All right. Good morning, everyone. It's a beautiful morning here in Copenhagen, and we're very happy to welcome you to this webinar called the Execution Smartwatch. For the next 45 minutes, we'll provide a perspective on the role of AI in converting strategy into action and value. And the program is fairly heavy, to be honest. I can tell you that. And we also understand that not everyone can join for the full duration. So, you know, many of us actually have a job to do. So we'll be recording the session and also sharing the presentation afterwards so that you can enjoy it at your convenience. Let's get into it. My name is Christian and I'm a partner in the strategy execution team. And what I do for a living is essentially to help our clients close the gap between their strategic ambitions and what they do to make those ambitions come alive. And with me, I have my co-host, Adam. And I'm Adam. I'm also a partner at Implement. And I help companies all around the world implement generative AI solutions in their businesses for fame and glory, basically. Fame and glory. That sounds fantastic. Cool. OK. Later today, you'll also meet Julian, who is an AI engineer that has helped basically on our journey towards becoming smarter with with AI. And you'll meet him later and he'll get a chance to introduce himself. So, as I mentioned, I've spent pretty much all my career focusing on the well -established gap between strategic ambitions and reality. And there are numerous studies that have tried to quantify how big is this gap. And we really don't have to get into a fight over the number that best describes this gap. But one of them points to a 37% performance gap. And this essentially means that just a lot of money is being left on the table as organizations try to execute their strategy. And in recent years, we've seen the pace of change accelerating drastically. Disruption is happening across industries much more rapidly than just a generation ago. And one very important disruptor that is probably on everybody's mind these days is AI. And I believe that AI is most likely going to increase the pace of change even further. And we're already seeing this happening around us. And if the pace of change increases, then the cost of slow execution also increases. So for companies that can't keep up, the strategy execution gap is most likely going to get even bigger in the future. OK, as many other people have. Oh, sorry. Let's see. Just one back here. There we go. Just trying to figure out the techniques here. Yes. So as many other people have also done, we have tried to put the, you know, on a formula how to close the gap. And I won't be going into the details of the model. We've done that in previous times, but there are essentially five disciplines that leading organizations master to close the gap. And as I mentioned, we've held several webinars and written articles in the past. Covering the model. So I encourage everybody to have a look at that when you have time. And you can, for example, see a screenshot here from the event where Alexander Strange went through the model. And we'll, of course, share all of that with you. But just a quick reminder here of how the model looks because it will be getting important for a later part of the session. So some of you might recognize this, but essentially the model describes the path from strategy on the left side to value realization on the right side. And overall organizations need to be good at two things. On the left side, they need to make sure that they're doing the right initiatives. And on the right side, they need to make sure that they're doing the initiatives right. And that's just a whole lot of consulting bingo, I know. So let's get just a little bit more specific. So organizations need a way to make sure that they're constantly clear on what the top priorities are. And that's what we call portfolio management. Organizations also need to be actively balancing the supply and demand of resources towards those top priorities. And that's what we call resource management. And when initiatives have been selected, organizations need to execute them in an efficient and aligned way. And that's what we call project execution. And without the right people and capabilities, the likelihood of success is just incredibly low, even with AI. And that's what we call capability building. In the middle of the model, we have governance and performance. And it's all about making sure that decisions are made on an informed basis, essentially. And the reason why the model is organized as a loop is because we see this as an interconnected ecosystem. So improving in one discipline means that you have to also improve in another. And in this ecosystem, strategic choices basically inform execution and learnings from execution inform back to the strategic choices. So that was just a quick reminder of the model. And with all of that throat clearing out of the way, we can now dive into the agenda. So the content of today will be heavily focused on AI, as mentioned. It's a topic that's very hot right now and something that we've been working on a lot since the ChatGPT moment in 2022. In a moment, Adam is going to take us through the latest news on the AI frontier. And after that, we'll be diving into the smartwatch vision, as we call it, and talk about the shifts reshaping strategy execution today. We'll also talk about the conditions that need to be true for AI to work in an enterprise context. And then Julian will come in to show us a demo of the use case, one of the use cases that we've been working with and the challenges we faced and the learnings that we have gained from building AI augmented workflows. At the end, we'll cover the potential implications for key roles involved in strategy execution. And we might just run a little bit short on time. We did try to squeeze as much into this program as possible. But as mentioned earlier, we will record the session and you can return to it to at any time. Anyways, let's kick off with a little tour of the world of AI in 2026. Over to you, Adam. Thank you. Keep getting better. And for all of you who tried using ChatGPT, you likely noticed that in the beginning, almost three years ago at this point, the models were fun and interesting and they could make little songs. They weren't particularly good and they could summarize little things. But sort of proper intelligence was still far off. And these labs, they kept promising us that it would get better. And I think most of us might have noticed that it has gotten a little bit better, quite a bit maybe even. These newer reasoning models that has the little thinking markers. But what you might not still have noticed is that they have actually gotten really, really, really good. True frontier class intelligence is still out of the hands of most people. If you have a paid subscription to some of these services, you might have been able to try it a little bit in your private life. If you have access to Microsoft Co -Pilot and you really know how to use it, you might have seen some of it. But actually using these technologies to their full potential is still quite rare. We in Implement have the fortunate pleasure of having access to these models and have the time and opportunity to actually sit down and work with them. And we are really excited about where these capabilities are going. There's a number of technologies and developments that interact. And I think that's what I will give you a rundown of today to be really crisp on why we think now is the right time to start looking at strategy execution, in particular with AI. For one is the sheer intelligence of these models, especially with reasoning models. These models that are able to think and break down tasks for themselves, allows them to handle much greater complexity and problems. We don't really know how complex problems truly are. We don't have a measure of problem complexity outside of math, maybe. But we're getting much smarter with it. And it turned out that something like strategy execution, I don't know if anybody on this call is surprised, but was actually a pretty hard problem. It required a very high level of intelligence to manage all the different aspects, to take in the large volumes of data, things that we thought AI models would be really good at, but only just recently have actually gotten the capabilities to truly do. The sheer intelligence of these models had to improve to a level where they could actually engage in strategy execution. The other side of it is what we call the agentic design around it or the agentic orchestration. Back in the days with ChatGPT, it was just a chat bot. You took conversations, turned with it, you asked a question, you got a reply. That's how most people still know AI solutions when they use them. But recently we've seen the advent of agentic AI, something you probably mostly heard as a buzzword at this point. Some of you might have seen instances of it where if you may be asked ChatGPT a question, it goes and searches the web a little bit and it starts citing its sources. So this idea of hallucinations are suddenly, I won't say it's disappearing, it's still a significant issue as Julian will touch further upon. But it's greatly reduced because suddenly the AI can figure out information on its own. It's also able to do a host of other activities where suddenly the model is able to take action in the world via computers, via programming, all of these different things. And it turned out that was immensely important also for strategy execution because it allows the model to aggregate information in different ways, to paralyze itself out. So if information gets so overwhelming that a single AI can't look at it, AI can start collaborating with other AI actors to achieve these much more complex goals where they aggregate information up. That's a really, really important breakthrough. All of this has led to what we call the agentic frontier of work. There is a nonprofit foundation called the Meter Institute that you can seriously recommend you to look up. It's a fascinating piece. They have a large database of different assignments, most of them mainly technical, where they measure how long time does it take for a human to complete the task and how long time does it take and how successful does an AI complete the task? AI is obviously much faster at completing tasks than humans. Most tasks are solved within minutes or maybe an hour at most of runtime in an agentic setup. What we're seeing now that with each model generation, the time it takes a human to complete a similar task is doubling. That means that the complexity of work an AI can take in is doubling about every seven to four months. We don't quite know exactly how steep that curve is. It depends a bit on where you measure. The very frontier of models in a true powerful agentic setup seems able to complete work that's almost a full days of work autonomously. And we're expecting that to continue. And that's again why we think now seems to be the time for a lot of strategy execution to actually start to get truly augmented. I think the last point to touch on is still the human element that needs to stay in this loop. I have personally had a lot of interesting experiences with this. A great example was, of course, I could take a lot of information. We'll see an example of that later today where we could take massive amounts of interview, massive amount of conversations and then ask the AI to summarize and display it. But what's really powerful is in these agentic harnesses to allow the model to output it in specific formats. You can teach the model to adhere to certain standards, to use different tools. In my case, we needed to report into a specific one page template. If any of you are sitting in PMO offices, you will know exactly what I'm talking about. And I'll admit this was a nasty template. Nasty template, 36 different parameters, over 30 different program project streams in the program reviewed every month to a steer code. Very, very harsh stuff. But we could get the information in and by teaching an AI how to fill in information into a particular PowerPoint slide. And again, nerdy details here, but that's important for strategy execution. It's a format that we're familiar with. We could teach an AI to interact with the specific template, extract information and output it directly into it. You won't see that exactly today, but you will see similar setups where we're able to extract information and leverage these templates. But crucially, I was still in control of it. I originally designed the template. I was the one designing the integration with the template. I had AI helping me along in all of these steps. AI were a large part of defining the KPIs, large part of designing it, but we continuously work together. I was always looking at what was happening. I had all the one pages. Yes, I read all of the one pages, of course, because mistakes can happen. So staying in the loop, but staying more on it maybe and staying at a higher level of abstraction is what's really, really powerful with these AI technologies. So with that, I'll invite Christian back in the studio to a rundown of what this actually looks like. All right. So one of the things that I love about hearing Adam speak is that I always pick up a new word and one that I've been picking up lately is harness. And I just love that word. That's a fantastic new one adding to the glossary. So a lot of new information. And I think we can all agree that it's a really interesting time that we live in. And what I'll do now is to try to bring AI into the world of strategy execution. And to do so, I'll introduce the vision of the execution smartwatch. And today, I think many of us already enjoy the benefits of a smartwatch for a personal health. And I invite you to imagine now that you have a smartwatch that could track the health of your organization's execution capabilities. And the smartwatch would continuously monitor your execution. It would proactively flag emerging risks and it would help the organization make better decisions faster. Now, wouldn't that be nice? I do acknowledge that that's probably sounding to many of you like a consulting fever dream. And I honestly say that it's not like we're an implement going to do a Steve Jobs on you and produce a smartwatch and get it out as a product. That's not the direction we're heading in. But the smartwatch is more like a vision and more like a vision of incrementally better AI augmented solutions that are sort of interconnected. And when you connect all of these various AI solutions to each other, you can kind of start to see that there's a smartwatch emerging from that. So each solution basically improves one specific part of the execution discipline and taken together, they form this smartwatch. So one is the PMO thing that Adam was mentioning. Another one could be could be something around risk management and so on and so forth. So let me try to bring this a bit further down to earth and explain what exactly it is that I mean. And to do so, I want to introduce you to Maria. Maria is our fictive head of strategy and execution in a large multinational company. She's living a very typical day that I think many of us can really relate to. She wakes up in the morning already feeling just a little bit stressed. She was hired to do great strategic choices and convert these into executable roadmaps with executive leaders. She was also hired to monitor the portfolio of initiatives and decide with leaders what the best path forward is for each. But instead, she often finds herself in back to back meetings where she tries to defend some numbers that she doesn't really trust herself. Because there's no single source of truth. She feels a little bit like everything is becoming more and more political. Everybody's sort of defending their viewpoint. And someday she even skips lunch to tweak whatever people have submitted to her ahead of an important portfolio review meeting. She gets pushback from the CEO, the CFO. They're all looking for more information to make those decisions because they don't feel like they have the full picture. And so she spends her afternoon trying to stitch together something meaningful for them. She gets home very late, you know, doesn't manage to beat the traffic, gets home very late and only just manages to kiss the kids goodnight before she logs on again. And around midnight, she clocks out feeling super exhausted and with a sense that she's not really doing what she was hired to do. I know this is a bit of a dark story and also a bit of a caricature, but I'm sure that you guys out there can relate to at least some of what Maria is going through. Now let's turn the page and see what Maria's day looks like after the smartwatch. So she wakes up and she enjoys an AI generated briefing of her day ahead as she packs the lunch for her kids in the morning. On her way to work in the car, she reviews a risk mitigation action proposed by a project lead. And the agent that she has working for her approves this thing and sends the approval via email. The first meeting, it's only about 20 minutes long because everyone has sort of been briefed by their own individual agents who have collaborated on a solution proposal that everybody can sort of initially align to. So they only take 20 minutes to finally sign off on that one. In the leadership sink a little later at 10, everyone is looking at the same set of numbers. The conversation has sort of shifted from what is the data and can we trust it to what are we going to do about what it says. And in the meeting, the team spots a resource conflict that for a key project and they actually solve it on the spot. That's pretty nice. Lunch is actually pretty enjoyable. Maria has time to ask about her team's, you know, life outside works. They actually have a collegial conversation. And in the portfolio board meeting in the afternoon, the board actually makes a decision and pauses a project and initiates the other. And the reason they can do that is because the AI agents have spotted a change in assumptions. And this means that it now makes sense to pursue the other opportunity instead of what they had already committed resources to. Now, everybody was already briefed ahead of this meeting. So the discussion runs smoothly. And everyone affected from this decision, they're immediately notified and follow up meetings are automatically booked in their calendars so the leaders can communicate it to them directly. Maria spends the car ride home brainstorming a little bit around the Q2 strategy scenarios with her agent. And she gets home and has plenty of time to prepare food for the kids, help out with homework. And at the end of the day, she's still exhausted, but it's the good kind of exhausted, right? It's because you've been using your brain for something you really enjoy. She nods off and she wakes up eight hours later, fully rested and ready for another productive day. To me, that sounds a little bit nicer than the first day. And this might actually be where we're all headed very soon. At least I personally hope that that it is. So the core claim that I'm actually trying to make here is fairly straightforward. And it's also kind of uncontroversial and it goes a little something like this. So all else equal, more data plus more processing power will give us better insights, which in turn will give us better decisions. I think this statement is actually both true with and without AI in the picture. But with AI in the picture, it has even more potential than there was before. And perhaps the single most important issue that organizations need to be focusing on right now is that of data. How to generate it, how to store it, how to govern it and how to make it accessible and useful for different AI solutions and agents to work autonomously with it. And the reason why data is particularly important right now is due to the compounding effect. Now, I'm a finance person. So just like in finance, when your early investment compounds with interest over time, data kind of follows a similar pattern. As you accrue more relevant data, your AI solutions become better. And when your AI solutions become better, the value you get from them increases. So you can kind of see this compounding effect. But the difference between the finance world and the AI and data world is essentially that in finance, you can kind of catch up later by just investing a lump sum. So you can kind of make up for what's lost by investing more. But with AI and data, this is a little bit more challenging. You can't just reinvent real data from two years ago. You know that data that has to be earned, so to speak. And so in data and AI, you can actually get ahead by starting early. And this can actually lead to giving you an advantage. So we recently looked at the whole spectrum of disciplines within strategy execution. And then we tried to identify a range of shifts that we see AI unlocking. I won't be going into all of them, but I'll just show one example to illustrate what we're thinking in this domain. So let's zoom in on the case of clear value, less inflation. That's a headline that we decided to give this one. But one of the core issues that I see out there in organizations struggling with is identifying and monitoring assumptions on an ongoing basis. So assumptions are essentially at the core of every single good business case. And when they change, so does the business case. So we also see organizations struggle with monitoring the actual value creation from the initiatives described in the business case. So our hypothesis about the future is that, first of all, AI can help monitor assumptions, conditions and value creation across the portfolio. And second, that AI can help flag these changes to the conditions proactively and help propose solutions. And if this is actually true, then it would give companies quite a lot more proactive control over their value creation. They can actually make those necessary adjustments as conditions change, which many companies fail to actually do today. So there's more shifts like this and you can read a little bit more about it in the article. And here we go. Turning strategy into value with AI, the white paper. There we go. We will now dive a little bit further into we need to move on to the next topic. I know I've been talking a lot, but we need to move on to the next topic where we dive into the conditions that must be true for the execution smartwatch or any AI solution really to become a reality. And for that, I invite Adam, our expert back to the stage. So here you go, Adam. Thank you so much. So in order to use these tools reliably, we need five things to be true. Number one, AI needs to be able to reliably synthesize data across multiple sources. I touched a bit upon it before, but part of what has been holding strategy execution back is that strategy execution. Legitimately is a massive volume game and it's not obvious how you aggregate that kind of information. If you are in a factory and you're working with sensor data, there's a lot of obvious ways to aggregate information in timestamps and you can do simple arithmetic. So unlike what's the average temperature or things like that. That's much harder to do in strategy execution. So we needed models and we needed systems that were able to take in not just dozens of pages. If you ever remembered back in the days on chat GPT, you would get this warning or error saying too much information uploaded. Nowadays, we have models that are able to take in hundreds of pages worth of information. And thanks to agentic harnesses, we can take in near limitless amounts of information if we allow to intelligently aggregate things. So number one breakthrough and number one condition that we now have that was required was this ability to take in near unlimited amounts of unstructured information. The second is that AI must be able to cite its sources. This is the hallucination issue. And while still not fully resolved, we've now clearly gotten to a point where it's reliable enough for a lot of strategy execution work. Some of the studies that we made internally actually finds that AI is more accurate than humans in most cases. If you just have a human writing notes after a meeting, they will make more mistakes than an AI that had access to a microphone, which I think most of us will find pretty obvious in just a few short months or maybe a year that of course an AI with perfect access to an entire meeting will remember more details than a human who has to use their very human based memories. And with that, we're able to cite sources directly down into the information. We'll look at that in the demo session coming up in just a few minutes. The next one is that AI must be predictable and repeatable. So once we start paralyzing things out, it's very important that different work streams gets the exact same treatment by the AI system. This is already today a problem in project management just with humans that, you know, different project managers or different stream leads might have slight different preferences, different styles that they want to want to lean into. We want to get as little of that as possible. Back in the days, these AI models were more stochastic. They're still stochastic models, so they will still do slight differences, but they're much more reliable. They're much more stable in the way they treat things. If you ask the model the same question, it's going to give content wise a much more reliable answer. And that's a very, very nice feature for us believing in this as an enterprise scale. The fourth one is that AI must be able to work at scale. So you must be able to access these models at a reasonable speed and a reasonable scale. I don't know how many of you have actually encountered this issue, but it is a legitimate challenge nowadays that there's just not enough compute power available in the world. We run some of the largest jobs that we have had in strategy execution where the legitimate issue was that companies like Microsoft would tell us that we can complete this job for you. There's just not enough computers available in our Swedish data center to complete it right now. And again, not because we're taking the entire data center, but of course, all customers are competing for access to these models. This is also getting a lot better. Models are getting more diverse and we can start delegating to models of different sizes and capabilities to actually work at scale, not just in theory, but actually in practice as well. And finally, AI must be able to act autonomously within human defined governed boundaries. This is back to that we need to be able to control these models. We need to be able to explain to them what are we looking for? How does the framework actually work? And these reasonably complex strategy frameworks needs to be something the model can actually comprehend. Again, in more typical machine settings, it's easier. You can define boundaries with literal strips of tape on the ground and say the robot can move beyond this point and you will have a sensor that detects the tape and the model won't go there. In strategy execution, it's not like that. You can't put tape on the ground in the same way. So you need a model that's intelligent enough to get these fluffy, unclear boundaries of what does governance look like. Again, it's easy for some things. Some things can be hard limitations, but a lot of governance and strategy execution work is really, really, I don't want to say vague, but it's very human. It's very upfront interpretation. It's full of exceptions and exceptions need good reasons. And knowing when to manage and when to move in that space is something the models only just recently have actually gotten capable of. Those are the five conditions. And with that, I'll hand it back to Christian. Thank you. Okay, so here they are five conditions that we try to hold ourselves to. I think this is a good slide for everybody out there to take a screenshot of and we'll also send it afterwards, of course. But this is a good thing to just keep in mind when building AI solutions. We'll now introduce the use case called the X-ray. So we're zooming in on a use case that we've been experimenting with for the past years. And it's basically a service that we provide to a lot of our clients. Essentially, what we do is that we do a bunch of interviews, observation studies and send out surveys. So we collect a lot of data and then we consolidate that into a gap analysis based on our assessment framework. And these could be different assessment frameworks. But it's a good way for leaders to get aligned on the key priorities in terms of what they need to improve. It's a good way to create some change momentum because you've actually listened to the people who need to do change. And it's a great way to establish a baseline for measuring progress. So this is just briefly what the X-ray is all about. And the reason for choosing this as a use case, what actually makes this a good use case? Well, first of all, it's an internal workflow that we have refined for years and we know very well. And second of all, we can sort of specify all the steps in the workflow and do what we call deterministic automation. So that's the least sophisticated agentic setup that you can do. And lastly, we believe that we were able to test those five conditions that Adam just went through in this single use case to see how the models would sort of live up to those conditions. So in sum, it made for a pretty good testing ground for our AI experimentation. And Julian will now take you through a quick demo and highlight the key challenges we faced and also the key learnings that we have. That we've gained over the years of experimenting. So over to you, Julian. Take it away. Thank you, guys. Hi, my name is Julian and I'm an AI engineer at Implement. I've been so lucky to be part of the team to develop the AI augmented X-ray in the latest iteration. However, it is really the result of the work of many others who have improved quality of the X-ray over the last two years since we first started experimenting with ways on how to improve this. And we made some great progress and we're actually on iteration number five right now. And we've managed to automate around 60% of the workflow compared to two years ago. And this is really due to all the great work by the consultants seen on this list here that are provided with their subject matter expertise and technical capabilities to drive this, identify challenges and keep improving. And one of the key things we got out of it is to identify the key technical challenges that we will need to solve in order for this to work at scale. And they are listed here and I would like to go through them because there are simply some problems that we encounter quite often when we work with AI at enterprise level. So one of them are hallucination, distinguishing pattern from noise, positivity bias, loss of context at scale, and generic language. And I would just highlight, try to highlight a few of these. So for example, hallucination, we've heard about that since the first models were released by OpenAI a long time ago. So whenever you are chatting with ChatGPT or Claude, you often see that they sometimes write something which is not grounded in evidence. That is not so serious when you're working with cooking recipes at home. But if you're working at enterprise level, the consequences can be quite grave. So that's really important to nail down. So positivity bias is less so a flaw within the systems themselves, but more a design choice by the providers. So they have this welcoming approach where the models seem quite positive towards you, but it can actually also give issues. And we saw that sometimes the gaps in the assessments were reduced and positivity around a single remark were inflated. So we couldn't really recognize what was said in the interviews in the final reports. Also generic language, quite common. When there is a lot of data, the models we saw could tend to converge towards some generic management terminology. And the clients in the end could not recognize the words that they use themselves. So also a super important challenge to tackle. What we can say is that the promise was really there from the beginning. But in order for this to work at scale, we had to narrow down and solve these technical challenges. Now I'd like to take you through a demo flow and then I'll try to double click on how we solve these challenges when we get to that in the demo. So now we see the screen that the consultant land on when they start the AI augmented x-ray tool. So on the left side, we have all these five steps that we have to go through. Of course, prior to this, the consultants have found stakeholders within the company. They've interviewed them and transcribed the interviews. They've done this according to two assessments type. One is the transformation program management here and the strategy execution assessment. So these are two frameworks that we will not get into in detail here, but that we use to assess the execution efforts of the company. Another thing you can configure here is the model configuration. And that is an interesting topic. So whenever you approach an AI problem, you have to choose which model to choose. And this is a balance between rising token prices from the providers. It is also about time complexity of the task. How much time can the analysis actually take? But it's also around maybe the most important one. What are the required capabilities of the task? What we found in this specific problem was that we had to use frontier reasoning models of the newer kind in order for really catch all the complexities in the analysis for this to work well. Right. So now let's try to upload the interview transcripts that we have here. They are marked, so it's not real data. And then we click continue and we start and we see we go to the next step of the process. So now we're in the pre consolidation step, which is basically a data pipeline that goes through all the interviews one by one in parallel. So here we take the docx format. We convert it to a more structured output that is usable by the LLM to see something like citations, the timestamp, the person speaking and roughly what they're speaking about. What we then do is go into an agentic loop, which is separated into steps. The first step gets an overview of the content of the interview. Then we classify, we identify and then we classify the core citations that belong into the framework. This will make more sense in a bit when I start showing the output. But then the final point is writing the observations according to the citations that we found. And this was a really core point in order for us to reduce hallucination. Because what we do here is basically manage the context that is sent to the model. So it's super important when you start building AI solutions that you are very narrow in your scoping of the tasks that they should solve. So that is what we did here by the stepwise approach, but also by just including the necessary data to solve that specific task to not overload the model with unnecessary information. But now let's try to look at what these reports look like after they've been processed by the model. So I'll just close this. And here we see all the different interviews that have been processed. So let's try to open a sync one so you can see what it looks like. So here you see all the dimensions of the framework. And this was the strategy execution one so they are all listed here. Within those specific parameters you have subparameters or subdimensions that you can look into. And let's just for the sake of this one look at portfolio management. So what has happened here is that the model has generated an observation like this one, which talks about portfolio visibility has improved through recurring quarter by quarter updates and so on. So where is this coming from? And that is another point I would like to highlight in this tool. Is this traceability and explainability that's really core in order for us to have a tool that works at scale. So what the consultant can do is to go into the sources and then see exactly what was said from the transcript, who said it and at what timestamp. So they can go back and trace where exactly this is coming from. And again, this was just a super important part of this is to have this audit trail so the consultants can trace it back. What it also does is provide us with a sort of quality gate. So whenever the consultant has seen this and they say, yes, that was what was said in the interview, it is captured well by the agent, then they can click accept and we can move on. So you will go through these one by one. You would make edits if necessary. And then in the end you would accept all. So let's just do that here for the purpose of this demo. So we click continue and now we see we moved into the consolidation step. So in the consolidation step, the design is much similar to the pre consolidation, meaning that we have an agentic loop with several steps where we handle data and model outputs sequentially. But here we are taking all these pre processed interviews and consolidating them into a single report, which means that we are working at much larger data scales. And for that reason, we have to be quite smart in how we manage this again, not to overload the models and cause hallucination, generic language and all these other challenges that we highlighted before. So what we did here was to have a read and write loop where the agent can access the different pre processed interviews, extract the observations and then it can go through all the interviews like this until it has all the observations written down. It will then keep doing this until it converges. And the way we define convergence here was based on two parameters. So one is cardinality, which basically just means how many times was an observation mentioned. For example, the portfolio management quarter on quarter cadence that we saw before. That could be one example, but it's also around novelty. So let's say only a single client mentions a specific purpose or specific thing within the framework that none other mentioned. It is still important to capture this because it's very relevant for the framework, but only one person mentioned it. So we try to balance this until the interviews are consolidated into this final report. So let us have a look at what that looks like here. So now we're in the final report. And what you will see is again that the parameters match what you saw before. So we're still adhering to the framework structure defined by the strategy execution framework. Again, now if we go into portfolio management, for example, we can see that in general, the sources have increased quite a lot. So of course, now we're looking at all the interviews at once. So therefore, there are quite a few sources coming from different interviews. So let's just try and look at one with seven sources here. So again, an insight around data quality and tool limitations. And then you can go in and you can see exactly where this comes from. So this was according to a single interview, two interviews, three interviews. And then you can see the person who said it again and the timestamp. So again, a way for the consultant to also, when you have the consolidated report, to go back and trace everything. You can also do this through the evidence map. I will not spend too much time on this, but just show that this is another way for a consultant to visualize exactly where this is coming from. So you have the parameter, the subparameters, and then you can go in. You can make it up in here. And you can expand that. So you have from two different interviews, four citations. And then you can go into the specific interview and see the citations. So again, another way for the consultant to get more transparency. So the final feature I'd like to showcase here is the chat and the agent functionality. So if we open it here and make the screen a little bit bigger, we see here we have the chat, which is basic QA functionality. So the consultant can go in and ask questions through findings in the report and really drill down and again, get some more explainability to where things are coming from. You can also use the agent to make direct edits into the report. But for now, let's just try the chat functionality where we can ask how many projects are ongoing. And we send that to the chat. And then we get an answer out. So again, a way for the consultant to drill down and see exactly what's going on. And that concludes the demo of this AI tool, which is this one example of how we use AI in our daily work as consultants at Implement. And the underlying foundation are really the conditions that Adam stated earlier that we have to adhere to whenever we do this. And what we find is that when we combine sub -big matter expertise together with engineering, we can actually build these AI tools to deliver positive impact in our daily work. That is all for here. Thank you for seeing the demo and back to you, Adam and Christian. All right. So I can see on the clock here that we're inching very, very close to the end of the webinar. So the last topic that I'd love to touch upon here is implications for the key roles that is typically involved in bringing strategy to life. But I think in the interest of time, I won't be able to go through it in enough detail. So we're going to carve that out for another good time and invite you guys for another webinar. So let me just skip through a couple of these so you can see what we were hoping to land on here. And we'll have to do that at another time. And fear not, you can you can read more about the viewpoint here in the white paper called Turning Strategy into Value with AI, the execution smartwatch. And to be honest, this is also a very emerging field. So I think the opinions that that we have on the impact for the different key roles is sort of moving on a day to day basis. So so being a little flexible here in terms of the agenda, we'll carve that last section out and and discuss the implications for the roles at greater length with more people at a later stage in time. But if you're if you're curious about AI and curious about strategy execution, then there are several options to get involved in the coming period of time. You can, for example, reach out to us to become a pilot for the the X-ray itself. We're looking for organizations who are willing to to try this this new way of doing it out in practice. We've already done it a couple of times, but always looking for more. Then there's an upcoming webinar. I think it is in September on on on project management and AI that that could be interesting to join. Of course, read the white paper. It has quite a lot of explanations that that dive into many of the details we didn't manage to get to today. You can also join the finance hackathon. It's an invite only. So maybe you're lucky to get an invite. And then we also have an upcoming webinar just before summer on on AI adoption that that you can dive further into. But that's it for us today. Thank you so much for spending your morning with us. And we hope that you enjoy the rest of the sunny day if you're in Copenhagen. Could also be sunny elsewhere, but have a good one out there. Thank you so much for watching.