and the next step. And then, we're going to talk about how you can connect your AI solutions with your company specific data and build these really cool contexts that were chatbots or AI solutions connected to your modern data platform. And we actually put already the first small example of how this looks on the screen you see here. So on the right one hand side you see kind of a generic chatbot that doesn't have access to any context from your company. And as you see it's a little bit less interesting and value adding than the one on the right hand side that actually understands the customer's specific request here and is able to answer. So what we'll be diving into today is how you actually build this and achieve this. But first of all, just a quick introduction. So I'm here together with Morten today. My name is Nikolai. I'm a partner in our data and AI team here at Implement. I've been building AI systems for the last approximately 10 years and the last four years with Implement. And I'm joined together with Morten today from our good friends at the Tech Collective. Hello everyone. My name is Morten and I have a background in business administration and information systems. I worked as an IT consultant for the last five years and I've been focusing on building modern data platforms from the ground together with different customers. So Nikolai and I share the same passion for technology and data. And one day we were thinking, so why aren't we hearing any success stories about generative AI and modern data platforms? Therefore we decided to create this webinar to maybe try to combine these two worlds. But without further, I will leave the stage to you Nikolai and step aside. All right. All right. And just a little bit of expectation setting before we start. So first of all, we will dive into a little bit around why generative AI actually needs to have contextual information from your business systems. And then Morten will come back and explain a little bit about how data platforms can actually help you solve this. And finally, we'll give some examples of how the path actually looks to be getting there. And just to align expectations. And just to align expectations, then we will not be going into a lot of technical details today. But if you want to hear more about that, we are really happy to take a follow-up meeting and dive as deep as you want into the technical details. But let's get started. So what everyone is looking at right now is to build AI systems, that being a chatbot or some kind of automation solution that is automated. And that's great. And that's great. And that's great. But maybe to sell you a little secret, that's actually the easy part. So the difficult part is getting the real actual foundation in place for the AI solution. And that's some of the stuff that we've been looking into for many years, actually getting the right data in place. And that we look at the base of the pyramid here, really the foundation is high quality, trusted data in your source systems. And that's where maybe boring stuff as data governance comes into place. So how do you actually control which systems have information about your customer, which system has information about your shipments, which system has information about your orders. When you have that data foundation in place, then you can kind of move to the next level here. And this is where you start to make data interoperable. So you connect all the different different systems and you're able to make one view, one single source of truth across different systems around. So somebody placed an order, who was that customer, where is it being shipped to, and so forth. And that's where the modern data platform also comes into place that Morten will cover a little bit later. When you have that in order, that's where we believe you can build AI systems that has much more business impact than the more generic solutions that we are seeing a lot of data. And the question I guess you can ask yourself is where are you today. And if we are to give one suggestion for what answer you will arrive at, we can look at some of the clients that we are working with right now and look at ourselves what we have been doing for the last couple of years. And since generative AI came out around 2022, we have been building maybe more generic solutions that are connected to a language model. And that's, of course, really fascinating. We can now talk to a computer and we can get it to automate much more complex business flows. But we also believe that really to take it to the next step, looking at 2025 and forward, then we need to be able to provide AI with this contextual information. This is really where you can take it to the next level, where you can build AI systems that are much more value adding and are able to automate business processes that humans would otherwise be doing or provide service information to your customers or or to even to your internal employees in these chatbot interfaces, where the chatbot really understands the context of the employee or the end user. So, one way to look at this is the Gartner hype cycle that a lot of you probably know. And I think we can agree on one thing over the last couple of years, if there's one thing that AI has been that is hyped. So, so there's been a lot of hype around this. It's maybe been the archetype of a hyped technology. And we've really been seeing that for the last two years. And maybe especially from executive teams, they have been saying this to us. So, we need an AI solution. But this is really a technology that's looking for a problem to solve. And that's fair enough. It is a really cool technology. And there are strong processes for what we can use this for. So, that's what we did. We tried it out. And suddenly you get this feeling, maybe you remember the first time you used chat GPT yourself. It's really fascinating. And you get the impression that we can probably solve all kinds of problems with this, because you can actually talk to your computer now. It's a completely different way of interacting and a new capability that the machine has. And then maybe over the last year's time is a lot of companies have come to the realization that it delivers answers that are maybe a little bit out of context. Maybe it does this hallucination you might have heard about. So, it's providing answers, but they're not really bound in the truth. It actually sometimes even lies is the feeling. And we're getting down to this kind of throf of disillusionment. So, maybe we are starting to doubt if this technology would really be useful for us. And what we believe now is that we are at the point where we are slowly starting to kind of connect these AI systems down to our data foundation. And this is where you can really provide the real value and where we can build the really cool AI solutions that deliver business impact. And where you start to automate some of the business processes that we were dreaming about. But it would be really interesting to hear a little bit from you guys out there. So, what are some of the biggest challenges in adopting the AI solutions at your company? And we've prepared a little poll for this that will come up in a couple of seconds. And some of the stuff that we are seeing out there is as with any technology. Now, I need to be careful not to bias you too much in your answer here. But it's actually not the technology implementation, but it's more related to the soft part of it. Actually, getting people involved in this transformation, getting them to use and trust the technology. And building the right capabilities to be able to actually maintain and build and deliver these kinds of solutions. And then we're getting some answers already now. And great. That kind of underlines the theme of today. Thank you for supporting that. But the lack of the data and the right quality and with the right availability is, I can see a major concern for a lot of you guys. So, that's great. That's why we are having this discussion today around how modern data platforms can support you in that. Perhaps the audience is also slightly biased in this regard. Those guys that are here today. So, all right, let's move on. So, what we have seen a lot for the last couple of years is a lot of proof of concepts where we try out to build some of these AI solutions. And if we look at some numbers from Gartner here, 30% of generative AI projects are being abandoned actually after the proof of concept due to poor data quality, risk controls or escalating costs. And I think that actually matches pretty good with some of the numbers we have seen out there. And maybe we have even seen numbers that are slightly higher at some companies. But we believe also that there is a way around this. So, if you take a slightly more strategic approach to this and actually before you invest a lot in a proof of concept, try to understand what are the risk factors around this use case. Do we have the right data in place? Do we have the right data in place? And so forth. That will save you a lot of troubles later on. And that can actually be done by taking a more strategic view from the very beginning. All right. And if we are kind of diving just a little bit into the technical details here around what are we seeing some of the challenges are when we try to deploy these systems. There are kind of three things we run into a lot. One is this highly normalized data and that's an IT word for when you're trying to look directly into the source system. So, you're looking into a database. That's basically optimized to drive a specific application. So, it's not built for AI. It's not built for analytics or business intelligence. It's built to drive that specific business application. So, that's the challenge. So, we need to have some kind of abstraction layer above that. And we will get back to what that is. Then, there's the challenge of data silos and limited perspectives. So, if we are not able to make data interoperable between the different systems, then you're getting these data silos where you might understand who your customer is, but you have limited visibility into which orders they have placed, which campaigns that they have received, what newsletters they have subscribed to, and so forth. And then, finally, the last big challenge we're seeing out there is the lack of accessibility down to these source systems. So, do we have a good API interface for actually getting the data out or do we need to go directly down to a database? And the key insight here is leading us to the next section that Morton will present in a moment. This abstraction layer, the data platform, the modern data platform, we need to have on top to be able to solve these problems and give your AI solution much easier access to the data that you do have in your company. So, finally, just to summarize my part of the presentation this morning. So, there are kind of three things you need to have in place to really build a strong foundation for your AI solution. One thing is you need to be able to provide real-time retrieval of business information. So, the AI models, they often interact with humans or with business processes. And they need to be able to get a real-time view of the single source of truth or the data situation across your different business systems. Then, second, data needs to be interoperable. So, what does this mean? It means that we want to have systems that can or have a view across the different systems and understand, as an example with the customer, what orders have they placed. What campaigns have they subscribed to? And then finally, and super important, then we need to also bring the security perspective into place here. So, you need to be super careful when you build these AI solutions. And we have seen many examples of that where you actually give it access to a little bit too much and maybe this information is floating to the end users as well, which can be super problematic. So, actually lifting those security models from your business applications and feeding those up to your AI systems as well. So, you don't have information floating to people that shouldn't see it. And we believe the modern data platform is a way to solve a lot of these challenges. And this is what Morten will dive into now. And before we do so, we have a little poll as well around, do you have the right data foundation for fully utilizing AI at your company? So, it's just a simple yes, no answer and interesting to see how many believe you have that. And probably based on your answers from the last poll, then this will probably be a challenge. at least what we see when we work with our clients as well. This is a challenge in a lot of places. All right. And I'll use this opportunity as well to just switch place with Morten. All right. Hello again. So, I'll just give you some seconds to finish this poll and then we'll continue afterwards. All right. Still seeing some answers coming in here. It seems like there's a majority vote towards that people don't actually have the right foundation for fully utilizing AI. utilizing the company. That's interesting. All right. Okay. But I will continue here. So, we have now seen why companies they might struggle with generative AI adoption, which of some of the things that Nikolai touched upon. But how do we actually solve these problems? Our answer here is the modern data platform. So, what exactly is a modern data platform? So, what exactly is a modern data platform? So, if you think about data as electricity, then the data architecture is the network of wires and circuits that transmit electricity from the power plant to the devices that need it. In this example, you can think of the data platform as the power plant transmitting data out to the different business units that need it. So, the definition of a modern data platform is a suit of cloud first and cloud native software products that enable the collection, cleansing, transformation, and analyzes of an organization's data in order to improve decision-making. This is at least how IBM defines it. So, it's not just really a classic data warehouse. It's more about making data usable, accessible, and ready for AI. And this is actually the keyword. But what exactly does this look like? Let's try to break that down. So, a modern data platform isn't just a database. It's an ecosystem that powers digital projects. So, a key thing in a digital transformation, including roadmaps, operating models, and change management, is having scalable data and AI solution. These kind of solutions we categorize into three different brackets. So, you have Gen AI and traditional apps, Business Intelligence and Reporting, and traditional AI and machine learning. To enable these types of solutions, you need a modern data platform. You can kind of think of the modern data platform as the architecture of houses. They come in many different shapes. They have many different dimensions, but they also share some dimensions. For instance, in your house, you will always have a toilet. You will always have a kitchen. You will always have a living room. And it's the same with a modern data platform. But it also depends on the size of the company and what the context they are in. We have given six dimensions here, which we believe are important, but they can vary. One of them is data integration, real-time processing, governance and security, and so on. And then, I think that we think is very important is having the same operating model. This means that the people in finance should operate in the same way with this modern data platform as people in HR, for instance. So, on top of the modern data platform, below the modern data platform is the operational systems, IoT, and sensors, basically every data that can feed into this. So, examples of operative systems, as Nikolai also talked about, could be CRM system, ERP system, it could be a POS, a point of sale system, and so on. The fundamental thing for this to happen is, of course, that you have a cloud infrastructure investment, and you have some thoughts about your data culture, and you also thought about being AI ready. So, in order to scale digital transformation within data and AI solutions, you need a modern data platform. But to truly understand its value, let's compare it to the classic data warehouse to see where some of the differences lie. All right. So, what differentiates a modern data platform from a classic data warehouse? So, in a classic data warehouse, the scope primarily focuses on structured and historical data. But in a modern data platform, it's not a data platform. It handles structured and semi-structured data as well. In terms of source types, in the classic data warehouse, you have data from operational systems. But in modern data platforms, it combines IoT, logs, social media, and operational systems. When we talk about functionality, in the classic data warehouse, we were more focused on analytics and reporting, but we are now expanding this in the modern data platform to apps, generative AI, and more advanced analytics. And architecture in the modern data warehouse. And architecture in the classic data warehouse would typically be around the cloud data warehouses, where in the modern data platform, we see integration platforms or data lake houses and so on. So, to emphasize, a modern data platform is designed to scale, it's flexible, and it's real time. It supports AI applications. So, we can kind of think of this modern data platform as a marketplace. At least that could be in a very fair, at least that could be an early vision. And we want to handle the exchange and fulfillment of data. So, on the one hand side, we have the suppliers. The suppliers are the operative systems that we already talked about. It can be sensors and IoT. It can basically be any data. And then on the right hand side, we have the consumers. Consumers can be decisions makers. It can be functional employees. It can be upper management. It can even be your CEO. We also have data analysts, ML engineers, Genitive AI. developers, software engineers, and so on. In the middle, we have the data platform. And again, I want you to think about a data marketplace when exchange happens. So, this decision maker submits a request for a data order, and the data platform team handles this request and uploads the data request. In a fulfillment process, you can think of engineers extracting data to develop, reuse of the data assets, and then store and process these in order to finally provide them through APIs or specific data access layers. An additional service could be that you feed some of this data back into some of the operative systems, but it could also be that you build some AI solutions on top of it. So, if you think about a modern data platform as a marketplace for data, it is not just a many-to-one relationship, as we have seen in the classic data warehouse. It's more like a many-to-many relationship. So, AI isn't just about using the newest available models. So, let's take a look at what the role of a data platform is in generative AI. So, we can look at it from a core capabilities perspective, and we can look at it from expanding capabilities perspective. So, when we talk about the core capabilities, we are talking about trustworthy AI. So, we want to ensure that data is accurate. And we're also talking about near or real-time processing. So, we want to provide up-to-date data at the point in time where we need it. That's the core capabilities. When we're talking about expanding capabilities, we are talking about a scalable infrastructure to support AI growth. So, with this, we're meaning scalability in the matter that generative AI is built on knowledge bases, and we want to be able to scale an infrastructure that can handle knowledge bases that can handle knowledge bases that can be used in the same way. When we talk about it from a governance and security perspective, we want to expand and enforce the compliance that AI operates on within the regulatory boundaries throughout the organization. So, this meaning that different business units, they have the same regulatory boundaries within this platform. So, without a modern data platform, AI will always struggle to have accuracy, governance, trust, and scale. So, try to imagine that you are a customer asking an AI assistant, where is my order, as we saw in the intro slide. The answer truly depends on whether AI has access to data or not, and if this data is real-time. So, if you take an example, without the data platform, you get the potential risk of limited personalization, a generic response, outdated information, or risk of hallucinations. And some of the reasons are that these large language models, they are built on generic training data, and at some point in time, it is cut off, and therefore, you don't have access to your proprietary data or real-time insights that are happening recently. If you take the example with the data platform, you can get the benefits of increased customer experience, better accurate responses, increased trust, and better problem solving, and so on. And this is because the generative AI application is connected to one source of truth of your data across your systems. So, Gen AI alone isn't enough. It needs to be connected to the contextual data in order to deliver some value. So, let's take a look at how generative AI models learn, retrieve, and improve, because this is an important thing in order to understand the context. So, when we talk about training, we have large language models that learn from large datasets, but at some point in time, they have a knowledge cutoff, and this means there's no updates after this period. This also means that the model quality depends entirely on the selected data during this training period. Therefore, it cannot access recent information or company-specific data. Then we have retrieval augmented generation, also abbreviated as a RAC. This is where AI retrieves and improves accuracy by searching relevant sources before responding. So, we often see enterprising having domain-specific models. So, for instance, there could be one in HR and one in finance and one with compliance and so on. And finally, we have fine-tunings, and this is where AI adapts for a specific task or format with a curated dataset. So, a specific dataset that it's giving to fine-tune the prompts. This ensures that responses match the desired style or terminology or structure that we're in. It is very useful for custom workflows, branding, and specialized activities, and so on. So, to emphasize here, a RAC is what makes AI relevant, but let's take a look at what a RAC is. So, RAC is a missing link that makes AI context-aware and dynamic instead of being static. So, let's break this down in three different buckets. So, we have retrieve, augment, and generate. Let's try to think of this from a stock trader's perspective. So, you can imagine a stock trader sitting with historical data about previous transactions. And then you can think about him also having access to some live data. So, current transactions, that are happening in the market, or other analysts providing some information. Together with the historic and the live data, he has an improved context. And this can make, can do better informed decision and better actions. So, to emphasize here, without a RAC, AI is just guessing based on some old knowledge. All right. All right. But for a RAC to work, we need to have access to the right data sources. So, let's try to look at the different knowledge bases we often see in generative AI solutions. So, on one hand side, we can have document stores. These documents store semi-structured and unstructured data. So, this can be FAQs in, if you take a retailer as an example, it can be FAQs. It can be return policies and specific product information, for instance. We enable these documents through vector databases. Vector databases is basically storing embeddings. And embeddings are like numerical representation of text. So, what it does is it understands a vague customer question because it retrieves semantically similar responses based on searching keywords. So, an example here could be a customer asking, can I return my order. And here, the generative AI RAC solution knows that return is the keyword. So, it's probably the return policy that needs to be returned in some text. Another example of a knowledge base could be that you want to use all your operational systems. And as Nikolai talked about previously, we often see them as siloed but structured data. So, and they are always in a specific, they have purpose for instance, a business application. But we want to make these into data products, products which is structured and uniform data. The key contribution of this is to aggregate and unify order inventory and customer data to provide some real-time updates and enable this through APIs. So, for instance, when a customer asks, where is my stuff? The generative AI application would know based on a customer ID, the specific order information, the time, the time, and when it will arrive. So, to emphasize this, the combining knowledge bases in a single platform is what truly makes generative AI powerful. So, let's see how all this comes together with a modern data platform. All right. So, how does AI actually retrieve and process data? This is where the modern data platform plays a crucial role. So, if you take a look at this RAC framework, we have put the data platform into it as the knowledge base. So, if you take an example where a user asks a question or provides some input in the prompt, the retrieval model will find the relevant data from the different knowledge bases. So, an example could be the keyword will be sent to the vector search and the return policy will be returned to the retrieval model. Or it could be a query sent with the customer ID. And the table for that customer ID with the specific order detail will be returned. And this will then be combined in the prompt with the relevant data to add some context. And hereafter, AI creates a customized and accurate response with the generation model and feed this back to the response and back to the user interface. So, without a data platform, AI lacks the accuracy, but this RAC here ensures that it stays relevant and trustworthy. So, now that we have talked about a very specific use case, we want to ask you a question. So, which AI powered use case will bring you the most value to your business? Take a second to think about this and we'll come back to this later. All right. I hope you had some time to think, but let's continue with the presentation here. So, now that we understand how data platforms power AI, why is now then the right time to act? So, we're seeing that customers and employee expectations are rising. And this is because they already have some installed solution and they are getting familiar with the functionality and the more advanced features of it. And therefore, the expectations are rising to these. And therefore, we believe the proprietary data is the necessary thing to stay competitive because generative AI is a commodity technology, meaning that everyone has access to it, everyone has access to it. And therefore, investing in your knowledge bases and building these data platforms is what really makes a true difference to stay competitive. We also believe that the technology is mature and ready to be deployed. And you can think of this. So, the modern data stack has evolved over the past decade. And this has meant that the types of solution and work that it takes in order to make AI solutions has decreased. And it's much more faster to build these kind of solutions now. So, to emphasize the AI racism, just about adopting AI first, it's about scaling it effectively with the right foundation. So, how do organization moves from experimentation to execution? And how do you actually go into production when you have already built your POC? We believe that building AI success requires more than just the latest models. It needs to be structured and that approach. So, here we have an example of a high level roadmap for how you get started with the modern data platform. We have categorized this into four different waves. So, in the first wave, we have the blueprint. This is where you identify the use cases and where you design the platform architecture and define the governance and metrics and plan the engagement and training and so on. Then you move on to the MVP. So, you move on to the MVP launch where you want to launch the MVP that delivers some specific business value and deploy it out in the systems. Finally, you want to implement the basic governance control and so on. And this is all to get some validation and learning from the MVP in order to scale this solution more broadly. So, scaling and role is the last step in the wave. And then finally, it ends up with a modern data platform. So, to sum up, you kind of build this MVP and then the idea is to scale it out to the different business units. So, to emphasize AI success isn't just a single project. It's an ongoing transformation and it requires governance, infrastructure and training. And at implement, we usually say also a tech collective that at least the same amount in development should go to change management. It's not always the case for customers, but at least the same amount in development should go to change management. It's not always the case for customers, but at least the same amount in development should go to change management. All right. So, can an AI model succeed without a modern data platform? We believe the answer is no. Large language models that lack access to proprietary data and real-time data. And you can ask the same question, is a data platform alone enough? The answer would also be no. The data platform has no natural language accessibility. But together, they create some synergies. You can kind of think of the large language model as the voice. It creates the interaction and engagement layer. And then you have the data platform, which is the brain. It stands for the foundation of knowledge and logic. So, together, these two in combination is very essential for scaling generative AI across your organization. So, AI without a data platform is incomplete. The true value comes from their synergies. And I will take Nikolaic. And I will take Nikolai into the stage here because we're ready to round off. All right. Thank you so much, everybody, for joining in today. It was a pleasure talking to you about this. And we hope you got some insights out of it. If you'd like to continue the conversation and maybe go deeper into some of the technical topics we've been covered today, feel free to reach out to both of us. So, thank you so much. Have a great Tuesday, everybody. Yeah. Have a great day, everyone.