And now it's interesting because we see that, SpeedLine well, that's actually the conclusion of the agent that SpeedLine is the highest spend supplier, but as a poor, on time delivery rate. And then we have Gear Pro, which is the second one, but also the second worst on time delivery rate. So we should probably definitely change or, mitigate the risk around these two, suppliers. I have not, continued this, this scenario, but we should hear do the a bit the same thing as previously that, okay, for a speed line, which could be the alternative suppliers to indicate this risk, or, same for Gear Pro, what type of products are shipping, and can we have some alternative supplier, for this, this product, that Gear Pro and Speedline are are shipping. And so here, the idea is really to, find the most the riskiest part of our supply chain and and try to mitigate the risk, which is the classic, problematic of the supplier managers. And the last scenario I would like to show you, actually, is, so the name is not great. Which superiors account for the largest, share of our total spend? So this is the first very simple question. It's just a group buy behind the scene, and we got this result that we already know, SpeedLine, your pro. And then I'm checking if we have any critical components that are sources sources from, a single supplier. And this is important for us because, we see that SpeedLine is a very, important supplier for us, but maybe it's it could be critical on on some parts. And then doing this analysis again, so reasoning step first, aggregating some data from the past historical data. And then, here we have some parts, with only one supplier and the total spend for this, part number. And same thing, we have a kind of conclusion on that. This part, for example, is, quite critical for us. And so, I would like to do a cost analysis. What is the impact if we change this part? And I actually, I ask for the top single source components, to put them to mitigate them around, like, two suppliers. And so here we have this new, again, recent parts, and we have, interesting results, which is the current situation for these three parts, and also a cost, impact, estimation. And and the interesting part is that for this part, we don't, have access to the catalog as well. And so this is we see that this is probably something we should add to this, this project because it's estimating, an industry average, actually, but it's not very, tangible, I would say, figures. It's just an estimation. But still, it's not hallucinating. It's not saying, hey. You would have this additional cost. It's just saying it should be around this, level of, increase, and it's, suggesting few, things like, risk mitigation and, also, try to negotiate the the prices. And here, the idea is really a kind of call to action to say, okay. You should contact other, suppliers for this part, to to try to have a a risk mitigation even though, the the price increase, would be there for sure, but still, that's important to, to have this. And so here in only a few prompt, you have, like, the top three parts you should focus on. Of course, you don't have because it doesn't have the data, but the important part is that it's not originating. But you have order of of many tools of standard, industry average. But, also, you know, three parts that you should focus on if you want to do some, risk mitigation over, several suppliers. And now I will a bit, deep dive into, this, this project. So you probably know that they could already. So this is, a classical project. And as any classic project, you will you want to go to the flow. If I go, yeah, zoom out a bit, we start with some input data. So, basically, what we put here is the supplier orders, and we also added the duty tariff into it. So we have this table in which, we have, some, parts and some orders for, each suppliers. We tried as much as possible to add some, description of columns. I think this one is better. Yes. So here, for the agents, it's important to know what part numbers, refers to. And so we try to put some description of the column everywhere. So it's really helped them, limit the risk of vaccination and also have some better results. What you see here in the oral history preparation is that we know from the user's question and from the use case we want to cover, we know that the agents will need some aggregated, data. So for example, I will take this example. It's the for each part number, this air filter, which which are the suppliers and what is the average price. So we can see the difference of prices, for the same part, along the different, suppliers. So we know that this kind of data preparation steps will be required for for this, project, and so that's why we already did them into the project. The idea is to avoid the agent of doing this operation by itself all the time. So it's really speed up the process. We have also, described the the table. So we know that the agent knows what's inside these different tables. So the idea is to have all this table prepared, and now the agent is ready to answer those question, and it goes, faster. And because, overall, in this kind of use cases, when a user is prompting an agent through this, web app interface I show you, it's very important to have, I would say, decent, at least, response time. And so this kind of data preparation really helps, on that part. And then, so here, it's for the demo. We for demo purpose, we we recreated the demand forecasting dataset, but, we we are we asked the, agent to use only the the scenario one, which is for each part number and each supplier, the the forecast quantity, estimated. This is just to show that also the, I don't show that by the way in in my scenario, but the the the agent could leverage this, demand forecast. And if we go back, to the flow, we also have this, agent part. So, actually, we see that we have several agent in in that project. The first one is this one is an agent that is able, just to, perform some simple SQL, queries and to actually check the different datasets. So we have limited, the datasets to the one we really want to, to to make the the agent to make some search on it. And then we have this other, pink part in which we have uploaded some annual reports from the different, suppliers. And from it, we have created oops. Sorry. We have created a knowledge bank. So we have this recipe. But without any code, we are just extracting the embeddings from the, you know, annual reports and storing them into what's that they could call a knowledge bank, which is basically a vector store. And then from that, we have, defined here a supply chain risk agent, in which we have several tool. So we have the main LLM of this agent, and then we have two tools, the knowledge bank search, with a few, short description of this knowledge bank, what is in this, knowledge bank, and when should the agent use this tool. And, also, we have a dataset query, so the the agent so this agent can call the other agent as a tool, to, perform some, operation around the, supplier holders, tariff, duty tariff rates, and all this data available in the in the project. And from that, we are using something that is prebuilt in Dataiku now, which is called agent connect, which is this pretty simple, chat interface you have seen. And to to create that, you need to go to new web app, and it's what's called a visual web app. So this is something that already exists to the. So you don't have to code anything to have this, or what? Let's call it a chat, chatbot agent, with agents, and then you just have to, add it to your project and configure it. So you have to, of course, in RNN, main RNN, which is the one creating the Anthos, configuring few things around historical data set and stuff, and then, you choose your agents. You can have several agents, and then that's it, actually. Then you have few other configuration because you can add some, preselected question like this one. But, overall, it's quite easy to do. In just few step without any code, you can create, this, Super Agent assistant really easily. That's really the strength of, of, data for that, around agents and and that. I think I've covered all everything I wanted to show you, so maybe you can stop sharing and go back to question if you have some. Yeah. Thanks, Romain. I think, yeah, we do have some questions in the chat. I've tried to answer some of those. Cool. Yeah. But I think there were there was thanks, from Alexander. Thanks, Romain. This is all based on index of Esrij's data or there is is there a text to SQL live search over some data like? That was a question we had. Sorry. I missed the question. Could you say it again? Yeah. From Alexander. Thanks a lot. This is all but is this all based, sorry, on index vectorized data, or is there a text to SQL live search over some data lake? Yeah. It's performing some SQL, to both, actually. Yeah. Because some of the data are prepared as I show, and some are, live, actually, SQL queries. So that's why I'm I'm using two agents. One is here to do some, SQL, operation, I would say, And the other one is here to, do the knowledge bank search and also to, prepare the the answers and do the, of course, the reasoning part. Okay. Hopefully, that let's see. I think we have a couple other questions here because we were done with the demo. Sorry. I was gonna, go through a few of this q and a. And we have, what what does see, does your agent directly connect to all ERP data? Do we need to create a middle, virtual layer for data that is available for your agents? In this particular okay. I guess, in this particular instance, we're using ERP data, for that demo. And I would say that, in terms of all that's a you know, I have to say it's a pretty inclusive term. I you wouldn't need an agent to connect to all the ERP data. It just depends on what you want this agent to do. In this particular instance, we're only really leveraging, vendor master data, supplier master data, and PO. In this case, it was actually open PO type, the history, I think, for the last three months. So hopefully, that answered that question. Okay. Let me see. How was the knowledge bank created? Did we I think did we cover that part, Homa? Yeah. Ma'am, maybe I cover it a bit quickly. But, basically, to create the knowledge bank in that I could it's very easy. You just upload some files, into a a folder, into your flow, and then you have this recipe. I don't remember the name of the recipe, but you have this receipt to create a knowledge bank. And, well, you you definitely have to select some models, some embedding models, but that's very, easy and and without code. So this is a quite an easy step. I think I don't have the details here. I think yeah, I'm checking, but I don't have the details of the, embeddings and stuff. But, really, that's something, that is easy to do. Of course, you need a embedding model, and you can you can select a different type of vector store, but this is really something that has been automated in the tycoon now. And then, Ron, we had a question about I'm gonna have a couple of try to tackle a couple of these. Are you using RAG and MCP for queries? Not MCP, because here, it's only into the, so it's not using MCP. It's really generating some SQL, queries. And, yes, we are using RAG through this, knowledge bank, creation. That's what's what do we call a rag, actually? So, this, tool and main, knowledge bank search is definitely, the the rag, agent, actually. Yeah. It's kind of a tool of an engine, but yeah. Okay. And we have a couple other interesting questions. Harry asked about risk of hallucinations. Was the likelihood of hallucinations? How do you minimize that risk? I'd still be a little uncomfortable at even ninety nine point nine percent accuracy. It sounds like it. I mean, that's that's common in manufacturing supply chain for making important decisions. Well, you could I can talk on the business side. You can talk about the Yeah. Manage hallucination risk. Yeah. Thanks. So do you wanna do you wanna go first on hallucination? Yeah. Yeah. Okay. Start step. Okay. Yeah. On the business side, I'll say that that's one of the reasons you if, for those of you that are here at the beginning, the majority of when we talk about agents and use in supply chain, the majority of it is for workforce augmentation, not for actually supplanting some of the decision making that your, you know, purchasing leaders and specialists are gonna make. I'm a strong and and I do a lot of, by the way, you know, core, like, manufacturing the make area use cases as well, and I advocate for the same thing there just from a trust perspective and, just the risk aversion culturally that happens with any large enterprise. So from that standpoint, it's still, like, one of the recent projects we worked on for we actually started with web scraping for optimized pricing for our suppliers, but then moved into, using some API approaches for approved suppliers. But even that would not necessarily, like, initiate a purchase order, for example, from the lowest cost supplier. What it does provide is the information and a link to that, buyer, for example, saying this is the lowest cost for this particular part. Now if you wanna buy, here you go. Hopefully, it said that from a business perspective, I guess, administrative control perspective, that's the way you would help mitigate some of that risk. But, Homa, you can talk about hallucination management now. Yeah. But I I also write, I agree with you actually that, overall well, we are in 2025 for the record. Gen AI today should not be the final decision maker, I think. So it's a decision support layer for sure. It really accelerates, analysis. I have shown flag some issues, but I think humans still validate before acting. I think it this is very key for, for today's, agent. And for yeah. To limit the allegation, I think it's also first depends on the model, used. For sure, I think I I think that also in the question, but I think it's important, not to let, the model generate sensors from, the open web. I think it's better to ground the model on curated data sources and also to control the scope just like we did here. Actually, we we we could say, okay. Search for information on our, suppliers, but it would be higher risk of hallucination. And I think it's much better to like, here, we uploaded some structured and, also, I would say, controlled, documents. So for sure, you have limited information in this document, but we know that this is a good source of truth. And so that's why I think it's important, to limit to this kind of document that I would say you, but the the the person who is creating this project is is mastering at least that just make sure that these information are correct and and verified. So that's very important. That's why I also created this rag into the project to answer that to ensure that these answers are are tied to the source document. Something I haven't shown right away is that we also always have the source document. So when the agent is answering something on a document, we have the source of it. So that's super important and really helps also, mitigate the risk. And then we have, well, we have some for production, we will have some validation steps. So, let's say we have a sample of maybe 50 question that typical question could happen in the chat, and we just check that for these 50 questions, the answers are correct. And, also, on top of that, knowing that it's only checking to we put that every everywhere in the in in the different description and in the different prompt inside the the product is that, if you don't know, or if you don't find one, information or another, don't use it. Just say I don't know it. And this is super important to add, this kind of guardrail a bit everywhere in the project to make sure that we limit the risk of, hallucination. So, yeah, I think that's that's a big topic, but, overall, this is how you limit the hallucination. We have a couple other one is, Rania, how is the agent better than a rag? It's not, better. It's, it's, yeah, I've seen that. It's not better. Agent is not better than the reg. It's, different. The reg is actually, it's a pipeline, for retrieving some information from documents and generating for it. And the agent is is different. It's it's adding, reasoning and orchestration, step. So, basically, in a rag, you are just doing generation from documents, from information, where the agent is the idea is to add reasoning. After this reasoning and orchestration step in which the agent decides to call, for example, the information from the rag because, from the description, the agent thinks that this result or this information should be in, the rag part. So that's why we use both rag and, and the agent. Basically, in the RAG, we are answering question that the agent, has, prepared. And so the agent generally solves some high level task where the RAG, answers questions. Great. We also had a client, and I think Thomas had a question about I would love for you to talk about this for a while. Are we able to have the agent format, the output as a graph rather than just an answer? I don't see. Yeah. We could have that. I don't think it's possible because as I said, what we're using is called HNconnect. And, so it's, built by Dataiku, and I don't think it's, available yet in, in the and I'm not sure even if it's still the road map, by the way. But, for sure, that's, something that could be possible by the models, not all of them. So you have to pick the the right one. But I don't think right now it's possible in in to the take home. Okay. Pavan also asked a question about chat history, user profile datasets in the web app. Are they created automatically, or do we have to create them before we create the web app? The profile, you said? Yeah. Remember the the you because you have the chat history, user profile, those datasets. Oh, yeah. Yeah. I'm not using oh, yeah. You have to create them before create the web app. But at the beginning, you can I think you can create them right from the configuration page of the web app, actually? And so, yeah, I think it's done automatically, and then, yeah, it has the right columns. So, basically, it's just creating a a new SQL database, data a table. Sorry. And so then you you have your web app set up. I think you need the history and the user profile, datasets, these two and but it's really straightforward as long as you have a secret connection. K. We also had a question around, do we have any standard scoring standardized scoring mechanism? I assume they mean, assume that's meant for supplier scoring, and the short answer is no. We don't have a yeah. We don't have a standardized what you saw was representative of, you know, the the reasoning power of the agent based on, you know, the annual report information, but also the the PO history and, on time, delivery performance. Yeah. But if in your company, you have, kind of standardized scoring mechanism, it's really easy to add it, actually to some table and then to, precompute it just as I did some data preparation, in some tables. It's really easy to add it and to, use, the agent, refer to this scoring system. Also, in a question, we can, see. If I came in operator so if somebody were to reject, like, say, one of the recommendations you'll provide feedback, I think is essentially here what we're going with. Low cost supplier, give reasoning back to the agent. How would this information impact future reasoning suggestion from the agent? I I will tell you I mean, actually, Houma and I could talk about how I kind of, I guess, audited to some extent, right, the early days when we first put this together and how we we collaborated on it. Dataiku Dataiku, that is actually an important part of of where we're going with our, yeah, our our product road map and adding a more kind of intuitive governed way of allowing SMEs or supply chain professionals, in this case, purchasing analysts to audit, provide feedback so that that could be used to improve, I guess, some of the prompt engineering as well as, the approaches from a algorithm standpoint. Romain, you could talk a little bit more on technically how you would incorporate that. Yeah. Right now, I'm not sure that by default, you have this kind of, I think you have this chat memory because you have the chat history that is saved, but I'm not sure it's, generalized for the model. So this is, something that would be would need to be implemented, but it's a bit more advanced because we want to if we want to share this memory or this information to all the users, this is something that need to be a bit more it's something more advanced that we have not, developed yet. I think you could do it in the table, but I would definitely use some code to maybe fine tune or maybe enhance at at some point the the the the agent. So probably fine tuning is not the right answer at that stage, but we would have to think of of this, this option. Yeah. I think some of it, on a product sample or was around some of the prompt engineering aspects, you know, in in improving the instructions. Yeah. Also, also, that would work. Definitely. But this is would this would be something, more manual than, you would not be, have something automatic, for this, improvement. K. I know we only got one minute left. We bring if we bring the standardized score, how do we embed that? This is a kind of just a general question there, but, we bring the standardized score, how do we embed that inside your tool so that would become part of that input, Right? And that data set. And so yeah. It it what Romain showed you, obviously, we constructed that or was an example from some, customers that we worked with. But, that would be another input potentially. Right? One of those recipes that or that that workflow, the analytics workflow. You would have your supplier score there, and you would just inform the agent that that's, the data source for scoring. Yeah. That we could also think of, another tool, that could happen. We can use create any other tool to evaluate, standardized, score. But, yeah, this is not something we have implemented so far, but could be something that, could create could be created quite easily in in that table for sure. And and, yeah, I apologize. I know we've got a couple more, questions around, but I know we're at time. So I wanna be kind of conscious of that. I'm happy we're happy to follow-up because as you, I know Alexander is a good one around AI and supplier management. Actually, we see a lot of interest in that just for some of the reasons I mentioned in terms of what's going on with supply chains today. And then we can also talk about leveraging AI to yeah. That yeah. That's some of the data consumers, business users. Absolutely. And that's a good focus for where and when an agent might make sense because we have some questions around that. But, I appreciate everyone's, time, consideration, and input. This has been great interaction, so we really appreciate it and hope it's been helpful for all of you. Thank you so much. Yep.