Inside Banking AI: Not If, But How.

Robert

Nice to see you, Mike. Thanks for having me today.

Robert

Athens to Atlanta, I think, right?

Robert

Yes, definitely. I think that was like one year ago or how many?

Robert

Oh my god, Mike, feeling old.

Robert

Sure. So yeah, working with banks, trying to provide technology for them that’s actually needed, right? The brunt of my work was customer facing. So I’ve been meeting banks in the past two years since we spoke. Like, maybe I wouldn’t give you a scale, but a lot of banks from a lot of different regions. So the learning experience from that was like extremely telling for us regarding how we should build our product, what should we strive for, and what are the actual customer needs, right? Because it’s very easy to you know, go into our R&D department and think about how should we solution, what should the platform look like, but nothing teaches you what two hundred conversations with customers teaches you, right? So it’s a very humbling experience. I’ve been traveling a lot, all of the Europe, USA and other other areas, and meeting with leaders and seeing what are the bottlenecks they are facing, what is their adoption cycle in technology. If they are stuck, if they are experimenting, if they’re mature, some of them, right? Right.

Robert

Yeah.

Robert

Yes, it was actually quite surprising. If you were to ask me, like, two years ago, what would be the adoption in banking regarding AI? I would say, like, banks are like, you know, the most like, how can I say, type of establishments that don’t like to adopt new things that are not very, very tested. Right. So I was seeing a slow adoption curve. I was, I was preparing for it and just like preparing for some extremely hard conversation regarding reliability, repeatability, deterministic outcomes and stuff. But we are seeing some, I would say, a lot of encouraging signals from banks that are trying to innovate and trying to adopt. I think some of them are market-driven, some of them are ROI-driven discussions. But I’m seeing a lot of openness, which was for me quite unexpected, right? I have a big background in banking, so I know how banks usually model processes and how they put them into production. So the level of scrutiny for any kind of modernization is like extremely high. So when AI was in the discussion, I was pleasantly surprised to see interest in the discussion.

Robert

I think they are embracing it for very, I would say, very clear reasons. Let’s call it that. So if you think about the banking process, let’s say whatever they have, policies, processes for mortgages, onboarding, all of the operations they are handling, right? So there was this huge era, what I would call like the classical automation era, right? Which is algorithmical development, one plus one is two, right? Which works right every time. If you do mathematics, one plus one is two every time. But you can push that into your banking practices to a pretty high extent of automation, right? Like you have step one, you have step two, you have step three, and then you move on. But there are limitations to this type of work, right? So those limitations would be whatever you have points in the process. Where you have to have more like reasoning capabilities, right? So if you’re looking for fraud, for example, investigation in a process, if you’re looking across 20 documents and trying to establish correlations between them, right?

Robert

Yes.

Robert

Yes, exactly. Because that specific parts of the processes were not dealt with with classical automation and they were dealt by humans. And humans, I love humans, but humans have a nature of being subjective in their work. And if you give them broad instructions of work, they will have different interpretation of how the work should be done in a specific point. So you would have pretty high inconsistencies of quality speed from one person to another, right? So not, of course, the biggest bottleneck is the operational overhead, right? So what a person could do in a few hours maybe with agents can now be done in a few minutes, right? So those were the parts. There were still black boxes, or maybe not black boxes, but still missing from the classical automation cycle.

Robert

Yes.

Robert

It’s funny that the picture you are saying is like, I still think about adoption when you’re saying that because banks have some sort of LLM capability deployed internally, so they have a chat interface. And in some departments, they are allowed to use it in some small parts of the work. But the work that you are actually doing in that situation is I have an application, let’s call it the loan origination system or core system. I do some work here, and then I go to the chat interface and I copy and paste and do something in what we call like an individual assistant way. Which doesn’t provide any kind of framework to my process that I’m working in. So it’s like grabbing context from here, putting it here. And that’s how you get a soup, right? So that’s why you don’t have any kind of guardrails or any kind of audit and traceability to the process that you are working on.

Robert

Exactly.

Robert

Sure. So what we found, and this is a trial and error thing that we did, is that you could go two separate ways maybe. So, the classical automation side, step by step, deterministic steps. And then there’s full agentic, right? Which is non-deterministic part of the work. The best combination that we found in banking is to harmonize the two approaches. So, you will still have some steps in the process which are deterministic and should be deterministic. This is a finger rule that we use in engineering. If it can be done deterministically, it probably should be done like that.

Robert

Yes.

Robert

Let’s pick mortgage, for example, which is so the usual suspects are the processes who are taking the most amount of time, and the reasons they are taking the most amount of time is because humans are still working on them, right? So, if you’re looking to optimize the process, you’re just what’s taking longest in a bank. Mortgage is a prime example because, especially in Europe and also in the US. You have government schemes, you have regulation, you have a lot of documents. So, average dossier size in mortgages in Europe is more than 200 pages per application. So, those documents are usually property documents, income validation documents. They could be a lot of information, right? So, to give you the parallel that you are asking for. The normal process would be: somebody goes into a branch, they have a conversation, they have the initial, I would say, financial qualifications. That is done in most part deterministically with some caveats. If you have documents for income, somebody has to check them. And you can imagine a bank statement that I don’t, I’m not sure how your bank statement looks like, but if you were to take my bank statements for one month, you will have like 30 pages, then you would see like a lot of lines there. Okay, how do you extract your eligible income from that? And then compare it to a payslip, for example, right? So all of that work right now is done usually by a person who just like looks at the paper and then copies information and puts it into a system and makes some checks. So just on the income side, you could spend like 30 minutes to validate that, right? That is part that was missing at the beginning.

Robert

So everything is happening in a specific pipeline. We have a logical order of how we manage those documents. The first thing that we do is called classification, right? So we have a process, let’s say it’s mortgage origination or type of servicing, right? And on that process, we are expecting some documents to appear at some point in time. And you know how those documents can come in. Sometimes it’s just one PDF that has like 200 pages, and that PDF contains 10 documents, right? So classification of documents is the first thing we are doing. We give us an artifact. And we try to see what documents are available in that artifact.

Robert

Very good. So, from the mandatory fields that usually they have to extract, it’s 100%. So, basically, they don’t have to open up the documents to copy information into their systems.

Robert

Everything. I call it like a it’s very simple term like a bucket. You just throw documents at it and he understands what documents are inside. If they are appropriate, if something else needs to happen, if they are expired. So basically, everything regarding correlation from the artifact to the raw data that we need to extract.

Robert

I think there is one thing that’s extremely important for this. And this is where we and the team have spent a lot of time in making sure that we get this right. Because, as you say, the risks are like pretty big if you don’t get stuff right. So, first of all, we do have human in the loops, right? So, we have what we call confidence scoring, confidence scoring regarding regarding the job that was operated. So, if for some reason the confidence score dips below a certain threshold that we set in testing with the customer, that’s what it will trigger a human review for the specifics of it. But the most important thing is this, Mike. I’ve been seeing a lot of OCR engines, all of them actually. And once you are doing high-level extraction of information, I think the only good enough solution is the one that has repeatable results. So, and I talk this with everybody that we have our customers: the test for us is that if I upload the same artifact twice, I should have the same results. Yes. Because if you’re talking about let’s make it better and how we can calibrate it better, if you cannot repro it’s like in bugs, right? If you cannot reproduce a bug, it’s very hard to fix it. It’s the same here. You have to have repeatable outcomes on the processing if you want to have calibration that’s meaningful in the process itself.

Robert

Yeah. And we only cover the first part, like which is the, I would say, the ingestion of data. The second part is even more cool. Oh, okay.

Robert

Well, this is a very, how can I say, easy conversation. There are multiple benefits. The first one, which everybody is interested in on ROI, is the operational efficiency, right? So how much are we impacting the workforce with our automation on the extraction of information? And that reduces the time from, I would say, on average of three hours work on documents per application to something close to five minutes.

Robert

Yes, yes. This is the average the average application processing time per documents extraction and interpretation.

Robert

That is just operational efficiency, and we should talk about also qualitative

Robert

Right? Yes, but also the manner in which the execution of that is happening gives the risk guys, right, and the compliance guys a lot more controls about how it’s how it’s actually being done so in those unstructured points in the in the process you could write something like a new policy and tell the guys how they should interpret it and how they should do it but if you’re removing the human reasoning part right and and you’re putting it by an agent you can actually test calibrate, audit, control, every kind of small calibration that you want to have on the process. So it gives you, speaking about your questions regarding regulations and stuff, we actually have very productive discussions with the regulators because our systems make the process safer than it used to be before.

Robert

Yes.

Robert

Yes, most of it is processed, right? And each bank has its own policy that they have to follow. There’s lots of regulation in place that needs to be happening, right? So what we are doing with agents is that we are not trying to change the policy because changing the policy is a very lengthy discussion. So again, I’m always going to adoption. Like where can we actually do some impact with the customers that we are having? And basically, we are not changing the steps too much. We are just automating them. So my goal is to automate ninety five percent of the work being done in each step. Changing the steps would be a different conversation, right? Just because it has a lot of novelty regarding risk.

Robert

Well, you’re buying a house, Mike. So if you’re buying a house, maybe I would even argue it should not take five minutes to buy a house, right?

Robert

So there are multiple reasons, right? So the bank has to assess the collateral, for example, because it’s going to be on their balance sheet, right? So how do you assess the collateral? There are multiple ways of doing it, but some of them include the person going to the house and checking your pipes. To see if everything is working, checking the marketplace. A lot of this can be automated with agents, right? A lot of it maybe still requires some human intervention, some human oversight for reliability reasons. What is being solved though is the financial analysis part. That is being like done mostly in real time. With most of the banks, the collateral part still requires a bit of like interpretation on on the field expertise and so on.

Robert

Oh, so I think having real-time approval on your financials is something the market was looking for for a long time. And we are seeing more and more adoption towards this. It’s like you can know exactly how your budget should look like for a new house acquisition in a few minutes, right? And you don’t have to go to the bank most of it for it. So then you can go shopping for a house and you come back to the bank and/or to the mobile app once you found the house. So I think that’s like a big step up that that’s being happening.

Robert

It is.

Robert

It’s moving, maybe just to add, Mike, like, what I’ve been seeing that’s helping customers a lot, and I’m talking also about the end customer, the person who’s actually buying home, is that if you have embedded expertise with the agents, so you give me a document, you give me information, and my agents are being able to react in real time to that data, I don’t have to have this long loop of information goes to a back office, to a lease department, to a legal department, and then it gets back to the customers.

Robert

Exactly.

Robert

Yes.

Robert

Yes. The way we talk with banks usually is like, having a Warren Buffett in your front office who looks at everything and gives you the best information and advice. So you don’t have to wait for those crazy loops to happen between departments and handoffs. It has real-time intelligence at the point of what they call data ingestion. So you give me information, I tell you if I need something else or if I need some additional stuff. And it’s good for the customer because he’s, you know, real-time feedback. You don’t have to wait for two weeks and watch at the ceiling.

Robert

What’s the big deal?

Robert

Oh my god. So. I’m not sure if all this is the scariest. I think like the level of digitization is the scariest. So I’ve been meeting with banks who still use Excel spreadsheets for making mortgage applications.

Robert

And they are like not like tier three. I wouldn’t see like some top tier banks who do this. And the operational risk that comes with that is like so. When you’re talking about gains in the level of scrutiny that you can have the process on, you can imagine moving from something like Excel, which doesn’t have any kind of audit.

Robert

Yes, you don’t. The reality is you don’t. Actually this is what we have been seeing in the last two years, is that the best path to adoption is to deploy it on top of the existing systems. Those systems are there for a reason, right? Nobody’s crazy. They don’t invest time and energy into those systems. They might want to replace them, but decommissioning something is extremely hard conversation. So deploying on existing systems is the way to go. And also, I would say the market to that can benefit from agents, I would say the repair replace market is like this. The deploying on top of legacy system is basically the whole market because the agentic era is still at the greenfield from an adoption perspective. So basically, all of the banks I’m talking to can benefit from deploying on top of the existing legacy systems. And I don’t see any issues with it. I think that’s actually the way for adoption. You should deploy on your existing stack. Use it, learn from it, and after some time you can decommission it at a later time.

Robert

Well, basically the question is regarding integration usually. So how do we integrate with those systems? And there’s also a question regarding interfaces, right? So there are multiple ways that we we achieve this and we we work with our customers, for example, in some situation. If they have an existing LOS or JSON system, we can basically embed our agents into their existing UI and recreate the experience that they are having in the container application. So, from an end-user perspective, you wouldn’t even know that you are on FlowX.AI, for example.

Robert

Yeah, so usually there’s a discussion. We are going to be complementing the existing solutions most of the time. That means some level of understanding regarding the underlying systems, like what is the API structure, for example, right? So we have a lot of experience of integrating with COBOL systems like you were mentioning earlier. We have done it all. So if it’s any kind of API native application, that’s the most seamless integration that we can have. But we don’t have that benefit, right?

Robert

Yeah. I think the best architectures that I saw, especially in the most modern days, is to have the MCP structure enabled. So basically some agent to agent protocol enabled, right? That would make our job when we are talking with those services a lot easier. But this is still, I would say, in infancy dream. Mike I think in two years’ time they will be the norm. Mostly, mostly we are still seeing all kind of batching processing in some instances, SFTPs, right overnight processes. And that’s fine. We can work with that. That might become a bottleneck as regarding performance, but it’s not like the end all be all.

Robert

Yes, I think so strategically most of the banks are having this conversation right now. And there are two paths that they are exploring, going AI native first, which does imply a lot of rip and replace, like you were saying, or how can we have a high impact ROI solution that is good today? I would advise the second one. I think spending too much time on architecting something that’s AI native and spending one year and then putting your chisel and start to work on it, I think the rate of novelty in the AI space, you will benefit a lot more by deploying on top of your legacy systems and just start to work with it, start to have results. Start to see how you can configure it, how you can calibrate it. And the experience you will gain from that will give you a way better picture on how to build the AI native stack. So that’s the adoption that we are seeing mostly. It’s easier because there’s so much low-hanging fruit, Mike. Like on mortgages, for example, like you can deploy that in two months, right? So it’s like so much low-hanging fruit that should be, you know, eaten up right now. And then you can talk about maybe AI native architecture from the ground zero and stuff. I think there’s a balance to be had there, but I think the ROI is there for deploying on top of legacy systems. So great.

Robert

Yes. I think there are a few things that I’m observing. And Mike, what I’m telling you right now might be very different from what I’m going to be telling you in six months. So the again, the technology novelty factor that’s happening and the cycles that we are seeing from a model to a model or from a platform version to another one, it’s like three months and things are evolving at a scale. But what I’m seeing right now, and again, this is what surprising for me at first is a thirst for more agentic parts in the process. So as I was just telling a bit earlier, banks still want to have very deterministic steps in it. And I think they should be, right? Because if it’s working right now and it can be done in real time, you should not introduce more volatility or more risk to the process, right? But I’m seeing a lot more appetite for more agentic interpretation. And I was a bit surprised by this. I was saying, like, why? But there seems to be a natural comparison between what the output work looks like from a human right now versus what the output from an agent might look like. And even if that work cannot be deterministic or not be 100%, but it’s 99%. The human work is sometimes 70%. And banks are making this calculation and saying, yes, we should automate this with agentic. Yes, maybe it’s not going to be 100% bulletproof on this process, but I don’t need to be. I need to be like better than what we are doing right now. Which, if I were to talk about this, Mike, with you, like four years ago, this would be very I wouldn’t expect this conversation to be had today. But they are seeing an opportunity right now, a lot more mandates to do this now. It’s also a first-mover approach. So in some countries, some markets, everybody was a bit skeptic until the first one did it, and they have inputs into pricing because of that.

Robert

Not just open. I think right now everybody’s running through them. My discussions with some CIOs is like, I have to do this because I cannot justify my operational costs. Some competition is doing it, so I need to find the most efficient way to do it. So it’s not a question of if they want to do it. Who to do it with? How should it scale? What would be the strategy for it. So that’s been, how can I say, refreshing to see, like a point of reality that we are sharing with them.

Robert

Mike, this was awesome. Let’s do it again.