Welcome everybody and nice to see so many joining. I know people are still taking in and thanks for all the input in the chat already. Good ideas. So let's get started for today where we from Implement would like to share our perspectives on where AI is actually taking supply chain planning in the future. So my name is Søren. I'm your host for today. This we are attempting to make not another AI hype session but more focus on a viewpoint on where planning is heading with AI and giving you some practical insights in what you can do. So what we would like you to be able to take home from today is a clear picture of how APS systems or systems and agents and planners should work together in the future. So what we would like you to introduce how we implement this, what it takes to implement it, introducing you to some of our key design decisions on how to address implementation. And then we have a special guest that will also give us some first hand insights on how UPM started their AI journey. So let's talk about who's here today and who will talk to you. So besides myself, we have, as I mentioned, a special guest star, Timo Zeppala joining us from Finland, from UPM. And he has started on the AI journey and he will share in the end of the webinar here some insights and experiences from that journey. Before that, we have Mike, who is an expert on planning and AI. He will walk us through how agents can help and he will also show us a demo with an example of this. We also have time for questions in the end. So if you have questions, please drop them in the chat. I know we can already see you can use the chat, but we will try to take some of the questions in the end. There's a lot of people joining today, so we might not have time for all the questions, but please put them in the chat and we'll try to pick it up at the end or otherwise try to follow up after today. Good. So let's get started on the topic for today. So here's a pattern that we see across many organizations. We see that planners are spending 80% on making the plan and only 20% on managing the plan. High performing organizations have flipped that. So they are working on managing the plan to get great output of the systems and getting the systems working for them and using that output. So if we double click on that a little bit, so it's not necessarily a bad thing spending time on making the plan when it makes sense. But often it means that the planners are really overriding the results from the systems because they don't trust it, because it's wrong, at least in their opinion, or because the input was wrong. So we had poor quality. So that stops us also from benefiting from all the effort we have built into the systems. What we would like to do is we would like to lift planning where we manage the inputs so we get great results out of the system that we can use and only override where it's necessary. And this is one of the struggles that we have in all planning implementations for the last many years. But it's also where we think that AI agents can actually help us. So instead of replacing this logic, we could actually make planning better with agents, taking a step forward and solving this problem that many companies have struggled with for a long time. So this is where I will bring in my colleague Mike, who will talk about this and maybe give us some advice on how to make this work. Yeah, thank you, Son. So before I start with the viewpoint, I would just like to make sure that we have the same understanding of the terminology used today. So Son also mentioned it in the beginning. So when we say APS systems, that's advanced planning systems, meaning systems that have the purpose of generating a demand plan, a supply plan and inventory plan. And when we say LLMs or large language models, that is basically the model architecture that is used for agents to be able to do reasoning. And it can understand, it can generate text, and it can also reason across unstructured data. So that's basically the technology side. And then when we say agents, agents is in essence just an umbrella term, but it can take into account context. It can decide on an action, it can execute it, and it can observe the result and loop back. And then within agents, there are different degrees of autonomy. There are, you can say, the low autonomy agents, which we refer to as virtual assistants. And what they do is basically they are supporting planners with a specific task, but at the end of the day, a planner decides. And on the other scale of the spectrum, we have autonomous agents. And that is basically a swarm of agents that can execute multiple tasks and decisions after each other. Right. So both of them are agents, but at different degrees of autonomy. So how we see the future of planning is that we see the future being layered. You will still have the APS system or technology that is supposed to generate the plans. You will have the agents as an interface towards APS systems. The agents will not replace APS technology. They will sit on top of it, turning its output into insights and improving what goes into it in the first place. And finally, we have the humans. And this is the part people, or at least the software providers, skip paths too quickly. So agents can't remove accountability. So even though you can outsource a task, you cannot outsource accountability. And the whole layer only works if all three systems and humans are working together. So you can make the plans in APS systems, insights from the agents, but accountability stays with the people. So how we envision it is that you still need the APS systems where you have the demand plans, the supply plans. And what's really important here also is the link to execution. So you are actually able to execute the decisions you have. The agents exist in various forms, either virtual assistants, autonomous agents. And then we have the humans, where we see the humans and future planners as being the decision architects. And what we mean by that is that it is still up to us humans to define what are the business rules. It is also up to us humans to define what are the guardrails and decisions that the agents are making. And finally, it is also very important that the humans define and monitor the assumptions that all the plans are made on. So as CERN also alluded to in the beginning, we see that a lot of companies are implementing advanced planning models, APS systems, but it's not all that get the full value out of the effort they do. Because in essence, to measure the true value of an advanced planning implementation, on one hand you have better plans, so a more correct result. But on the other hand, you also have the business adoption. So if 50% of your forecast or 50% of your supply plan is overwritten by planners, you will not get the full benefit of advanced planning models. We see there are fundamentally two reasons why this is the case. One of the reasons is that the advanced planning systems are producing the wrong outputs. And that is not due to that the algorithm or the machine learning model is wrong. It is more often due to that the data going into the model is wrong. Or at least that the data has drifted or changed since the model was implemented. And here we really believe that agents can increase the business adoptions, because you can use the agents to improve the input going into the models. And it can question it and it can synthesize and making sure that when the model starts or when you start planning, the data is in proper place. Another area where we believe agents can help in supply chain planning and increase business adoption is to explain the output. Because one of the other reasons we see that the full business adoptions of planning systems is not there, is that the outputs are right, but they cannot be explained. So if a planner has to explain to all the stakeholders why a forecasting model is 100 and not 150, if they can't explain that, then they cannot, you can say, put forward the plan to all the stakeholders, and then the planner will start overwriting the plan. But the agent can help explain both to the planners and to the stakeholders via a certain model decided to be 100 instead of 150. So in other words, we really believe that the agents can by itself increase the business adoption and improve the input to the plans and thereby increase the return on investments of advanced planning systems. And if we take a double click on planning, then there are normally three planning domains. We have demand planning, supply planning, and inventory planning. And we still believe that all the algorithms to make the plans belong into the APS systems, being it machine learning forecasts, network optimization, or multi-ethylent simulations. But before the models run, you can put on agents to improve the input. So for demand planning, it could be that you have a product lifecycle assistance that ensures that all the face-in and face -out profiles are maintained correctly, it can flag anomalies, and also ensuring that other data patterns are correct. It could also be for the inventory planning that all the assumptions you made for your inventory calculations, it can track whether you are actually following that performance that you expected, or if something is drifting. And on the output side, you can also create assistance here to help you. It could be an assistant to prepare the demand review meeting, it could be an explainability assistant for supply planning, or for inventory planning. So again, the assistance by itself might not be groundbreaking, but it's part of the very important stuff to making sure that all the plans you make are actually being used. But this is also the important steps, if you would like then to have a future autonomous supply chain. So, an autonomous supply chain will not happen overnight, and it shouldn't. We really believe that you should move stepwise. So, today, humans, they are fully in the loop on every step, meaning, before the demand planning or forecasting runs, the human reviews the input, and the human also reviews the output before you have a demand viewer meeting. And that's the same for inventory and supply, and for the SNOP decisions, right? What we believe is that already now, you can start improving the input, so you only have humans judging the output of the plans. And in the near future, we will believe it will be humans on the loop, so you can actually have autonomous preparation of the data, you have automatic calculation of your plans in your APS system, and you can also then have agents evaluating the output, and only escalate to humans when needed. And that is then basically the path towards a fully autonomous supply chain planning process. But at the end of the day, some of the decisions you make in your SNOP process, of course, can have a big impact, so thereby, you still need decision makers in the loop at the very last end. So that is why everything cannot be outsourced. But what is then the new role of the planner? The planner is no longer about making the plans, but they are now responsible for the decision rights, and to make sure that this process flows. Okay? Good. So, I prepared a small demo of how an agent can support you. And here, I will bring a demo of how to prepare an SNOE meeting. So, in an SNOE meeting, there is the preparation part, there is the execution part, and there is the follow-up. But we also know that SNOE meetings, which happens every week, are often under a lot of pressure and a lot of decisions need to be made. So, therefore, we show how an agent can help prepare that. So, I will just find that demo here. Good. So, I will have here an example of an agent demo. So, on the left side here, you can see this is what the planner will see. So, this is the SNOE cockpit, where you have the jobs waiting. You will have an agent chat, and so forth here, which we will come back to. Then, on the right side, because it is a demo, we will show here all the steps that the agents will take, so you can follow this. Good. So, now in this case here, there is first a job trigger. When should the agent start? In this case here, this agent has been trained to learn that two days before an SNOE meeting, the agent should start preparations. So, that's basically the trigger point that is this calendar invite. That then awakens the agent, and it starts the tasks that it has been asked to do. So, that means it first needs to pull the right data for your various systems. So, first of all, it needs to have the data from your APS system. So, that could be what is the plan status, what are there any exceptions, are there any open demand, are there any supply shortage and risks, and are there any disruptions on the way. So, that is one data point. But, in this case here, there are some exceptions that need to be handled, but it's not clear which customers you need to allocate to. So, it also takes into account what is the data from our CRM system, and what tier is it, and what is the contract here. But, as you can see, there is not necessarily a unified way to prioritize between these customers here, since the contract cannot be compared. Then, it also reads the action list from the last SNUI meeting, because that is also a way to follow up that all the actions have been completed. And then, when it has extracted all the data, it starts analyzing it, and if there is something that exceeds the guardrails, where the agent cannot take decisions, it then escalates that to the planner. So, in this case here, there are some findings that they would like to email to you as a planner. So, in this case here, you have the SNUI cockpit. So, now there are the findings here. There are two decisions that need to be made. It generates an exec summary, gives a brief overview of the demand signals, supply and shortage risk, the allocation conflict here, so which customer should we prioritize, and also go through what are then the open actions from our last meeting that are overdue. That means that you, as a planner, can go through this, but you can also basically see what are all the steps here that the agent has taken, and you can use that as an interface to ask follow-up questions, or ask the agent to perform additional steps. But once you, as a planner, are happy with the results, and you approve the findings, then there is another agent that is then called, that starts drafting the email to prepare for the SNUI meeting. When that draft agent is done, it here proposes a mail to send, and you can then distribute that to the meeting invites. So, this is one example of an exercise task that we know a lot of companies are doing every week, which can be automated, but also actually help support or improve the decision made. So, this was one example, but of course that requires some implementation. So, Cern, I will hand this over to you, so you can talk about the implementation. Thank you. But first, Mike, can you explain the demo we just saw? Yeah. Is that a finished product or what? So, what you just saw, that's an example of a demo we have. It has access to our own APS system. It can generate emails and generate PowerPoints. But this is, you could say, technology we implement as customers. So, what we then do is together to define what are the skills needed for the agents, what are the connectivity, and also what kind of technology is needed. Right? So, it's not a product, but it is stuff we implement at customers. Okay. Yeah. So, this implementation, can you elaborate a bit on that? Yes, I can do that. So, building on what Mike is saying here, you have seen the demo. And then, we think that when implementing AI agents, we should consider that this is not an AI project. It's a three-dimensional implementation challenge. And those three dimensions are covering some of the topics we talked about. But first of all, this is new technology. We are learning about agents, but we also have a layer below that. You just heard Mike say, needs to be connected to our APS system. We need to have other data. We need to find a way to get templates in there. So, it's not necessarily on existing infrastructure. So, we need to make that. And we also have a new dimension that there's a cost of use, tokens. So, we need to accommodate or consider that in the implementation. So, that's the first part, the technology part. The second part is, this is also a shift of operating model. In a way, we can say, this is the same as we are hiring a new planner that is sitting there. And we need to hand over tasks to them. We cannot hand over ambiguous tasks or unclear tasks. It needs to be something clear. We're giving to the agent, so the new planner. We need to know, understand what they're doing. We need to give them guidelines. We need to figure out when do we escalate. When are they okay to do this in an autonomous way? And how do we use the result? You could also say it's the same as moving a planning process to a shared service center. And we all know that is a difficult task. The more process steps we move, the more complexity it is. So, the more we want the agents to do, we also need to be clear on what are the steps we are passing on. What are the steps we are sending to the new planner or the new shared service center. So, that means also governance and escalation paths. They have to exist and they have to be clear. And then lastly, back to the trust and adoption. So, if we just give the agent this and we are getting a result out, but the people with accountability do not trust what the model brings out, they will not use the result. So, then we have an agent that is corresponding to the new planner we put in the corner or the shared service center delivering a result to us. But we will have the same adoption challenge as we have had before. So, we need to make sure that the people using the results, they trust it. They know what the agent is handling. They know when the agent is escalating something. And this is not some trust is not something we earn overnight. It's something that we need to spend energy on as part of the implementation. So, we also think that one of the foundations for all of this is some of the foundations that we have struggled with for many years. And that is establishing clear business rules for our business. So, what you see here on the hexagon slide here is the example of some of the business rules where you see on the right side we have the customer dimensions, the service level agreements, the customer segmentation or prioritization and allocation principles. And on the left side is supply chains response to that, the end to end segmentation, planning concepts and policies and inventory policies. And all of these should be clear. But yet, at a recent conference when we asked how many have clear allocation principles in terms of shortage so clear that they can automate it, nobody raised their hand. So, we are still struggling with that. So, it's difficult to give that task to an agent if we cannot write down the rules. So, we still need to get these basics in place. And maybe we can get the agents to help us with this. So, this will be the foundation for what we call the decision rights and the guardrails that we give the agents. This common foundation that we can share. And that leads us to the model that we have defined for implementation of agents. So, we have four areas. Eleven key design decisions that are addressing this. We have the IT infrastructure part. We have the agent construction itself with the guardrails, etc. We also have the knowledge and the tooling. And here, what you saw in Mike's demo. There's some specific knowledge that we need to apply. How does the S&OE process work? What is the context? What is the standard agenda for the S&OE? And how to interpret data, etc. And then we have the governance part with the decision rights and the skills. We teach the agent on how to send mails, how to identify an issue, etc. So, all of this is not a complete. But this is the key design decisions that every implementation within planning should address. So, that's the conceptual part. Now, I want to turn to someone who has done something in real life. We have Timo sitting in Finland joining us from there. And Timo, you can maybe just introduce yourself and Upjom a little bit further and tell us what you have been doing. Yes, thank you very much and thanks for the invitation. Hello everyone. My name is Timo Seppalä. I come from a company called UPM Kymmene, based here in Finland. And I am part of the business called UPM Fibers. We operate a fairly complex value chain. So, we source our wood from different forests in Finland and for example in Uruguay. We process that wood then through different assets and produce sown wood for construction, energy and other byproducts. But then our main product, wood pulp, which is the raw material for paper, tissue and printing paper and packaging and such. So, today's topic is about this optimization challenge. So, my job is being responsible for maximizing the value from all of this. What I just explained. And we do have advanced planning systems for all of the three, let's say, businesses. So, wood sourcing, timber business and then the pulp business. However, on top of that, we have created in-house, coded ourselves and an optimization system that optimizes this whole. And which is a really, really a complex thing to do. So, what we are doing today then is we are utilizing AI agents and LLM to help pretty much in similar manner than Mike and Søren were explaining. So, first of all, we are building an agent to validate the data. All of the assumptions we have are using it. Again, we have a massive amount of master data and assumptions when we do different kind of simulations and scenarios. But we are also implementing an LLM that then helps us to be more efficient in understanding and explaining these different optimization results. And that's pretty much in line with the 80-20 that Søren was talking about. We are not changing the process. We are not changing the algorithm. But we are trying to be more efficient and faster in doing this. And so far, it looks pretty good. It's really interesting. And the speed of development is pretty fast. But that's where we are. Thank you, Timo. So, maybe if I can ask you here. Of those three implementation challenges that we discussed before, so with the technology operating model or the trust, what has been the hardest for you? What has been the most difficult? I think that's a really, really good question. And I think the most hardest one is the same than whenever you implement an APS or optimization or whatever. So, it's trust. So, it's not a challenge. When you have good people, it's not a challenge to build something, create fantastic technology. But then what comes out, that ensuring that people trust it. When you use massive input and massive amount of constraints, it's pretty easy then to question something. But hey, why is this? Is this really so? So, it's really the trust that is the biggest challenge in my opinion, not the technology or something else. So, have you managed to build the trust now? And what have you done here? Yeah, I think we are, well, of course, we are not perfect there. But I guess the key there is that we have good people who understand the technology, the AI that is moving forward so fast, all of the LLMs. And we have people who understand the business, where the value comes, what are the key challenges. So, when we combine these people and put them in the same room, ensure they talk the same language, understand each other. That is really the thing. You have also spent time and energy on this and putting people in the same room, as you say, to establish that. Yes, yes. And I guess what I mentioned, this optimization system, that's something we have been working already for a year. We are building it ourselves. We coded it, all of the code and documentation and everything is ours. So, we know it really, really well. So, therefore, it's easier perhaps then to explain also to the people who don't know that system so well. And again, that's one of the, I would say, one of the value of having this kind of a tailor-made system is that then you can input the code and the documentation also for the LLM. Because when it knows how the system works, it's better to explain the results. It's easier to find inconsistencies in the data and so on and so forth. But basically, we are just continuing to work with the help of AI we have done for almost a decade. Yeah. Yeah. So, you already had a good foundation to stand on. Maybe, what's one thing you would tell others that to get right early that you would have liked to know in the beginning? I think there are a couple of layers to that answer. I'm pretty sure everyone is familiar with the fuss and the hype around AI today. Everyone is speaking about AI, these agents that to kind of really focus on what is the business challenge, what is the value adding challenge and then asking if AI can help improving in that. So, kind of managing the expectations and going to business value first and then asking if the technology can help us. Not the other way around, but hey look, this is a fancy LLM or fancy something else. And kind of first things first. That's really the case. Okay. And you had a question for us earlier that we agreed that you would casually ask us without preparation. Maybe it's time. No, no. Because the presentation that you made, which is really, really good, kind of got me thinking that all of us, the professionals in the audience and everyone in this field, we talk about the maturity assessment. How much is this and this process, how much is your SNOP and so on and so forth. And we have been building on that foundation and developing to get to the better steps of maturity in one form or another. So then the question is that how will all of this change that kind of the whole assessment and whole, let's say, scaling and dimensions of it. So that's an interesting one. And thanks for giving us the question earlier. So we had time to think and now we can pretend that we are quick thinkers on this. I think, so of course, we discussed this in the maturity assessment. I think that there's some elements of the maturity assessment that we would do or the maturity that we are working with. You know, the assessment itself is not important. It's just a tool to show us a temperature, right? But it's very clear that some of the foundational elements that we would normally measure. So are we following the same rules here? Do you have nine demand review meetings that are acting in a different way or are you at all adhering to the process? And the data, are you using the data? Are you trusting the systems? All of these will be even more important when we try to put agentic AI on top. But I think also when we hear Mike's demo, you could also say maybe there's some shortcuts in the maturity that we can do. Because if we use the agents to improve the input data, then one of the traditional problems that we had that the data is wrong, we cannot trust the data, nobody is monitoring or maintaining it, we can use the AI to actually skip a couple of maturity levels. We could also maybe use AI to skip some places where maybe the markets are doing demand forecasting in five different ways. But maybe we can apply an agent that helps us harmonize the input to the system. So maybe it can help us, you know, skip a maturity level or two. And then maybe last, if we should make a maturity assessment, I think we would probably add a new dimension now to say, you know, application of AI agents. We're starting with, are we using, you know, co-pilot to do meeting summaries, but later there's virtual assistants and agents or autonomous agents that are doing actual planning task force. That would maybe be the step ladder we would add in this. That resonates quite a lot. Me personally, I think it was one year ago, I think I said it aloud that kind of starting to implement AI without kind of having bulletproof validated data. makes absolutely no sense. Now, now, like exactly like, like, explained already in this, that if we are able actually to utilize AI and the agents. To validate that is something you just pointed out that maybe it kind of helps in, in the matter it is as such. So it's interesting times we are living. Yeah. Thank you. And Mike, do you, do we have one or two questions that we should discuss here at the end? Yes. One of them maybe you already talked a bit about, that was data quality. So, so I think you touched most upon it. I don't know, Timo, has it been an issue with you with the data quality or anything? What is required to get there? Obviously, it's, it is an issue and it continues to be a very, very important part. But I, like I mentioned, I believe the agents that can be utilized to check the data, to validate the data and really use all of those methods really, really efficiently and much far more efficiently than, than human beings. Not only in the input data, but then when you get results and you validate the results of, for example, optimization. So, yeah. But it's, it's obviously, it's still a big deal. Of course, I mean, from our perspective, we also think that it can improve the data quality. But of course, if, if you don't have the lead time or if you're missing the bill of material, right? That fundamental stuff you still need to have in place. But maybe if the lead times haven't been updated for a long time, you can create an agent and actually compare actually lead times versus planned and then see, okay, how should you then update it? So, yeah. Yeah. We have one more question, Mike, that we should do. Yeah. Yeah. How to make sure that AI correctly analyze the impact of new events. So, we're living in a world with a lot of disruptions and if an AI agent should help you here, how can you actually make sure that the conclusion that the AI agent gets to is correct? Is correct. And what would you say? So, I would say here it's important that you distinguish between what the AI agent does and the APS system. Because I then think, I believe that the AI agent can transform that into a digital twin in an APS system, simulate the consequences and come with an output. But I don't see the AI agent do all the calculations, but it is turning the new event into a scenario. And then it can compare the new event scenario and the baseline scenario and then come with a recommendation. But then it's still up to the planner to do the final adjustment. So, therefore, I think it's important that the AI agents, you could say, yes, there has been historically that they can hallucinate a bit. But I think both of all that has been improved. But you also only ask the agent to do the stuff that they are good at. I don't know anything to add. No. Do we have one last before we just have a message here at the end to summarize? Yeah. So, it's a bit connected here, but how can the AI driven supply chain architecture empower the planners in a world with frequent disruptions? And another one asks, like, the whole, not issue, but with many planners are spending the time on is to identify an issue and find the root cause. How will this make the life of a planner easier? Yeah. And here, maybe I will point back a little bit to the demo Mike showed, because we saw that we had an agent that was connected to our APS, meaning it has the data from the APS. It was also connected to the CRM, so it could actually look up some additional information, some of this data. We are paying expensive implementation money today to implement data or interface data between systems. But if we can have this in a data layer where the agents can access from multiple, we can actually ask it to reason or perform the same effort that a planner would do looking up data in three different systems. So, I think this is at least a way where the agent can actually be trained to do some of the analysis for us. I see we have a question to you, Timo, and I think we can just have time for that, right? So, it's from Andreas here. So, that understood that matching the people with tech and business understanding was key for success. So, do you have internal tech people or was this relevant for success or what kind of capabilities was needed from your end, both on tech and business understanding? Yeah, we absolutely do. So, like I mentioned, we have been developing this optimization system already years. So, we have people, data scientists who actually be happy doing the coding. But then we also have very, very disciplined documentation around it. So, even though these people might be leaving to other responsibilities, the kind of knowledge is there. But that was kind of one point I made, that we are able to give that documentation and that code to the AI and actually speed up even the development ourselves. But then coming to the fundamentalists, that it doesn't really matter what the system is, as long as the people who know the system and the people who don't know the business, I think that is the critical thing that they communicate and do things together. Good. Thank you very much, Timur, and thanks for joining us today and for sharing your experience here. So, you could say, wrapping this up or maybe trying to highlight our main conclusion here again. We think that we are in planning, we have struggled for many years with the problem that we are making the plan manually. The planners are overriding, not because they are trying to do a bad job, but because of the reasons that Michael was explaining. The input data is too bad, so the output does not have the right quality or they don't trust or understand the complex results that they are getting. And this is where we think that the agents actually give us a possibility to take a maturity step forward by improving the inputs. So, maybe, and we showed some examples, it was not, you know, a full list, but some examples of agents that can help us improve input for forecast, for supply planning, for inventory planning. But there are many more than what we showed, to help us improve the input, to get better results out. And once we are running a planning run, we can also get help to explain the results, to identify issues, to perform some of that work, so the planners can actually start using it and not overriding because they don't trust it. So, that's it for today. Thank you very much for joining everyone. Thanks for all the questions and for the input. So, the session here has been recorded. We will share the recording along with the slides, so you will have that. We might try to answer a few more of the questions because there were quite a few in the chat, so we will try to do that. But thank you very much. Thanks again to Timo and thanks for joining us today. Bye. Bye.