Will There Be Bankers Left in 2030?

Read the Transcript

Will There Be Bankers Left in 2030?

Mathias Schutz: [00:00:11]:

I think we are on air. Hello and a warm welcome, everyone joining us for today’s webinar, with the title, Will There Be Bankers Left in the Year 2030? My name is Mathias Schutz. I’m hosting today’s session with my colleagues here, which I will be introducing a bit later. I am CRO and Deputy CEO of Natech. Natech Banking Solution is a leading banking technology platform in southern Europe. We’re headquartered in the emerging technology hub in Ioannina in Greece. We serving financial institutions across Europe with future-proof, cost-effective, end-to-end design banking solutions.

As you have seen from the invitation, Will There Be Bankers Left in the Year 2030, this panel discussion aims very much to be a bit provocative, a bit of provocative thought leadership discussion around the future of banking technology with a focus on the real world trajectory of AI in financial services. Expect a vibrant discussion. For the very first 45 minutes, my guests will be discussing with me various questions around AI, agentic banking and stuff like that. We will be keeping it very interactive. For the last 50 minutes of the session, we have reserved some time for you also to ask question. Of course you can ask questions at any point in time. Feel free to add it in the chat and we will be picking those questions up for answering.

For administrative purposes, please note that this session will be recorded. That means for the sake of good order, if you stay in the call it means you will consent with the recording. Thank you very much.

Now it’s a pleasure to introduce my guests in the panel to you. I’ll be doing this in alphabetic order of the company that they’re representing. I’ll ask each one of the guests to do a quick introduction of himself, including the company they’re from. To start with the letter A, Additiv, a warm welcome to Yann Kudelski. Yann, over to you.

Yann Kudelski: [00:02:37]:

Thank you, Matthias, and great to be on this panel today. As you said, my name is Yann Kudelski. I’m leading strategy and business development at Additiv. Additiv is a Swiss-based financial technology platform, and we serve clients across Europe, Middle East and Southeast Asia. What we basically do is we offer an orchestration layer that runs business logic, abstracts, process logics, and orchestrates all processes and workflows on our platform. We embed AI across our processes and workflows when and where necessary, to automate the workload as much as possible. Back to you, Matthias.

Mathias Schutz: [00:03:20]:

Alright. In order of the alphabet we don’t continue with B, we continue again with A. We have from Atfinity, Thorben Croise. Please introduce yourself.

Thorben Croise: [00:03:31]:

Thanks, Matthias, it’s great to be here. I’m Thorben Croise, I’m the CTO of Atfinity. We are a FinTech startup in Zurich that automates business processes for banks, like onboarding or KYC reviews. It’s great for us that is AI hyped, because we’ve been using AI for almost 10 years when I founded the company and nobody understood what we doing, and now it’s this two letters and everybody has an idea what we’re doing.

We’re not using exactly Gen AI to automate these processes, but we’re using a different AI that we also custom developed 10 years ago. But you can imagine any business process, a huge flow chart, and we create this flow chart automatically. We don’t force them somewhere to stand there and put it on the whiteboard so a development team can do it. From the outcomes of a process, we can automate the process itself. Yeah, we work with AI a lot. We are adding Gen AI to the platform now in three or four different products at the same time. I’m looking really forward to see what everybody has to say, how we can use that. Thanks a lot, Mathia.

Mathias Schutz: [00:04:32]:

Great, thanks a lot, Thorben. We are jumping to the letter D, Deloitte. We are having Federico Crecchi with us. I’m very pleased to have also a representative from a big consulting company showing the expertise from the market side. Please, Federico, over to you.

Federico Crecchi: [00:04:50]:

Yes, thank you, Mattias. I’m Federico Crecchi, I’m a director in AI & Data in Deloitte, Italy, and I focus on the banking sector. We support banks in both revising their processing and their organization for adopting AI and Gen AI, but also working with them to develop a solution that can be put in production and used internally or towards the customers.

Mathias Schutz: [00:05:28]:

Good, thank you very much Federico. Last but not least, we also have a very prominent representative from the investors and AI investors front with Margaris Ventures. Over to Spiros.

Spiros Margaris: [00:05:44]:

Thank you, Mathias. I’m also, just to state here, as on the advisory board to nt. I’ve been in the AI space for quite a bit in terms of an investor, and since 2017 as the thought leader, FinTech much more. It’s a very exciting time because AI has been, as Thorben said, it’s been around forever, especially FinTech and everywhere. But it’s a different AI now and that’s about the things we’re going to talk about today. But thank you very much for having me here.

Mathias Schutz: [00:06:18]:

Alright. Again, a warm welcome to everyone, and pleased to start out the discussion. To start our session, we want to demystify the true headline of our session today. The headline is Agentic Banking.

Let me maybe briefly for people who have not yet been having much exposure with this topic, give an outline of what agentic banking means. It refers to a very new paradigm in banking which is driven by artificial intelligence, and in essence it comprises AI that can act autonomously, make decisions himself and execute very complex and multi-step tasks with minimal human interaction and intervention. That means as a differentiation to the traditional AI, which makes things more in a rule-based automation way, it really is a leapfrog and the next step in doing things, removing the interface, the user interface for it. It can be removing both on the users, on the customer client side, but also on the bank side.

As two examples, on the bank side it could be let’s imagine an AI agent which sits on your e-banking and monitors your accounts, maybe even monitors the accounts over multiple banker relations, and optimizes the savings, moves funds around, and thus everything to maximize your returns. Or maybe another example more on the bank side, where an AI agent would be real time assessing data and updating risk information just to make sure that a bank is regulatory compliant at any point in time. Just two very basic examples to show the dramatic dimension agentic banking, and that it could really remove human interaction to a large extent.

Having said that, I would like to initiate the very first question and to pass it over to Thorben. Looking at agentic banking, do you think we have the right architecture to support such AI agents? Is the world already ready for such a move?

Thorben Croise: [00:08:56]:

If you say architecture from a technical perspective, I would say not yet. We are getting there, but the first step to have an AI agent is to run an AI model somehow in a way that’s compliant and works for the banks, so to say. Rightfully so, banks have data that you don’t want to just send to the ChatGPT interface somewhere in the US and you don’t know where it’s deployed. You need to run the AI in your own infrastructure.

We’re getting there, but right now it’s not exactly like a standard completely accepted way to easily and cost-effectively run an AI in your bank or any kind of organization for that matter. It’s just that banks have higher requirements, so it will take even longer than for any other organization. There is still something, like in the last 6 to 12 months you see things that are happening there, but in the next 12 to 24 months or so at some point we’ll see it’ll be relatively easy to do that. But right now it’s not relatively easy. There’s not a standard way to just run the AI already. From that perspective I would say not yet. It’ll come, but right now it takes a lot of effort to do that.

Mathias Schutz: [00:10:03]:

Yeah, I think that was expected. That was more from now from the technical architecture side, but in such movements there’s also a governance dimension to it. Maybe here Yann, you can explore a bit how you evaluate the situation around the governance readiness for supporting such agentic banking cases.

Yann Kudelski: [00:10:30]:

Yes, thank you. I agree with what Thorben said. Also on the architecture side, I think additionally we also see some challenges with data silos and particularly integrated architecture that hinders as well to get there. But on the governance side, I think we are seeing a bit of a governance gap. That was done quite a bit on the Gen AI governance framework to get started on that end. But on the agentic AI, as you said, especially when it’s about autonomous execution of tasks and decision, that’s the big next step that would happen on the governance side. Because regulators and banks, both of them expect explainability of the decision that’s being taken. Explainability and traceability is key there. But in the AI models it’s not as simple as in a traditional rule-based risk scoring type of decision frameworks, engines to explain it and trace it, especially for humans in the way we’ve looked at it. Overcoming that and make it really explain a bit, I think it’s the key to getting there.

However, I think as Thorben said, with architecture he said 12 to 18 months, I think on the governance side we will get there. But at the beginning what we’re seeing is the more human in the loop approach. You’ll still have the agentic AI but you will have a human oversight at first human in the loop, and gradually you will get more autonomous agents.

Mathias Schutz: [00:12:10]:

Maybe just one additional question to that, out of curiosity, how do you think Yann, can this explainability be improved? What are the key factors to make it work, the explainability?

Yann Kudelski: [00:12:27]:

I think it’s also an evolution of the way models are explained in the past versus how models are explained in the future. I think you cannot overlay the exact same approach framework on a new approach. That would have to evolve and have a bit more also principle-based approach to it. But I think it’ll evolve as well as we continue.

Spiros Margaris: [00:12:56]:

On this I would like to say, people have to realize, even old men, even they don’t know how it works. They know how large language model works, but they can’t explain how the outcome comes out. They have a new approach now chain of thought. Basically they try to evaluate to see how does the large language model come to a certain answer. Although they see those steps, sometimes it comes with different answers.

That’s a risk in our business. If we risk so much reputation, money, et cetera, and you still don’t know exactly how it’s done and if it won’t give another answer than it should, that’s a big risk for reputation risk and money and risk in general. Everyone has to remember, nobody knows. Even the gurus, they don’t know how at the end it works.

Mathias Schutz: [00:13:55]:

That sounds scary.

Spiros Margaris: [00:13:56]:

That’s part of the whole thing. That will hold it a bit. At least for critical things, even autonomous cars, for each five autonomous cars you have a person watching.

Mathias Schutz: [00:14:12]:

Interesting. Maybe this is a good thing to move on to the next question. Spiros, I’d like to hand this over to you. We all know about two and a half years ago that the whole AI wave started to come, even though AI was there previously but with the introduction of ChatGPT the whole thing set. But so a lot of people are talking, a lot of companies see that as a big topic on the agenda. It seems somehow that still also due to the hurdles that you just explained, like observing all those automations, somehow it seems that the implementation is slower than expected. Is there anything you want to add as further roadblocks why it’s getting slow or what would have to be done to be removed?

Spiros Margaris: [00:15:08]:

As I said, those most advanced models, horizontally it is seen everyone uses it. But we haven’t seen the benefits in terms of as company in dollar terms. People have a lot of use cases, JP Morgan over 250 use cases they’re going to try and experiment with. But it’s hard to see the benefits dollar wise to measure it.

Everyone is waiting a little bit to see, what does the other guy do, what does this guy do company wise, and try to copy it. People are working, I think they’re executing with the hand brakes on. They will use it for correcting emails. We’re talking about large language ones, we’re not talking about where AI was used for fraud detection, gen tech, et cetera. I think people are still waiting and trying, experimenting. Companies are not committed yet. Even customers of ours, or everyone here, I think they talk to talk but they don’t do the walking.

Mathias Schutz: [00:16:26]:

Do we have any other views in the group on this topic, on the reason why it’s slow?

Federico Crecchi: [00:16:34]:

On this I completely agree with Spiros. Another things that I can add is that even the banks that are trying to walk the walk, we come back to architecture. It’s like sometimes when they start with a pilot, with the idea, and they can do a prototype in a sandbox, everything looks fine. But then they build the business case, and for the use case to be useful very often it needs to be integrated with the current system. More likely than not, these costs of integration are extremely high, and so everyone becomes much more cautious to go forward with it or not.

When the use cases are more in a greenfield, everything is much quicker. But when you start to touch processes that are more historical, that have a system that is in place maybe 10 years, the cost of integration, at least from our point of view, very often becomes a major if not showstopper, kind of opposed. Because maybe they say, okay, this makes sense but we need to evolve our system, and to evolve our system they go into the backlog of our IT and we would talk again in one year on this. This is another aspect that we are seeing that is slowing down adoption of AI in core processes.

Thorben Croise: [00:18:19]:

Yeah, I think what Federico is saying makes a lot of sense. As fascinating as AI is, it’s not just alone, it’s often not the solution to some of this, especially not in an agentic context. You still need a surprising amount of engineering around that. You need integrations systems potentially. You also don’t need just one model maybe; you need maybe multiple models that control each other. Again you have to engineer that. You need an engineering team that figures out what’s the problem exactly is, what’s the solution here.

It’s different from what we just done before, but just saying let’s use AI for this, it’s not enough. You need to enjoy and engineer the complete solution around that, and this takes time. As you would introduce any other software, it will take a while to figure it out correctly. The project takes a year easily.

Spiros Margaris: [00:19:05]:

You want to be sure if you do it, because it’s time-consuming, resources, money, that as a leadership decision that at the end it will work. Basically they’d rather see somebody else do it and say, okay, let’s copy it; than finding out after a year that there was a nice experience, we learned a lot, but it didn’t get any traction.

Mathias Schutz: [00:19:32]:

Yann, over to you.

Yann Kudelski: [00:19:36]:

Just on what Thorben built, I think it’s an important point, especially if you talk about core processes that touch various items. It’s not that you just have one agentic AI replacing everything. You need to embed it in all the workflows, processes, connect all the systems. There’s a lot of work to bring it together. But the benefit, there is a case, it’s a lot of investment, but we still see there’s a lot of interest in getting started, hashing it out. But I agree with everyone what they said, it’s not a short-term exercise.

Mathias Schutz: [00:20:14]:

The question as we discussed right now, the pace of implementing it was in a rather generic environment I asked. But now let’s look at the banks. Banks are typically by nature risk averse, they don’t like to take risks. Going the agentic banking route means you have to take risk also. You need to accept that there could be errors and so on.

There is already a question I would like to answer, which is very much fitting into this topic around. It was raised to Spiros, but it could be to anybody in this group. Will the industry ever accept an agentic banker? Explainability is almost impossible, it’s the dream of creating an autonomous bank that adapts itself to the market conditions. Is this still possible? Maybe here on this one, the first thoughts?

Spiros Margaris: [00:21:20]:

Look, I am probably the oldest here. When Amazon and all those companies came, people thought, it’s impossible, people are not going to buy over the internet, et cetera, and people are not going to pay with a credit card over the web. But very quickly people got accustomed and it becomes second nature.

I think maybe also agentic banking will be not too long from here. I wouldn’t say in 1, 2, 3 years, but it will happen that people will just think it’s normal. Even people make mistakes, your banker could tell you something wrong. Autonomous cars crash, people crash. Maybe we’ll assume, okay, it makes a mistake, maybe. But there will be fewer mistakes than by humans. I think for younger generation and us later, because everyone uses it, will be okay, why not?

Mathias Schutz: [00:22:25]:

You mean the risks will even out the benefits or will be evened out by the benefits?

Spiros Margaris: [00:22:31]:

You’ll disclaim and say, yes, I don’t mind.

Mathias Schutz: [00:22:34]:

Alright, any other thoughts on this before we move on?

Yann Kudelski: [00:22:44]:

I think I support Spiros. I see the same. It’s not over 1, 2, 3 years, and also define agentic fully autonomous banking or it’s basically supporting first initially some processes and then it continues to grow. But if we talk about, we set on the invite 2030 or 2030 plus, let’s put it that way, then that’s definitely a possibility and on the horizon, yes.

Thorben Croise: [00:23:11]:

I fully agree. With that time horizon, there will be lots of, it is not the obvious path to start with. We’ll probably first see AI a lot in things that don’t touch the customer, but eventually you’ll also see that.

Spiros Margaris: [00:23:25]:

And it’s not going to come from big banks. Remember Google, OpenAI, it’s a small player who has nothing to lose, who will implement it if it’s not perfect. Because bigger you are, the more you have to lose. Remember ChatGPT, the Lambda which it is all based on, was invented by Google but they had more to lose rightfully. But remember it’s not the big banks, big banks will come afterwards, but they will copy it perfectly.

Mathias Schutz: [00:23:59]:

Now looking also a bit on where we stand on the road of implementing AI agents, I would be interested to hear the view from Federico, from Deloitte you’re exposed to many clients with many industry players, what do you observe? What’s the maturity level already there in banking regarding AML onboarding, lending? Can you share some tendencies on where we stand, where the players stand now?

Federico Crecchi: [00:24:33]:

Yeah. I get back to what was discussed before. Given the risk conversion, the precautions and everything, there is a big difference in implementation and solution that look at the employee of the bank or look at the customer of the bank; that look internally or externally. For sure, the solutions that are thought for an internal use are at a much more mature level.

For instance, we see all that is for IT and software development, they are in banking as in other industries probably the frontrunner in adoption in developing agent with a diverse different level of autonomy, for coding development and for the whole software development, life cycle analysis, testing. This is where we see the major banks have already adopted an implemented solutions. Entire banks are working on it and we see a lot of work.

In general, others that are looking internally, for instance the anti-money laundering is not something that should be exposed directly to the client, but in supporting all the analysis that is required by regulation. In terms of KYC, in terms of transaction monitoring, here again we see a lot of investment and already a mature solution.

When we go through on the client direction, we see much more caution in there. We see that there are explorations, maybe they are starting to think of something that could go to the customer, but doing it at first for employees. For instance, a first level support, let’s use it for an internal desk and then if we are okay with it, let’s move to the customer help desk. For sure all the credit life cycle, so lending and stuff, it’s an area of attention. For commercial banking in Europe it’s the core.

But we see there is the attention, there is the willingness to use agentic AI as a way to review the processes, make it faster, reduce the time to yes. But we are at an earlier stage compared to other areas.

Mathias Schutz: [00:27:26]:

Okay, very interesting insight also to see how they’re doing it, starting with less exposed stages and then rolling it out to do more towards the client.

Talking about investments, every AI investment in AI agents and needs, it’s a case. We somehow see that those investments feel to be calculated differently. How can you make the return on investment as tangible as possible in such a project? I think that’s where a lot of players struggle. Also here to the group, maybe starting with Thorben, how DO you see that? Why is it so difficult to measure the tangible results of such an investment?

Thorben Croise: [00:28:31]:

I can maybe say something about the time saving. You can say if you save a lot of time, that you can calculate it to money, it makes sense. In two areas I tried to estimate it.

In the last two months or so, a few times with a lot of software engineers who tried to see away from coding how much time do you save when you use AI? I was in a similar panel a few weeks ago, and the answers varied totally. Then the funny thing is the answer vary totally in the same team. There’s one guy who thinks it’s 5%, the other guy thinks the team saves 30%. A little bit is a bit like when you have something that saves your time, usually it is called Parkinson’s law and it’s very prominent in software engineering. Basically when you estimate something will take two days, then it will take two days. If you’re done after a day, you just increase the quality. People do that, right? You save five minutes for a task, and then you see that other parts of the task suddenly have a slightly higher quality. If you’re not very careful, it’s hard to measure the time that you save.

We tried it also, we work with compliance teams and if we can estimate how much time they saved, it’s very interesting for advantage to sell the software. We tried it a lot and it’s extremely complicated. For years it has been very complicated to see how much time did it really save? It is hard to estimate time savings without AI. I’m not surprised that had to figure out for things where saves like 5 minutes here and 15 minutes there, that it’s very hard to estimate it. I think it’s also a different area. It is not just automation. It automates things that are unusual. Suddenly there’s semantic tasks that were only done by humans, and we have maybe experience of estimating things that are classical automation, but this is a new type of thing you would have to automate. There’s also not a lot of experience of how to do this.

Mathias Schutz: [00:30:17]:

In other words, if you want to come up with a business case, you can make up the numbers as you want. It’ll become true, right? Is that what you’re saying?

Spiros Margaris: [00:30:25]:

But if we look at some FinTechs, I don’t want to mention the name now, but the large one that wanted to go IPO and pushed its AI capabilities, what they were saying very simple that everyone understood is, “We saved marketing costs, 30 million. We will not use as many people. You can calculate. If you had 2000 people, maybe you need 1,500.” It’s very quantifiable, much more than, Thorben, I’m with you, the other way it’s so hard because as you said, they waste time or they expand their project as requested.

But when we look at all those tech giants, et cetera, they fire lot of people, they use less. So far the closest to find out how useful it is for them in dollar terms is the reduction of people and maybe marketing, or whatever. It’s easier to quantify. We don’t need an ad agency. I’m not saying to proclaim that’s the case, but we need fewer people. That’s how they measure the success. Because a lot of great AI behind the scenes, compliance, a lot of things people don’t see, but if we wouldn’t be working they would feel it. But it’s very hard to quantify.

Thorben Croise: [00:31:56]:

Beforehand. Afterwards you can see it, you see now it seems to work and feel. But yeah, beforehand, we make this project here, especially when you say let’s do something small in the background.

Mathias Schutz: [00:32:10]:

Any further thoughts from the group?

Yann Kudelski: [00:32:13]:

I think the challenges come when Thorben said as well, if you do it in small increments or small pieces here and there, then the 5%, 10% saving might disappear somewhere else, as Thorben so eloquently put it.

There’s also a different approach. If you re-engineer something front to back, end to end, and take a new AI first approach, then you’ll have a chance to see bigger benefits upfront than as a one-time, as opposed to benefits sneaking in there. That’s one. The second what we’re seeing as well, yes, it could be reduction, but it could also use to make advisors more productive, effective, have a higher load, do more cases. I think that’s a direction that we see a lot as well.

Spiros Margaris: [00:33:02]:

Federico, a question to you as a consultant, how do you convince your customers to go that path? Although they will say, “I read so many things that nobody still has this killer case,” but how do you convince them regardless of that to spend the money?

Federico Crecchi: [00:33:24]:

It’s the Holy Grail, this question.

Spiros Margaris: [00:33:28]:

Oh, I shouldn’t have asked that question!

Federico Crecchi: [00:33:29]:

No, no. Right now what we see is that it is very difficult to convince a smaller bank that used to be laggard, that they are waiting to see, tell me that this has been done by one of the major banks and replicate what they have done. There are very few cases in which we can do this. On the other hand, it’s going into the major banks and push on those department that have some chief that are less risk averse and with a greater business case.

The point of the software development cycle I was mentioning before, that is also because there are lots of literature that say that this technology brings a lot of efficiency, so you can build an ROI. Also it’s tech for tech. the people are usually more prone to listen and to understand limitations but also the possibility.

The second step is internal operation. Internal operation are those where even a small improvement in large processes for major banks that employ lots of people, can bring an ROI that makes sense. This is the reason why it’s easier to go there. While for instance, as I was saying in credit, everyone is interested, but building the actual ROI is more difficult and so on. It’s still in a phase of, let’s analyze, let’s discuss, and let’s see how it goes in the other areas.

Mathias Schutz: [00:35:36]:

Very good. We are moving to the million-dollar question and the topic of this session today, and that is around, will bankers be obsolete? We’d like to dive a bit deeper on this one here.

The first question, is automation, agentic banking, capable to replace most frontline and bank office banking roles? Maybe a short statement from everyone in the group saying, yes, no, partially. Starting with Spiros.

Spiros Margaris: [00:36:21]:

I think if it develops with the vision we had at the beginning, it will automatically replace a lot of bankers, because it’s in the interest of the shareholder. I’m not saying it’s nice. It’s just going to happen.

Mathias Schutz: [00:36:37]:

If you say a percentage of the workforce, what would you say?

Spiros Margaris: [00:36:45]:

90%.

Mathias Schutz: [00:36:48]:

Large bet. Yann, your view?

Spiros Margaris: [00:36:51]:

Everyone wants it, the ones who want to make money. I don’t say it’s a good thing. It’s not a good thing.

Yann Kudelski: [00:36:58]:

I see it probably a bit more partially. Yes, they will be replaced. But if you look at what roles would be replaced, I see less on the client facing front and more on the back office side. I see also quite a lot of role changes that might happen. But yes, there will be a percentage that can be either replaced or can be used to be more efficient. But I would say it’s south of 90%.

Mathias Schutz: [00:37:35]:

Okay. Thorben?

Thorben Croise: [00:37:36]:

We work with a lot of private banks in Switzerland, and I think for the front office I would not see it so much. They also now claim that a lot of their clients like to talk to humans; I don’t see that so much changing. I think that they will have just more time to talk to the client, make them even happier. For a front office, I don’t know, I can see it totally happening. In more retail banks I can see it’s going much higher. But in a private bank with a front, not so much.

For the back office I also see that many of our clients struggling with hiring enough people, because they have an influx of cases. In that sense, in the short term, maybe not. They could not hire people, but so it will replace people. But for private banks it’s probably not as dramatic as it used to be, because the client contact is important, and the really good client service.

Mathias Schutz: [00:38:25]:

That’s a very good point. Maybe I can also bring one example with Snappi, which is a joint venture of Natech with Piraeus Bank, which is about to go live in Q3. They were clearly focusing as a challenger bank, but also bringing in a clear human support service as a differentiating factor. I think there is also those cases which can be highlighted.

Federico, your view on the question?

Federico Crecchi: [00:38:56]:

I agree with the idea that it’s different, it’ll be very different from the back office and the front office. In the front office. I believe that probably all the transactional services there will be, we already seen before the AI a reduction in the workforce dedicated to that. I believe probably the 90% is a fair estimation.

While I believe that in the advisory, being wealth advisory or credit advisory for smaller and medium enterprises or everything else, for the time being even if the technology would enable an automation, I think that the customer base would continue to want to speak with someone, with a human touch.

On the back office, I want to add the point, I believe that everything that is now called operation will see a major reduction. But on the other hand, I believe that for banking to become an agenting banking, it means to go along the road of becoming a tech company. Becoming a tech company would mean that the software engineering component, the tech component will need to increase in size. It’s still software. Although we will be faster to write software, it’s not the same doing text analysis as it was before LLM. It’s easier. It’s an integration of external components and so on, but it’s still tech.

I believe that in the transformation towards tech companies, the banks will continue to see their share and the size of their tech components increase, also because they will need to hire people with these capabilities and skills.

Spiros Margaris: [00:40:58]:

I would like to add something here, just your thought. We always think of the moment we are at now, but the younger generation, actually there’s a trend, don’t want to see their bankers per se, they don’t want to face people. That’s the one thing that will give a shift. At the end of the day it’s a quality of advice. Think of Ironman, if you have ever seen in one of those Marvel movies. If you believe that agentic AI gives you a better answer, why would I speak to a person? We still reflect at the moment of our generation. But if the quality is better through AI, why would I speak to a human? I will speak outside with humans because I need that touch, but when it comes to money and decisions, I think eventually there will be very few humans we will speak to.

Mathias Schutz: [00:41:57]:

You think the human touch will not play a role, Spiros?

Spiros Margaris: [00:42:00]:

For some people, yes. But for the majority, a very good point, look at cartoons. They’re not humans, but we relate to them. Kids relate to them. They’re not humans.

Mathias Schutz: [00:42:18]:

Interesting point indeed. But as we had this round of answers, we had the very extreme 90%, and the more mixed, I think there will be also something like in between, semi-autonomous banking. Here I’d like to ask one more question to Thorben. How would the human role in a bank change in such a semi-autonomous banking world?

Thorben Croise: [00:42:51]:

I think maybe it’s interesting to see what short or medium term happen. Maybe you can see what happens in the next 20 years or so. We think the tech systems are already slow to be adapted. But governance systems are even, especially if you talk about laws, it will be even harder and even slower to be adapted. In the end, if you take decisions, somebody needs to be responsible. You want probably a human to sign off.

We talk a lot about compliance systems. There’s also in the backend where you check, should we really give this credit to the client? Somebody needs to say yes, we want to do it, and make a decision; or no, we don’t want to do it. I think that for a long time we a problem to give this to an AI. It’s simply a legal problem, and this will take a while. 2030, I think this is probably too short term to do this, to make laws that would even go in this direction. I guess for a long time you probably need humans to make decisions.

On the one hand, and I see it a bit differently, I’m not so old but I still like talking to humans, so I’m not quite sure. I think that the human touch could still play a large role when you go to a bank, and I think the responsibility. Both things that you probably see people shifting to going more. You can take AI also a huge workforce you still have to control. AI is usually not the boss. AI is the intern that knows a lot. You still want to control this a little bit and you sign off what the AI does, and you check it a little bit. Maybe make it much easier to check it, but you will still check what the AI does. Then you go back and talk to your client.

Mathias Schutz: [00:44:24]:

What you’re saying is that agentic banking is an army of interns, right?

Thorben Croise: [00:44:30]:

Kind of, yeah. It’s very smart, but still controlled somehow.

Mathias Schutz: [00:44:34]:

Okay, very good. We have 45 minutes of our call and I have already seen a couple of questions have lined up in the Q&A. I would like to slowly move over to those questions. If you have any further questions to the group, please feel free to add it in the Q&A.

I would like to start first with, just looking at the question, from the bottom one. I leave the second one for the end because I think that is an interesting to want to close on. The last question is, is there such a thing as too much autonomy in the financial services context? Any thoughts on that?

Spiros Margaris: [00:45:26]:

I would love to do a follow up question to the person who posted it.

Mathias Schutz: [00:45:31]:

What is your follow up question?

Spiros Margaris: [00:45:33]:

What does it mean, this phrase? Yann, do you know?

Yann Kudelski: [00:45:39]:

It’s probably related to the autonomous agents, right? I assume at least.

Spiros Margaris: [00:45:46]:

I see, okay.

Yann Kudelski: [00:45:49]:

There it goes back to the explainability, regulation, risk appetite thereof. Right now I think as we all discussed, it’s hard with the explainability. There’s always a human in the loop, and I think we’ll see how we get to, are we okay with a principle-based explainability framework on top of agentic AI? Financial services is a very heavily regulated industry in IT sense. That by itself gives an answer to it. I’m not sure if some fellow panelists want to add anything.

Thorben Croise: [00:46:37]:

Yeah, I guess if you make autonomy completely free. I don’t think you want it. Going back to this example, if you think the AI agent as an intern, you will not give the intern unlimited budget to do anything he or she wants. You will give it good guidelines.

Also talking always when we think about the future as a really well implemented agentic system, it probably has some boundaries of where it cannot go. In that sense, I would say yes, there’s boundaries. You don’t want the agent to make decisions that seem too extreme probably. You probably also don’t want the humans to do that. Why would you do it differently?

Yann Kudelski: [00:47:11]:

Yeah, it would have to be within guidelines and then you have triggers and you have a supervisor almost with a cockpit looking at what’s happening. Yes, going outside of the tight boundaries, probably not a good idea.

Mathias Schutz: [00:47:25]:

The next question goes a bit into the same direction, and it says, can AI agents realistically navigate compliance and regulation or are humans always required in the loop? This very much goes into what you said, Thorben. There will be an army of interns that need to be supervised. There is somebody that needs to control it, that is somebody who needs to have a control on top. I think that’s pretty much answering this question. Or is there any other thoughts on this question?

Thorben Croise: [00:48:00]:

You always think about anyway in level of defenses. You think there’s the AI as the first level of defense. If the AI figures out this is a very high risk, you maybe put a human in the case. You classify everything. I think that you can remove a lot of humans from the loop, but you probably need somebody to maybe sign off in the end. For a high risk case, yeah, it’s worth having a human still look at it.

Mathias Schutz: [00:48:23]:

Very valid point. Good. Then we come to the next question. It says, a recent study from Apple says that current LLM models are really quite stupid, as the moment they are tasked with a problem outside of the scope of analyzed data they are completely dumb. Companies attempt to reduce number of databases and replace them with Cursor, Perplexity, et cetera, but quickly realized that this cannot be done. Question, will AI ever be able to solve problems that have not been solved before by humans?

Here I can bring in another example, which I saw by a comedian in Switzerland who explained that as an affinity for AI. He made a very nice picture in front of an audience. He said, “I was asking the chatbot to prepare picture of a horse on a man.” First he said, “Make me a picture of a man on a horse.” Typically all those pictures that were coming back were the Marlboro pictures, the Western type man on a horse. Then you ask the agent, “Build me a horse on a man,” there was no result returning. It goes a bit into this direction, so dumbness that is in the model. Any thoughts on this one?

Spiros Margaris: [00:49:51]:

I know this study by Apple, and there were a lot of smart people behind it. But as we’ve seen so far, everyone remember the six fingers people had, the first AI pictures? That disappeared. Those problems are real, but there are companies that try to solve them. It’s good that we have such issues, but they all will be solved, then we will have other problems. It will become very smart.

Artificial general intelligence is defined in different ways, where it’s things like humans, the expectation is it’s not going to take that long. Even the godfather of AI, they didn’t think large language model would have these capabilities now. When we see a problem, don’t think it’s going to be there forever. Because they’re always small startups that think, wow, here we can differentiate ourselves if we solve it.

For us, every great problem, I believe it’s just something good that we’ve noticed it. There will be companies that will address it and we will benefit from it. That’s my view.

Mathias Schutz: [00:51:05]:

Any other views in the audience?

Federico Crecchi: [00:51:06]:

I have to say I agree, because I believe that we are so within this that we’re not realizing that since ChatGPT has been released, it’s two and a half years ago. The quality of the models have improved incredibly. The newest GPT is doing things that seemed impossible to be doing, in terms of the length of the context you can give to these models, the ability and multimodality, the easiness in which now we can build agents and integrate with a web search or whatever to increase the capabilities. It’s happening so fast. If we look back just one year, we are in a different world. I believe that we need to see some perspective.

To bring a bit from another industry on something that’s not been solved, it’s not a new problem, but if we go into the pharma industry, there are papers and stuff about new molecules identified and invented by generative AI. That is fueling drug discovery. I’m a very optimistic person on what this technology can bring us in the coming years. I believe that it would be a mistake to think that the current shortfalls that exist will be shortfalls forever.

Mathias Schutz: [00:52:53]:

To summarize both of your answers is that the boundaries will be continuously expanded. There will be always new capabilities coming up. It’s just a matter of time and the limitation will not be there forever.

Spiros Margaris: [00:53:10]:

The thing, as Federico and everyone mentioned before in the whole conversation, is nobody doubts that everything will be possible. The fear that comes is, do we have the control, keep the control? I would never worry that it’s not going to be possible. The fear that all those people have is, can we control what we produce by the Pandora box? But on my note, everything will be possible. Almost everything.

Mathias Schutz: [00:53:47]:

Very interesting, the statement here, Spiros. Looking at the time we have four or five minutes left. I would like to move on to the last question that was raised in the audience, actually one of the first questions. I would like to do it in a way that everyone is asked to answer it. It says, Spiros mentioned that JP Morgan has identified over 250 use cases to test AI. What does this group see as the most promising or viable use cases for AI in banking today?

To end this panel here will, it’d be very interesting to see, where does every one of you see the most promising use case? Who wants to start first?

Spiros Margaris: [00:54:44]:

We’re thinking very hard.

Mathias Schutz: [00:54:46]:

It’s an unfair question. Otherwise I’ll roll the dice instead.

Thorben Croise: [00:54:55]:

I can definitely tell you, when we talk a lot to our clients to see what we can do, one of the more promising things is in compliance. There’s a lot of things about compliance, about policies, about text, about things you have to understand semantically. They are very hard to automate and do anything about it. You can do certain things with rules. I think that it’ll also have hybrid systems because then they can be more explainable. But I think you can get a lot of things to understand the content that you have produced.

For example, when I talk to compliance when we do an onboarding, they tell me, human understands that if you say you live in Switzerland where you have an Albanian phone number, you should ask why is that happening? This is something a large language model will just do out the box to check such radical basic things. I think that’s to just go through a package of an onboarding or a KYC review or any process you want to go through, and check, is there anything wrong? Do you see inconsistencies, so to say, inside the data? I think this will reduce a lot of checks that seem crazy futuristic, and it is relatively simple with an LLM to do.

Anything in compliance, like understanding policies, doing something with this and analyzing client data with the policies in mind, this is an amazing use case. You can do so many things with that. It’s relatively simple. When I explained this, now you probably all have an idea how to do this, right?

Mathias Schutz: [00:56:15]:

Cool. Thorben is putting his bets on compliance.

Thorben Croise: [00:56:21]:

Yes.

Mathias Schutz: [00:56:22]:

Who is placing the next bet? Maybe Yann?

Yann Kudelski: [00:56:24]:

Yeah, I can be a little bit on Thorben, but I also like what Federico said about the credit side. I don’t place it on bets, but on actual client work that we’re doing with clients, and we’re spending a lot of time in the mortgage origination process. There it starts. You have a lot of documents to look and review. You can extract a lot of information more in unstructured documents, get the information out, validate it, have all those cross checks that Thorben just mentioned. I think that’s we’re working on. You then have a bit more rule-based decisioning framework behind it. But once you have an ETP, an exception to policy, you can have an AI review it, prepare a file still for human intervention, but prepare it already with some reasoning behind it and making sense out of it. You can then use it for credit monitoring, to identify patterns going forward.

We’re building a smart AI agent to help the advisor throughout the mortgage process too, because for people who don’t do it on a daily basis, they might have questions themselves. We have a smart agent supporting them, helping them, guiding them through the process. But the key there, it’s all advisor-facing, in a sense to enable and make the advisor more productive and speed up things.

Mathias Schutz: [00:57:53]:

We’ve heard compliance, loan origination, advisor. I would leave Spiros for the last one because you’re the most extreme. No, Spiros, you’re the last one to answer because you’re the most aggressive with 90% to be replaced. I’d like to hear first, Federico, what’s your view of the most promising use case?

Federico Crecchi: [00:58:14]:

I cannot say what has already been said, but I agree. I will get back on the use cases related to the change in processes and IT. Everything that is related to defining business requirements, transforming business requirement into a mockup or a proposition and stuff, and then developing these solutions. I believe that we will see that these processes will shrink in time in a major way. Because at every step of the process there are some ways in which generative AI and agentic AI can help.

Mathias Schutz: [00:59:08]:

Okay, so it’s a lot about optimization process improvement. Spiro, the stage is yours. You said 90% are of the banks are replaced. You need to be very clear with your use case.

Spiros Margaris: [00:59:24]:

No, but I always say in this case that hope I’m wrong. I’m a FinTech thought leader forever. Top guy. I’m more holistic. I think AI in general, all what the gentleman mentioned will play a major role. But at the end of the day, it’s providing better services through AI to our customers, cheaper services, and drive democratization of financial services.

Rich or poor will have indistinguishable services. The only difference is the bank account; but it’ll make no difference if you’re very wealthy or very poor. That’s I think what AI will enable us. I think that’s a nice high note. I truly believe this, the organization of financial services. I always tell in my speeches, financial services, all these things will be like air – you will breathe it but you won’t think about it. It’s just there. It’ll tell you, you can buy this, you’ll buy this, you can afford this. Things will be done for you. This only be done through AI.

What we see now, in 10 years we’re going to look back and say, “My God, did we really think this was cutting edge?”

Mathias Schutz: [01:00:47]:

Very nice. Nice picture. With that, I would like to close the session today. I would like to very much thank Thorben, Federico, Yann, and Spiros for participating. It’s been great fun. For all the guests who have been here and have been asking a lot of questions, thanks a lot for your participation. Thanks a lot for the Natech team and marketing for setting this session up. Very much appreciated.

I wish you a wonderful evening. Thanks again, and goodbye. Bye-bye.

Thorben Croise: [01:01:16]:

Bye-bye.

Yann Kudelski: [01:01:16]:

Bye-bye. Thank you.