NVIDIA's Georgios Kolovos on Revolut's PRAGMA, Transaction Foundation Models & the Agentic Banking!

Speaker A: Foreign.

Speaker B: AI native bank or AI native fintech? Is it a distant reality?

Speaker A: I'll say those who adapt survive.

Speaker B: Okay, and what about agentic? Is it a matter of adapt or die?

Speaker A: Or invest?

Speaker B: Or invest? That's a good one. You were at gtc, obviously this year we which industry showed up the loudest,

Speaker A: had the biggest presence of course, financial services. And for Those don't know, GDC is the Nvidia's developer conference that happens once a year in the US and we bring it into Europe. So October in Berlin, so I hope.

Speaker B: Oh wow. Oh, in Berlin. We in financial services are celebrating this moment where Revolut has published its foundation on model pragma. Is it a game changer for the industry or only just for a few players?

Speaker A: It's a game changer because the insights are uh, for me the moat for the future where companies will win or will lose.

Speaker B: And what about tier one banks? Is their future, their near term future really off the shelf foundational models? Yes or no?

Speaker A: Depends. This is a strategic decision that each bank needs to make.

Speaker B: Okay, and what about a financial world model for finance? Is that a fantasy or do you think that's the vision?

Speaker A: I'm sure that people that are building such models. So if you're building a model like this, call me, okay?

Speaker B: Okay. Hello everybody. I am your host Effie Pilarino and today I'm sitting down with uh, Georgios Kolovos who is Nvidia's, uh, EMEA Payments and fintech leader. Georgios, welcome. Thank you for joining us, Effie.

Speaker A: Thank you for having me. So lovely.

Speaker B: You're one of the clearest voices. You work with the uh, industry players that are actually biddling as the tokenized world says. There's a lot of building going on behind the scenes and I'm sure we'll have the opportunity to, to talk about that. Let's start with what happened at GTC this year because there were, I would say more than half a dozen important fintech and with a broad sense players there give us a sense of the atmosphere and what it really means for this vision of an AI factory for

Speaker A: our industry, for those who uh, haven't been there definitely need to experience it. So GTC takes over the entire San Jose. So the entire city is painted green, the Nvidia colors. And this year was over 35,000 people that joined the event, including everyone and all the key and AI companies. Financial services typically has a very strong presence, but this year was particularly strong as it was the largest uh, industry represented in the event. And we had a very clear idea of where it's going. Financial services for Nvidia, uh, typically we looked into a number of different three segments. The one that I spend most of my time is in payments and fintech and in payments in fintechs. We noticed there was a few very interesting signals. One is this notion of converting data into insights and the other is the agentic revolutions, how this insights then are used by agents. So at the event, as you rightly said, we had a very strong presentation from Revolut and Adyen from Europe. We have Stripe and PayPal coming also to talk about each of these areas. And all of that comes into the, as you say, the AI factory framework. For the people that are not familiar, AI factories, how do you look holistically? How do you go beyond just the application layer? How do you set your infrastructure, your business to be able to benefit from the broader, uh, shift towards AI enablement of these AI capabilities? Looking at the application, looking after all the models and this can be the LLMs, but also can be these foundation models that we will talk, I'm sure around the transaction foundation models, but also how are you thinking about the infrastructure, how do you think about the GPUs that you need and how do you build this full stack? Jensen talked about five layer cakes. So this is the five layer cakes, the companies that will be the winners in this AI world with the companies that invest across the entire stack.

Speaker B: Interesting. The names that, that you shared are familiar names and we know that they are really leading beyond the incumbents that are also investing heavily. But let's go take a step back because I want to get your perspective. When we say we want to become an AI native player, be it an M AI native bank, an AI native fintech, paytech, uh, and others use terms like an agentic bank. There's already books out there, the agentic bank, and there isn't any consensus. I think different people understand or think different things when they hear this. I wanted to hear your perspective. What do you think is an AI native financial services player or an agentic? Do you make that distinction? What's in your mind for me to

Speaker A: make a distinction is how do you use this technology? Are you going to buy things off the shelves or are you going to build things yourself? That gives you the competitive edge. So the two options are quite viable and this is a strategic decision that each organization needs to make. Is this uh, agentic first? Is it AI native? For me, all companies, it's one form or the other. And M, we did A interesting survey around financial services institution around the adoption and evolution of how companies are investing and where they are looking into AI. The numbers were pretty telling. Every, pretty much every single organization is looking to invest more this year than in the previous year. So what label you put on top of that is I don't want to say irrelevant but it is. What do you do underneath? How do you prepare your platform to what you discuss about the AI factory But also what are the type of talents that you bring into the organization? But also how do you upscale your existing talents? How do you augment to work with AI tools and agents that help the businesses to be more productive? And we've seen this evolution coming from it initially was a lot of a this will bring a lot of efficiency. This will help us to manage costs, to manage operation more efficiency. There was a then big foray of different use cases experimentation. There was a few famous reports there quite a lot of pilots didn't POC came to fruition. Now we are moving to the phase and the more and more I talk with the board is around how do you embrace it? How do you actually start thinking about changing your business models? What this new capability come to provide you? Are they created in. If you are uh psp how do you create new value added services to deepen your relationship with merchants? If you are in kind of more in issuing on a fintech side how do you add more value to your consumers with these capabilities? Again for me the labors yes there's interesting and interesting way people are positioning themselves but ultimately it's what you do underneath.

Speaker B: Interesting. We've clearly seen some of the most common use cases be it in fraud, be it at the call centers, be it at customer service, be it in the marketing area. I think those are the more common use cases that have been but fraud

Speaker A: pretty much every engagement that we have starts from fraud. And fraud has been an area that machine learning has been used for a long time and the latest deep learning techniques are always starting from there. The evolution that we are seeing is yes in fraud there's a lot and our platform can accelerate a lot of this. So Nvidia is not only a GPU companies but also have a lot of software library that accelerate a lot of the workloads. So when you apply this for fraud you can see the faster you process the data, uh, the faster you train the models the better model you use better protection you provide for the company. So immediately this is very clearly you can articulate the ROI from this from these initiatives where it Becomes quite interesting is how do you take it to the next level? How do you share all this goodness that comes from and the experience, the models, the insight that you generate from the fraud side, how do you spread it across the organization? And this is some of the, the work that Revolute and I forgot to mention when I mentioned all the companies in GTC MasterCard was also there. They made very interesting announcement the same. So it's not only for fintechs, it's also for companies with quite established infrastructure

Speaker B: with quite incumbent in a way or. Yeah and natively born hyperscalers like Revolut or Nubank and so on. Having said that we uh, see this emergence of these transaction foundational models that are trying to be more unified and multipurpose models and we'll talk afterwards about uh, Pragma and some others but share your thoughts around this tug of war between specialized models and these more multipurpose models. Do you see a world where we'll have both some companies really using these multipurpose developing them and using the multipurpose ones or will most of the industry be with these specialized models and somehow have a uh federated network of agents that are managing these. Where do we stand?

Speaker A: So it's yes this transaction foundation models the essence of these models, they're predictive models that work on predict what is the next transaction, so what is the next action based on the prior historic interactions that the customer has with this, with this particular financial services institution. So obviously to build these types of models you need to have a critical mass first of customers and also you need to have a uh, relationship with these customers over time. So it provides you a number of different touch points which will enable you to make a prediction based on the priority behavior. If you don't have that, if it's your business is a one off engagement with a customer there are other techniques that are much more relevant for you to build and for example graph neural networks are very good that you don't compare and try to predict based on your prior engagement with the customer. You're trying to predict the behavior of the customer based on how you compare this customer with all transaction without a customer transaction with similar patterns. Right. So this tabular um, foundation transaction foundation models are not for every use case and on the Pragma paper you see the uh, Revolut team specifically called out uh AML as an area that these models are not necessarily performing better that are more a uh, different methodology. Just to be to clarify these models provide the context these models will be Able to articulate to help you to understand what potential next transactions will be with a customer. But these embeddings that hold this information and then used to infuse all the other models that today exist, this will be maybe your fraud model or maybe your cross sell model. So it is not one universal model, it's the embeddings that power a lot of the models. What is why this. There are two main things that are quite interesting for these models and uh, this is why we see a uh, surge in company adopt this first allow you to break silos. And anyone that works in financial services knows how many silos are there in any organization. The defense size, the risk compliance, very rarely interact and talk and share data and insights with your growth sites. So doing these transaction foundation models enable these two sides of the organization and there's tension in between to interact and provide a common context and operate from the common baseline of the interaction of the customer. And then the other important uh thing is that this is a self supervised model. So traditional models, you train the models, you test it and you deploy it. Right. These models continue building and evolving as the, and learning as more transactions are coming. As I said, they are not easy. This is not happen just out of the blue. This transformation is built over, let's call it 30 years of evolution of using machine learning and deep learning in financial services. But in beyond with the evolution of the transformer infrastructure, et cetera, what we are seeing there was a signal when last year Stripe, Nubank, ATI and all were talking about this. And now you're seeing the next wave of this. Companies are building on top of the research. And this is quite interesting. A lot of the research is public, right. And each step in bringing us to a little bit closer to a better model to finding a new insights. Where do we need to research more? Where do we need to invest more?

Speaker B: Nvidia I saw has published research on these transaction foundational models. Right? And then there's Revolut that also published its own research paper. Not obviously with the weights and everything, but it's a blueprint. I think Nubank did that but with a lag and maybe we'll see see more of that. What I'm hearing from you is that these transaction uh, foundational models are not replacing or disabling other specialized if you want models or existing ones. They are actually connecting to them and enriching uh, intelligence or the context. Is that correct?

Speaker A: Correct. This is the intelligent layer as you say. So we are talking before about data layer. So these New models are converting this data into insights. You can call it connected insights because comes from a number of different sources but also the more you use it in more use cases, the richer it becomes and the opportunities are endless. Right. So everything that has happened so far is very much into the remit uh, of the existing organization. You can imagine how these embeddings can become also a potential new revenue streams for companies, how these embeddings can be used to connect to third party propositions. Let's say you want to deploy a uh, voice AI customer service. You can use some of the embeddings to provide the context of the customer. So there's a lot of opportunity and there's definitely a lot more research and a lot of more experimentation that need to do that. But the signal is very clear these silos that exist today in financial services doesn't help the companies to be competitive and provide better services being to merchants or to consumers. So these transaction foundation models are ah, stepped into the direction to unify the insights, provide the context and become much more personalized in the interaction with the customers.

Speaker B: So do you see this trend becoming an industry standard? You mentioned before that it's not for everybody, it's in areas where you have large volume and frequent. So do you see that big tier one incumbents just to throw some names, JP Morgan, bank of America, BNY Mellon, those type of size of players will go in that direction and maybe they are and it's not just public. Do you think that's the industry trend?

Speaker A: A lot of the banks are already investing in intelligence in one way or the other and some banks are quite ahead of the curve. So it's not only that this is happening in isolation in the financial services in the fintech space. Also a lot of banks are investing uh, and has been investing for time. What was interesting actually the one of the most. I don't know but one thing that I'm quite happy with the work that Revolut did that it's not only about when you talk about, when you refer to transaction foundation models, you think about transactions as the endless source of insights. What Revolut has done quite smartly is adding events. Data and events can be anything, right? So your interaction with the app, uh, the receiving of a marketing communication, your application, your customer calls, then these data sets become much more richer and provide you a better, better context. Because transactions, yes, they're good but it doesn't give you a lot more so the behavior comes from these events. So if you think of transaction foundation models Being built not only on transaction data but also you'll be able to augment them with other events. Then becomes not a question of how big is the organization but how many touch points the organization has with their customers.

Speaker B: Yes, that, that's why we're talking about an event sequence based uh model for Evolute which is probably not the case for other models. For example mastercards very evolve fraud model I don't think is based on Revolut

Speaker A: is quite, quite interesting. It's a very different context. Right. MasterCard is a network is a uh, customer B2C proposition that has quite a strong suites of products and subscriptions. So have a lot of different touch points with the customer. The context is very different. And this is where uh these foundation models you need to find what's the purpose if any organization thinking to building it. What is the purpose why you build. Are you building it to deepen and make your relationship with the customers better? Are you building them to actually help merchants become better? Merchant acquiring is another one and Adyen has been doing some amazing work on um, on their models and the foundations of the insights. And merchants acquiring is also an interesting space to explore and you can take it even further. You can see what Shopify is doing for example which uses similar technology that connects actually from one side you look how to help consumers to find what they're searching but then on the other side you help merchants to figure out what are the customers that are coming at a good or bad customer. So the application, the concept and the methodology that is used is quite interesting. You've seen Plaid also made an announcement recently about the use and they are using it very much to help them to classify transaction and engagements in a better way. So you see a lot more experimentation. You can see these foundation models move to uh, different domains. Maybe there's a uh foundation models, Transaction foundation the B2B space that will come. Maybe there is other financial services domain that will leverage the same methodology which is hey if you have sequential. If you have a time engagements with the customers which provides you a lot of insight. How do you convert that uh to provide you the context of which all three or other models and your overall interaction with the customer operate.

Speaker B: You're talking about what British people call Plaid and others may call Plaid a classic uh. What's the right pronunciation? Is it the British one or the American?

Speaker A: If you want we can record again can call.

Speaker B: Depends who you're talking to for this. Which brings me to a question that is important to address. Obviously legacy financial services providers Even those that have gone through um, let's say successful digital transformation over the past two decades, they still have legacy, they have silos, they have fragmentations, despite any digital transformation. However, what about digital natives, the revoluts of the world, the ones that were born 10, 15 years ago, Cloud native, modular and so on. What difficulties or challenges are they facing in their AI transformation journeys? What are you seeing?

Speaker A: I don't, I would not say it's only unique to this digital native or cloud native fintech businesses. The world is quite complex at the moment. You have so many different technology that are emerging, AI being one of them. You have everything happening on stable coins and sovereign rails and complexity of operating. So it's, I think that one of the bigger challenges is, I see is how do you make the strategic decisions? How do you decide where to invest an investment? I don't mean only investing in the infrastructure, in the GPUs. Uh, it's also investment in talent, investment in education, your existing stuff. Because any decision that you make around those technology means that you need to shelve something else. Um, so I think this is the bigger challenge and you add into that all the uncertainty politically there's tons of legislation, changes that happens. It's very difficult to make the decisions. My, my view is that you need to make a decision otherwise the decision will be made for you. So the company that we were discussing for, they have a very clear directions where they want to go and very clear, clear goalposts around. Um, what kind of, what are the steps to get there? The company that has a little bit trying to, oh, we want to go there but we will test a little bit will not put 400%. This is normally the tip companies either that not moving fast or they're just completely lost, losing their path. And there's different strategies. Right. So you can be, you want to be leading the charge, you want to be seen as innovator, you want to invest or you may take a different approach. You may say I want to be a uh, fast follower. So I would not bet very heavily on everything that is hot. And in the financial service in the tech industry there's always the next hot technology that is around the corner. Uh, you may wait and figure out. For me it's not the distinction between exist the established players and fintechs. It's more about the clarity and the vision and beyond that is the also the execution. Right. It's good to have the vision but are you actually investing along the way to make this happen?

Speaker B: Yeah, indeed. Georgeous Gentle one talks a lot about world models especially in physical AI, robotics and so on. Should we start thinking about a uh, financial world of model? Is that something like a vision that should exist that eventually we will get one of those financial world models?

Speaker A: There's so many things that I would love to do. This is one of them. Financial services is very unique. Uh even if put the umbrella financial services. What do you mean by financial services? Is that the capital market? Is it the payments infrastructure? Is it B2C B2B. So I think there's plenty complexity, plenty of challenges to overcome. I'd uh love to be able to do even a digital twin of the entire payment ecosystem in a particular country. Imagine how rich this data can be and how powerful this can be to predict not only if something goes wrong, what will happen. So you simulate different scenarios but you can predict movements of people, you can enable it to develop, accelerate innovation in the particular market. I think it's a complex way but there's if you can scope it in the right format, I believe there is an opportunity to do something interesting.

Speaker B: Yeah, I've always dreamt uh on board of Star Trek with a uh dashboard that you really see all the money flows and can really do you need

Speaker A: a dashboard in our days. I uh. This is one of the things that is fascinating for me. It's your bank uh app applications right. So mobile app still everything is very static. Everything is dashboards related where there's a very few banks that completely has embraced this chat based interface with agents. Not only for customers service but also for anything else that you do with the bank. So your starter AC analogy for me is static. You need to look at the dashboards, you need to make a decision where there's another word that someone can help you. You can have a beautiful window look after out of on the in the space and someone tell you what what

Speaker B: you need to pay attention to be attention actually what you say. I only know of one bank in Malaysia that has actually launched this. It's not only conversational for customer support, it's like for everything you want to do and maybe we will see more of that. The question for me of course is how much intelligence is behind it because that matters more than the customer uh

Speaker A: experience and it's also for the customer adoption. You think about how long the take to customer to accept contactless payments. You see a wave of evolution of that. But when you see the little button typically in the bottom right corner of every banking app thinking about you kind of customer service chatbot. I think this will evolve and soon will overtake the entire screen. Soon is a, um. Is it 12 months? 18 months, two years. But I think this is the direction of travel.

Speaker B: Yeah. A hundred percent. Georgios, thank you so much for sharing your insights. And let's talk after GTC Europe.

Speaker A: I hope to see you there.

Speaker B: Uh, okay. Thank you.

Speaker A: Thank you.