Transcription
Warm welcome to this first AI seminar at Norris Bank. And this is just a huge moment because we have never seen a technology like this. And it's moving not in a straight curve. It's continuing to curve up and it's nearly vertical what this technology can do now. And so the issue is what we call the technology overhang. Are we able to utilize all this technology? And that is really the tough part I think is to get the organizations to absorb and to utilize what we have.
So we are as you know the most transparent fund in the world. And why is transparency good? Well, I think it's good because people can look in and see what we do. And we believe that that creates trust. But I think equally important is that we can look out and see what the world is doing. And I think that's why we have been very very keen on absorbing this new technology because we also talk to leaders across the world and we see what it can do if you apply it in a correct way. And then we thought it was great to invite all of you not because we think we got it right but because we also want to learn from you. And we we think that if we share with you what we know, you will share with us what you know. And the cool thing, we just do not compete. We can cooperate and share good practices across the board and across the country. And I have this dream of being able that we together, you know, the private and public can basically lift productivity in this country. That's a good ambition.
I think we have a lot of user cases across the firm and we could have picked uh you know many many different ones. We picked 10 which we think gives uh a kind of a taste of different things we do. Some of them help us make more money, some make us save money. Some improve efficiency. Some improves accuracy and quality and some just prevents us from doing boring stuff because I think in this new world we should not spend our time doing boring stuff and hopefully you all agree with that.
So what we're going to do first take you through why we are in the situation we are now with our technology briefly by big the journey we've had on AI and how we have tried to make it to permeate the whole organization Ludia is going to talk about uh the the framework we have for doing this in a correct manner because of course it needs to be compliant and it needs to be correct and safe and then Tron together with 10 of our colleagues will take you through a lot of user cases but first big over to you.
Thank you Nikolai. Uh so uh we've done many transformations in MBIM since 2015 but today I want to talk about three major steps that we have done that has built the foundation for our AI strategy. So the first thing we did was to ins insource operations. So before that we had an external vendor that took care of settlement, corporate actions, fund accounting, valuation, everything. But when we scaled into new markets, we wanted deeper expertise and richer data. So the solution for that was to bring it all home. We own the process and we also wanted to own the knowledge.
So the next big thing we did was to move all our IT infrastructure, all our IT systems to public cloud. So before that we used to rent space in an external data center, outsource the technology with it. But what we saw was that we had a data uh ceiling in a way. But what we wanted was to have a data horizon. We wanted instant scale on demand and we wanted to get away from server refresh cycles. When we had moved over to public cloud, we saw quickly that our old databases which we had also moved over did not meet the same requirements. We were not able to utilize this scalability that the cloud provider gave us. So what did we do then? We decided that we had to move our old database solutions over to a modern setup so that we could have the same scalability.
>> And is it fun to clean data?
>> It's no fun at all.
>> So it's the most it's the most kind of boring job there ever is. Does anybody thank you for cleaning your data?
>> No. How do you get people to clean data?
>> You basically tell them that 31st of January we are going to turn the old old data off. Yeah.
>> And if you sit there the day after and have no data, you are going to look very stupid. So we had a lot of late nights and an enormous work by the whole >> that was a lot of tidy up and rewrite of code for basically everyone in the company but uh now we have one place for internal and external data with high quality and it's tidy up and it can be used for AI and it's called martium core. So this is how much I love our new data warehouse. Um so these are the three things that we have done that has been the foundation of our uh success within AI and I will now give the word over to Stian who will talk you through what we did after that.
So I'll I'll take you through our AI journey and having a good data warehouse and a lot of compute in the cloud has absolutely been necessary for us to get on with the AI journey. And it all started about two years ago when Nikolai had Sam Alman from OpenAI and Dario Amodai from Antropic on his podcast and he found out we should be 20% more efficient. So he said can you make that happen? I said thank you. That's an easy target to fix. So um to see how he actually got that working is that him being a a battery durel battery that never goes out of the energy on the top is pushing the organization throughout the last two years and he has pushed everyone. So everyone has gotten the tools they gotten the time they have experiments they created a lot of different project on how they can use AI to improve themsel but that is not enough. If we're going to be 20% more efficient and we're going to start using AI, we have to change our habits. So, we needed to push everyone a lot and nudge them again and again and again. So, we created a uh add upskill program for everyone. I'll take you through that. And we also created an ambassador network to help people get going here. And to keep the nurturing going, we created a this tech year 2025 that was full of activities to help uh bring AI uh to the focus of everyone.
So first of all, we created the ambassador network and you can call it champions if you like it doesn't matter. There were 20 uh volunteers from all over the organization. uh they were given the task of finding the one use case in your team that you think can be a valuable project with AI solve it with the help of the other ambassadors the AI team but also with the help of entropic. So twice a week Antropic help us get started by doing uh creating training for these ambassadors and the AI team for two months. So as soon as you know the project would keep rolling here the uh the ambassador sold the project the turnout showcase their team but also the rest of organization and we can soon see you know the value of AI in you know all the corners of the company.
So we also had to create this tech year and the important message with this slide here with a lot of different bullet point is that whenever something happened in NB NBIM in 2025 it had AI in it. If you had a gathering AI was on the agenda. We had this large tech tech day in London, Oslo and Singapore which focused entirely on technology, tech stack, cloud, the data warehouse and absolutely on AI. But if we had a leader summit, AI was on the agenda. So it kept rolling and rolling reminding people to use AI in the daily work. We trained the ambassadors but we also had to train the rest of the organization. So we've created seven 30 minute sessions with different topics AI and NBIM meet claude prompting critical thinking and responsible AI and so forth and this was done for everyone and this was this is this is mandatory. Now do people like mandatory? They hate mandatory because it's going it's like going back to you know primary school. Can it be voluntary? No. Because the people who don't want to do it are the people who need it the most. It has to be mandatory and you have to be on them like a wasp. Okay, a bit like my wife on me.
So, everyone got the training, they got the time, they got the experiments and they also uh got a help bit of help of support from the AI team. We started off with just three people and we now grow to 10. We are the catalyst. We are not doing all of the AI and NVM. It's definitely the organization that is uh doing it and you will see that from the use cases. We are trying just to enable AI. We are providing the tools the AI and envelopes platform and I'm really happy that we're now at 10 because it's uh only that is a lot of work and feel free to reach out to the guys there in in the audience there afterwards.
So tools first of all we are using claude. Everyone in NBIM is using cloud on a daily basis. With cloud comes cloud code and more than half of the employees are using cloud code to create solution. So more than half of the people in NVIM are actually creating code. We just got Gemini this month uh to complement claude and already uh more than twothirds have signed on and started using it. Curser is also a development tool and about 70% is using cursor but more and more are moving over to cloud code there.
So we have really been through three different phases of our AI transformation. First of all we gave everyone the tools. We gave them the training. We gave them ample time to experiment and from bottom up approach it created a lot of different project in NBM. We had thousands and thousands of project in in um cloud projects just to try it out to get people going. And the second stage is really since we're going to be 20% more efficient. Is there this one use case that can improve all of NBIM or do we have another high valuable project for that we can really um find? We interviewed the chiefs, the global heads, uh, uh, Nikolai. We had workshops. Uh, we found 171 new project. We did not find the one, you know, great AI use case. So the good thing is that we weren't that inefficient before we started this. The bad thing is that we had to do all of these smaller project actually then become more efficient. So a lot of work.
So the last phase is we need to deliver on all the things we have been given the tools and the experimentation and you know the project we want to deliver but already then in the autumn we saw that AI is moving so fast the upskilling pro program is already a bit out of date. So we had to do a second round of upskilling for everyone and we also did a lot of focus on uh cloud code and more the core developers to get that rolling because we saw that that had a ripple effect throughout the company and we think we need to continue upskilling as you know AI changes. Just last week we had a two-day hackathon with the focus on on the core developers again and this just keep on pushing.
The last thing I want to mention is that traditionally we have based our project culture in NBM on scrum methology from Jeff Sutherland and Ken Ken Sber from the 1990s eight developers and one from the business working together creating or solving a a business case they have a lot of ceremonies they have daily standups they have sprint retrospective it takes a lot of time with AI we see that it doesn't make sense anymore it's It's better to have get rid of all of the more or less all the rituals in scrum. Take two developers, one business person and have them work together autonomous but also empowered to take all the decision they need in a project to get the speed up to a whole new level just by utilizing AI here. But what does that mean when we trust AI to do so many things? It means that we need to ensure that we have you know good quality good code quality good deliverance on other aspect from the AI as well. We need to trust what it's doing. We need to follow the rules on how we should use AI in a compliant manner and we have created a framework for that and for that Lydia our AI compliance officer will take you through it.
>> Thank you.
>> Thank you. Good morning. AI is moving fast. It's changing the way we work, how we interact with data, and how we make decisions. Now, at the fund, we're fully aware of the importance of ensuring the AI we use is always done responsibly. And that's why we've built a responsible AI framework. So, what does that mean in practice? Let me show you.
So, first we make the rules. Our responsible AI guideline sets requirements for every employee when we're buying, building, or using AI. The guideline aligns with law like the EU AI act as well as globally recognized AI standards. It addresses some key areas like protecting people's data and ensuring human involvement in all AI systems that help with investments or people related decisions. Now the guideline takes a riskbased approach. That means the way we would handle a simple email filtering system is completely different to the way we would handle an AI system that impacts people.
So the guideline sets the rules, but how do we make sure they're actually followed in practice? Our operating model, this is a document which translates the AI guideline into a functioning governance structure. It sets out key processes that we all need to follow from the development of an AI system all the way through to deployment and beyond. It addresses risk management, legal compliance, security, and lots more. Now, at the very center of this governance structure, we have our AI governance working group. It's a team drawn from representatives from across key teams at MBIM. And the role of the AI governance working group is simple. make sure responsible AI isn't just something we talk about but something we actually practice.
Now the working group keep up to date with regulatory and industrywide AI developments. They talk about AI issues and they find solutions. Oh, I've gone the wrong way. Sorry. Now, one thing that we're very aware of, a governance structure is only as robust as the people within it. And that's why we've trained all our employees on responsible AI. We want every single employee to know what AI currently can and can't do. We want them all to look critically at AI related output and to feel comfortable raising any concerns because after all, responsible AI isn't just a compliance function. It's all our jobs.
Now, the technology is moving so quickly and our governance structure needs to keep up with it. But we're quite confident that we've built something that works for us. A guideline which sets the rules. An operating model which translates these rules into actionable processes, a working group which keeps all of this alive, and most importantly, our people at the fund who with the right training and culture practice responsible AI every single day. Now, these elements together have allowed us to build a culture of responsible innovation. one where we can work smarter, work faster, make bolder decisions, all while staying on the right side of the law and holding ourselves to high standards. So now I'll hand over to Trund who will talk about AI strategy and show you some of our use cases.
Thank you.
>> Thank you. So uh so the key question is how do you turn this foundation into something that has true business value right and then it starts with you know what is your objective and in MBIM it's to achieve the highest possible long-term return in a secure responsible cost efficient and transparent manner that's what we aim for so every year uh not not every year every third year every year would have been too uh frequent we formulate a more short-term strategy for how to do this right what do we set out as our ambition for the next three years. And so in our recently released uh strategy plan uh we mention AI a lot across every function across every department, every individual. This is just one example of a pretty bold ambition if you ask me to cut all our manual processes in half by the end of 2028. So we'll see how we get there. The key is really what we've been through I think uh already the foundation that we have that we have a uh cloudnative infrastructure basically our uh infrastructure is code not physical hardware uh we have a data lake or a cloud data in snowflake uh we have the tools that showed us we have upskilled the organization we have the competency and we have the proper guard guard rails so now it's really up to each and every one of us to make the best use of this new technology.
So we'll take you through in rapid succession uh 10 different use cases three minutes minutes each and we'll start with uh the core of our business which is investing. Oola will talk about how they with a team of five people manage two trillion uh kir in European equities.
Thank you. Imagine being contacted by Goldman Sachs. Ferrari's largest shelder, wants to sell shares worth 30 billion corners. That's more than 3 weeks of normal trading volume. Goldman is reaching out to a few to to a few big investors and wants to know if we're in, by how much, and at what price. They need an answer within one hour. Would you buy? In our team, we get about 200 these requests every year. And with time, these transactions has contributed billions to the funds excess returns. And we clearly see that the better we are at using data to decide when to pass, when to participate, and when to really swing big, the more money we make. Because these transactions, they come with risks and every deal is different. So to decide on the for our deal, there's a lot of things we need to understand like who's selling and why. Was it anticipated by the market? How are similar things gone in the past? What is the fair price? Will this trigger force buying by index trackers and so on. But the challenge is that we have very little time and the data is everywhere. It's in external internal sources. It's in text and numbers. It's in databases, web search and algorithms. And the output of one might affect the output of another which makes this very difficult to automate because you cannot solve everything by code and you cannot solve everything by a language model. We need both.
So we built agents, specialized AI programs with dedicated tasks and tools working together. And to give you an example here, you can see one agent searching the web to find out who's really the owner behind this holding company. Another agent takes a deal text and pulls out the most important data points. And it's all sent through to a third agent which runs an algorithm to calculate whether this will trigger an index effect or not. In reality, there are more agents with more tools. But the point is that in very short time, we have the full decision bases ready with more data and better analysis than before. We started this by building a prototype within the investment team. before we got help from a very talented developer Yan right here who helped us take what we started and build something real. So when Goldman calls we spend less time gathering all data more time analyzing it which leads to better decisions that we make more money. Thank you.
So we're moving on to communication. Uh transparency is a big thing for the fund. Probably the most transparent fund there is. Uh what's cool about our communication department as well as being good communicator is that it's become a lot more datadriven. And to explain a use case saf will take us through. Echo is a live overview of all of our communications activities across channels. We're not developers, but we built this ourselves using AI tools. And the last year, we've been working on taking this from providing stats to actual insights. In 2025, the fund was mentioned in almost 50,000 articles. So far this year, more than 5,000. For a press team of only two people, it's simply impossible to keep track of everything that's being written. That's why we built an AI powered sentiment analysis to help us do exactly that.
So, this is an agent-based system where each article goes through a main agent that delegates to specialized sub agents that classify the article's sentiment or participation in the article, the priority of the media outlet, the article type, and how prominent the fund is in the article, as well as topics and people being mentioned. And all of this data is stored directly into our data warehouse, Snowflake. Existing media monitoring tools are expensive and honestly not that good. So building this ourselves is cheaper and we're able to display the data exactly the way we want it. And here you can see one of the sentiment pages that we built in Echo. This is a little while back and as you can see there was quite a bit of negative coverage at the time. So we built a timeline where you can easily drill down to see exactly who has written what. We also built an insights feature where we're using AI to summarize uh the coverage and provide us with the key highlights. This enable us to faster get an overview of everything that's being written, what's driving the coverage, and where we might need to act. Lastly, we built a chatbot sitting on top of Echo being an expert on all of the com's data. So instead of digging through the dashboard, we can simply ask like analyze engagement on social media. Echobot then dives into Snowflake, fetches data on LinkedIn, Instagram, and YouTube and generates a report. And what's important here is that this is not a predefined view. It analyzes across channels on the spot. It can identify trends and come up with strategic recommendations. Previously, this would require us to log into each of these platforms, fetch the data, and put this together ourselves. So, we essentially automated some of our internal reporting. So, AI has enabled us to build our own systems, automate analysis, and make better decisions faster.
So uh essentially when you run a fund like this you are uh risk managing uh so you've seen as of late markets up and down energy prices up and down that is something that we can stomach that is actually something that we can play to our advantage but if there's one risk that potentially keeps us awake at night it's cyber security risk and how do you keep us in We'll we'll see. Um so part of what I do uh as uh working with cyber security in the fund uh is that I think about how will it look like if somebody's trying to attack us. How will it look like if somebody is trying to steal money and basically defraud us and part of this is that I maintain together with my colleagues basically a giant network of invisible trip wires across all of our digital infrastructure. And this is quite a big uh sort of data collection operation. Just to give you a sense of scale, we collect roughly a trillion data points about MBIM and MBIM's operation throughout the year. And then from those trillion, we surface down maybe somewhere in the range of a million to 100,000 things that might be considered suspicious. And then we have to take uh you know a broad stroke on that and surface only a very very tiny portion that we believe is of high value uh down to myself and my colleagues.
So to give you an example of what something like that might look like is that uh for example an employee who likes uh football might be using their computer to stream uh a football match. They might have browsed to some shady place on the internet and gone to a site that we have intelligence on that says that this website uh also does bad stuff. Um and then what might happen is that we might get an alert. I might, for example, be woken in the middle of the night uh and uh I have to figure out what's going on. I will usually then just get an alert that says this computer connected to a bad place on the internet. And gathering all that context and gathering all the sort of surrounding information from the trillion data points that tells me that this is just a normal user that searched for something on Google and then went to a website and basically constructing that full story. That is sort of a human judgment task where I have to decide in what direction do I go and uh what do I look at what's important and what confirms my hypothesis that this is benign or what is an indication that this might be bad and when I get this phone call in the middle of the night at the same time we also have an agent that starts at the exact same time. So while I'm rolling out of bed uh this agent has started working immediately and it does the exact same process. It's trying to figure out where do I need to look, what data do I need to pull in, and what's important and and it's making all these small judgment calls. And ultimately, it produces uh a report or an investigation, which is very similar to what I would do that I can use to my initial triage. And it does this quite well. It's gotten really really good over the last year. Uh and just to sort of put it into perspective, I would say that it does what would have taken me roughly half an hour in in 5 minutes. And the other very nice part is of course that it's never lazy. So stuff that might be repeating and be very similar, it's uh does it with the same amount of work every single time. Thank you.
Uh so what one of the uh key advantages I would say that the fund has is its size and its long-term investment horizon which makes that makes us an attractive partner for our invest. And so we have very good access to the companies and have dialogue with these companies. We meet with the chairpersons, the CEOs and the lead groups. And so when you have those meetings, you want to make sure that you make the most of them. So in preparing to do for those meetings, we have developed a AI use case that Christina from London will talk about just now. Christina,
>> hello from London. Um, so we've been running a very exciting AI project over the past few months and it really improves what is a very important process for how we invest in equities and engage with the companies in our portfolio and it's been a close collaboration between the AI team, the portfolio managers and the stewardship team. In 2025, MBIM held more than 3,000 company meetings. Each of these meetings take about 3 hours to prepare. So this is nearly 10,000 hours every year that we can spend more effectively. And this is something that was very important to us to build ourselves. Firstly, it plays to our competitive edge. As Tron said, we're a large long-term investor and that means we get unique access to company management. And secondly, we have a very distinctive approach to company meetings and we have extensive training in interview and interrogation techniques and that's something leading external solutions couldn't really replicate.
So here you can see an early version of our solution. It pulls up the companies that were meeting over the coming weeks. So you can see Meta for example. The model loads data that only we have access to investment hypotheses and meeting notes. You can then choose which AI model you want to use. You can add guidance. You can attach documents and it then feeds into a multi- aent system. So you have one agent builds a plan. Then three to five sub aents go out and research different sources. A final agent gets the output and this agent has been trained on our very best meeting prep examples and internal interview technique materials which then evaluates the input and decides whether it's good enough or not. So here you can recognize in the output the prompt that we entered. You can check the sources that it references so you can make sure there's no hallucination. And you can also recognize our approach. So questions that build rapport with the company and that keep a long-term focus. And we've also then made sure that we can iterate on the agenda in a chat. And we're going to continue to develop the product and we'll soon add a simulation component. This will use podcasts, part past meetings, and company communications to predict what the counterpart is likely to tell you. So, this makes sure that you'll be able to both refine your plan, make sure that you really get what you're after in the meeting, but also get feedback on how you run the company meetings in general using speech to speech. So in short, 10,000 hours better spent. The model gathers information and builds structure so portfolio managers can focus on the strategic questions. It will be higher quality. We've trained the model on our very best examples and the simulation will make us all better. And finally, it helps us make the most of our competitive edge. Thank you.
Okay, so the fund has uh millions and millions of transactions in uh more than 60 markets every year. These are highly regulated markets. So we want to make sure that all of these trades that we do are legit and comply with the market regulations. So Oscar, how do you use technology to improve on that one? The risk is real. Cases of insider trading, market manipulation are not rare. We read about them often. We're also seeing active enforcement uh in the Nordics. So, this this is a this is a big deal. Market integrity is vital for any market participant and for an investor like MBIM, documenting proper compliance is foundational. So, how did we get here? Well, with increasing uh focus on stricter regulation, the buy side uh like MBIM was now required to demonstrate Odin train monitoring capabilities. Traditionally relying on the sell side banks uh to cover this for us. MBIM selected an external system in uh 2018 that uses advanced market risk models and spits out alerts that the compliance team can manually investigate. But this system, it doesn't know MBIM. It doesn't know if a trade was as a result of a rebalancing, an index event, or if we met the company the week before. That's context that's carried manually by the compliance team today. And alert clearance, to be honest, is process work. You're checking for the same type of things. You can get fatigued. We're spending time on false positives. And this is what we're now changing. Introducing our AI surveillance team. Uh we have six sub agents that are each reviewing all alerts spit out from the system. Uh they specialize in trade context, index rebalancing, company news, industry news, timing patterns, and company interactions. Uh and they assess every single alert simultaneously, the same way all the time. Never get lazy. They all feed into a master agent which we call our enhanced vigilant agent or Eva. Uh she's always on, completes a full audit trail for every case assessed. Um uh every alert is assessed simultaneously and she's an expert on pattern recognition. Uh and there are always three cases where uh our master agent hands over to a human and that is when the case is genuinely ambiguous. Um it's also when you uh can't automate judgment and lastly uh a decision has to be owned by a human and that's where we come in and those cases get sent to compliance. We have the same team but with vastly better coverage. Thank you.
So uh the starting point for our management of the fund is a benchmark that is given to us by the ministry of finance. That benchmark contains roughly 7,000 companies. So if uh we were to implement that benchmark, we would buy in equal share roughly 1.5% of each of those companies. Now the question is, do you actually want to own all of those companies? And Martin, do you think so?
>> Probably not. So um in forensic accounting, we have a little bit of a challenge. The challenge is to remove the bad apples from the fund. And on average, we spend each analyst spend about two weeks doing a deep dive on each of these companies, looking through the financial statements, the footnotes. There's millions of pages to read in all the 17 7,000 companies that we we look to sell from. So what we need to do is to remove the financial makeup that uh makes these companies look better than they really are. So, we need to clean up the accounts, so to speak, 16 years back in time for all of these 7,000 companies. And then we need to train the machines to learn to spot these um this type of financial makeup. So, how do we do this? Well, one practical example here is that we we're looking for different keywords. So here's a payables extension is what we're looking for. So we have found a keyword here in the in the footnotes and then AI is picking out relevant sentences uh in those couple of pages before and after and then it picks out the number we're interested in. So in this case uh a doughnut producer has extended her payables by $745 million then we store that and we learn from it. So we're making many different kinds of agents that is uh trying to spot this kind of financial makeup and uh this is uh the companies moving revenues and uh costs, earnings or cash flow back or forth in time to look better than they really are and the changes can be um quite substantial for some of the adjustments. Then we're having a machine learning model to learn from all this data. So what we have made inhouse, we have quite a unique data set where we have gone through every forensic accounting research shop and gotten all of their historic cases where a company has um put makeup on their accounts and when the market found out then the share price has fallen quite a lot. So we're having thousands of these cases and we're training the machine learning model to spot uh similar cases. So the output of the model is uh the percentage likelihood that a company is going to develop into such a forensic accounting case and the share price is going to fall. So this model is in production and we we have it used daily in cooperation with Ulusha and and his team and we're working full speed ahead to make more and more of these agents to spot more and more of the financial makeup. Thank you.
And now for financial statements without makeup I guess. Uh so this is um Turus. He will explain how we uh on a regular basis disclose uh a fulls compliant uh financial statement based on the millions and millions of transactions across uh a multiple array of different instruments. How do we do that? Every quarter we produce NBIM's financial statements, notes and analysis. It's a solid process yet resource intensive. Think complex Excel workbooks with long formula chains and lots of manual overrides and adjustments. This um sorry keeping the quality we expected from ourselves took a lot of time and effort. This time and effort came at a cost. We spent so much time in production mode that little were left for deeper analysis and insights and critical knowledge lived within few individuals creating a dependency we really wanted to move away from. It's a high quality process that actually deserved a better infrastructure. Sorry, this is the slide I was supposed to show. Now on to the next one.
Good processes can always get better. Ours were no exception and AI gave us the tools to do so. We decided to build from scratch. Together with fund accounting, we started with the underlying accounting data to establish one single source of truth. We made sure the data was clean, structured and reliable so that the AI tools could do what they do best. Our team of two are not developers. So we use cloud code and cursor to write the code. Even the most complex calculations and aggregations now run directly in our underlying customuilt data sets feeding automatically into our note not note production, financial statements production and analysis. We kept human control throughout bringing accounting expertise and business logic to make sure the output was correct while maintaining internal control even with a new way of working. The result is a platform that's already delivering better analysis earlier and faster production. Our analysis are now available even before the accounting system closes on business day 10. FX and tax at the push of a button on business day 2. Secured lending and borrowing on business day 7. All of this gives us a time to actually investigate and correct before it's too late. Full automation will free up eight days across 2.5 people in our little team previously dedicated entirely to production. This time we can now spend on analysis control quality assurance and audit can move earlier in the process. A good example of this is note 14 collateral and offsetting. a note that I think very few actually reads. Yet, one person spent one full week producing this note. Now, it's done in only a couple of hours. We're just getting started. Across NBIM and together with Noius Bank, we're implementing a new reporting tool. With this in place, we'll have fully automated processing from transaction recording to official external reporting for a large part of the portfolio. uh to showcase even further where we actually are today. AI enabled us to automatically produce note four income and expense from equities, bonds, and financial derivatives and notes 11 foreign exchange gains and losses in the annual report 2025. So these are beautiful notes. Go in have a look. I recommend it. Thank you.
and and and without makeup.
>> Thanks, Chris. Uh so we go to London again for Christina. Uh responsible investing is a big thing for the fund, right? And so we have deployed new technology also to improve our risk monitoring when it comes to ES and G factors.
>> Hello everyone. So, as has been mentioned, we invest in over 7,000 companies across 60 countries. And as a responsible investor, we need to know whether any of these companies are linked to serious issues such as forced labor, child labor, deforestation, corruption, and so on. And this means we have to screen all of these companies. But if a human analyst were to do that, we estimate we would need 3,000 analysts working an entire weekend and our team is only eight people. Therefore, we have to leverage AI to help speed up our process. And today, we're going to walk you through one example of such a screening process.
So, today with AI, we can screen companies at a scale we simply couldn't before. The concept is simple. For each company name, the AI searches across any publicly available source, think uh news, financial data, government records, and local media, and outputs a structured risk assessment. The key to making a system like this really valuable for us is that our expertise is involved in every step of the process. The screening process runs in two phases. For the first phase, we use a lighter and faster AI model. And here, the AI looks for any indication of a company's involvement in the issues we screen for. And as expected, most companies clear this stage. But when anything comes up, it triggers phase two. In phase two, we deploy multiple AI agents, each responsible for researching the company from a different angle. This means that one agent could be tracing supply chain links. Another agent could be looking at direct operations while a third agent looks at financial relationships. Once all of our agents are done working, they come together, summarize their findings, and give each company a risk score. So for step two, our human analysts jump back in and make the decisions. We review every company that the AI flagged as high risk. This means that us humans, we verify the sources, we check the reasoning, and we make the final decision. Now, if the risk is confirmed, the high-risisk company is flagged in two of our internal portfolio management systems, Polaris, and in our investment simulator. In addition to being flagged, we can also choose to engage with company management or in the most serious cases for the small companies in our portfolio, we can choose to do a riskbased divestment. So today, we can use AI to screen many more companies so that our team can focus on what actually matters. We can of course not promise that we catch everything. But what we can say is that we can screen more companies, look at more sources, more languages, and go more in depth than ever before. So that when we see a risk, we can act.
So uh one department that is really all over this new technology is our legal and tax department. So Christie will join me now on stage and talk about the use case from our department.
>> So what if you could walk into a negotiation already knowing the other side's strategy and how to redirect that to get the terms you need? When I was working on an important contract, I saw an opportunity. If AI can model language patterns, then it can model negotiation patterns. So, I created the negotiation simulator. It has two key functions, planning and simulation. Planning mode. It creates a written strategy plan that helps anticipate the arguments and find an optimal combination of terms and concessions. And so far it has delivered. We're able to predict over 80% of the arguments. We achieve a higher ratio of our key terms and these are the terms that have significant risk and revenue impact on the fund. So it matters. We also create a written plan a lot faster than we used to before. So we save time and we can focus on more strategic issues. But simulation is what sets this apart. In voice mode, I can simulate a live negotiation in a risk-free environment. I can also ask for feedback to improve my skills. And I can even switch roles to see how AI would handle the situation if it were sitting in my seat. Um, here is a short video showing you, I think, aha, how we do this in action with the scenario for a price renewal negotiation with a software license vendor. We're ready to begin the live simulation for your software license renewal negotiation.
>> So the situation is that we have had some service issues in the past year.
>> 20% reduction is quite substantial. That's a significant ask for us.
>> I need you to add on to that a higher rate of discount.
>> I'm not going to say it's impossible, but if I'm going to agree to an exit clause, I need something concrete in return.
>> Okay. So he's a tough negotiator, isn't he? All right. But that is just one way that I'm leading my group to go beyond automation to innovation. When I saw our contracts portfolio, I saw not just one more opportunity, but multiple opportunities. And why is that? That's because everything we do here connects to a contract. Contracts contain a lot of important information of course, but if you look at those contracts at scale, you can find patterns, changes over time, connections that you could not find otherwise. So now we're able to do that using AI and we're able to do it at scale. For example, we're looking at force masure clauses and our combined impact under different scenarios. We're also trying to look at how different triggers across our investment contracts might impact or delay our access to credit, collateral, or other issues. So, those are just some of the ways that we're using AI, but we're not just using it to help us negotiate contracts. We're using it to help us extract strategic information that adds real value. Thank you.
And so for our uh 10th and final uh yel will talk about how we inside the fund manage all the transactions that cross different portfolios. We have 250 different portfolios. So portfolio managers choosing to buy or sell something uh you know at different points in time. We have cash coming in and out of the fund and we have this benchmark that also changes on a monthly basis. So to control over that and even better actually try to make money out of that or not lose money. How do we do that? Thank you.
Now the fund has grown substantially over the years luckily right but also means that our trading needs grow along with it. We are now in a place where we ourselves become our own worst enemy in the market. When we buy, we buy so much, we push prices up. And when we sell, we sell so much, we push prices down. This footprint in the market, this market impact as we call it, we estimated last year to 14 billion of region. That's a substantial number. Okay, it's not a accounting cost. So it doesn't so show up in your notes. So it is but it's a significant and real value destruction. It hinders us in doing the assignments that you all have a task us for doing. So what do you do? Here are some ideas. We could for example predict the market. AI helps us find patterns in the market and give us probabilities for which stocks will probably rise in price and which stocks will probably fall. So when you give me an order, I look at it to say, okay, you ask me to buy something, but probably it would rise in price. I need to step up a little bit and be a little more active. But on the other side, if I predict, AI predicts the stock will actually fall in price, I can be patient. I don't have to gun it. I don't have to push the metal to the pedal to actually do that with a lot of impact. I can step back. Okay. So, predictions build patience.
Another thing we can do is to look at all the internal portfolios. As mentioned, we have multiple different strategies internally and sometimes they touch the same companies. They touch the same stocks for fair reasons. But if someone is giving us an order on a Monday and we have reason to believe that we might get the opposite later in the week, we don't go to the market. We park it instead. So right now I checked the parking lot this morning. We have 10 billion parked right now. Last year we parked more than 120 billion and we don't go to the market. We don't pay tax twice. We don't pay commissions twice. We don't push it up first and then push it down later. We just sit on it. We utilize the balance sheet and the risk bearing capacity we have to just take a step back.
And finally, this is super exciting. The way AI helps us improve the way we work. Okay, so everything we do has always been driven by a solid investment process as a foundation and then human understanding. We have great people. They really understand the market. What's driving the various company? what's going on in the news in the macro in politics and we understand deeply what's going on and then AI sort of came in recently as the icing on the cake was cute right but what's happening now is that that arrow of knowledge is going down right AI is helping us as human beings to understand more faster deeper red teaming analysis providing different perspectives looking at things from multiple angles and then also we are improving our core fundamental processes not only how we do them i.e. automation, doing the same stuff faster, but also working smarter, doing better work, right? That's quite amazing. So, then we're going back to where we started. The problem with the 14 billion, is it a lot or a little? It's clearly a lot, right? But had we run 2025 with the cost structure we've had just a few years back, that number would have been closer to 20. Is AI part of uh is AI all of that saving? No. But it's a significant part of how we have saved four, five, maybe even six billion croners in the way we trade for the fund. Thank you.
Thanks. So that was uh 10 different examples. I hope uh it was useful. At least that's where we are today. Now this technology is moving so fast. So every week, every uh you know month there's a new model out, there's a new opportunity out. So uh stay tuned and we will certainly certainly tune our organization to make the most of this uh technology.