GenAI And The Future of Financial Services

Amit

Hello everyone, to yet another episode of Build AI by GI Ventures. Today, we have Andreas, who is the Head of Industry Principal Architects for Financial Services at Google Cloud for APAC. I'm looking forward to this discussion given that Andreas comes from a big tech company, and so far we have interviewed a bunch of founders. So he brings a new perspective, a different perspective, and I'm looking forward to hearing that.

Meet Andreas Neidhart-Lau: His role at Google Cloud

Amit

I think it will be a good start to hear from you: what is this role exactly, and what do you do? And also, how did you get pulled into this intersection of what seems like cloud, AI, and financial services?

Andreas

Hey Amit, first and foremost, thank you so much for having me on your podcast. I'm very excited to be here today. So, as you said, myself Andreas, I lead the Industry Principal Architects for Financial Services in Google Cloud Asia Pacific. I'm based out of Singapore. I've been in Singapore for more than 10 years now and have worked most of my career in the financial services industry.

What does a Principal Industry Architect do?

In Google Cloud, the Industry Principal Architects is a team of very seasoned, senior industry experts for financial services—both banking and insurance—that bring a wealth of experience on the technical side as well. So I have people in my team that have been Chief Digital Officers in banks implementing large-scale data platform transformations, as well as industry architects from the banking side themselves.

What we do in Google Cloud is we work with our customers to give them an understanding of how the capabilities that Google Cloud offers—whether it's infrastructure services, product services, or most recently with our advancement in AI in terms of large language models and those types of products—can support their KPIs and objectives. We help them map those capabilities to the digital transformation agenda that the banks and insurers in the region have.

Google Cloud's tech engagement with FSI clients

Amit

Right. Very interesting. And I'm assuming that this organization within Google Cloud is there in other parts of the world as well, apart from APAC, right? So is this a symmetrical organization in the US and Europe as well?

Andreas

We like to say Asia is always a bit specific, but yeah, there is obviously a strong alignment to the global teams. So this type of organization exists in Europe and North Americas as well. There is a global community within Google Cloud that continuously exchanges their experiences and success stories so that we all can learn from each other.

Amit

This seems like a very interesting role. Your background—you've worked in financial services and consulting firms. Was there a specific reason or a specific moment how you got into this particular role at the intersection of AI, cloud, and financial services?

Andreas’ transition from energy to banking tech

Andreas

I think there was no specific moment; it was more of an evolution of the work that I've done in the past 20 to 25 years. I've always worked in a deep technical role, mostly on the enterprise solution side. Whether it was at the start of my career in the semiconductor business or later on in the oil and gas trading industry, it was always about how to implement those enterprise solutions in the context of a customer environment and ensure that the solution is deployed in a highly scalable and performant way. Back then, that obviously happened largely on huge server farms.

I've done implementations in Europe where we implemented trading and risk management solutions for the largest electricity producer in the northern part of Europe. As part of their day-to-day operational processes, the solution had to complete certain calculations and end-of-day processes. These kinds of calculations could take days to complete. Implementing a solution that connects hundreds of servers together in a grid-like computing setup to help accelerate that number crunching was part of the role I was doing.

As I subsequently moved to Asia and transitioned more into financial services, the problem statement obviously remained the same. Financial services institutions, banks, and insurers are number-crunching a huge amount of data every day to come up with predictions, reports, and other information. Those outcomes need to be calculated and generated in a way so that the end-user who is consuming that outcome can receive the information in the most efficient and shortest possible time.

Rise of cloud-native banking and career milestones

When we did these types of implementations, this new technology emerged which was cloud. Quite interesting at first, but within the highly regulated environment of banking and insurance, it was more of a topic that was known to the CIOs and CTOs of the banks but not actively pursued because the services were not mature enough. There was still a lot of uncertainty from regulators across the region. Policies and compliance teams didn't know how to approach this type of technology. But things have obviously, as we all know, changed over the years.

As I progressed my career, the solutions that have been developed have been developed in a more cloud-native kind of fashion. One of my career highlights in the past 25 years was working for a UK startup that developed cloud-native core banking technology, which was very progressive and new at the time. Trying to bring that technology into Asia Pacific and convince a lot of very traditional banks to make use of it, and also load their operating and processing of data from their on-premise environments into a public cloud environment, was at times very challenging but also quite interesting for me to learn.

I would say that the move into Google Cloud was then just a natural evolution because I've been working on this type of technology for many, many years. Now, sitting directly at the source, having access to all those innovations firsthand that are being released almost daily in the pipeline of Google and Google Cloud, is actually quite amazing to be a part of.

Generative AI in BFSI: effectiveness vs. efficiency

Amit

That sounds like a fascinating journey. I think you've seen this industry now from various angles. That's a good segue into my next question. I think the billion or trillion-dollar question these days is: what are the most common use cases of generative AI in financial services? I know you have a vantage point in giving that answer. So please tell us, what are you seeing in terms of the most common use cases?

Andreas

That's obviously a great question, and it's something I'm faced with almost on a daily basis. Generative AI, but also predictive AI, has been around in banking and insurance for quite a while now. We've spent an enormous amount of time and effort with our customers and prospects to identify use cases that can be applied to this technology.

If I try to summarize it or bring it down to maybe two keywords on how you can identify those use cases, I would actually call them use cases related to effectiveness and use cases related to efficiency.

I would say in the last 18 months or so, most financial institutions have experimented with use cases that improve the efficiency of their internal processes and operations.

Document processing, code assist, and reporting use cases

So, we've seen a lot of use cases specifically around how you document processes, process documents, and summarize the content of those documents in a meaningful way so that unstructured data becomes structured and can be used for further processing.

We obviously see most recently, with the advancements in large language models and the capability of providing code-assist capabilities, use cases around code generation and development. So again, improving employee efficiency, specifically the developers within banks, to help them write new code faster and more efficiently, is something that is quite useful.

Another use case specifically in that category is around report generation. How do you extract the vast amount of data available in the financial services universe—all these industry reports, company filings, pricing information from exchanges—and consolidate this information to create a report so that a human can digest it in a fast and meaningful way? That is obviously one of the key use cases in this category.

Amit

That's a very interesting framework. You said basically effectiveness and efficiency. I think from what we have experienced at GI Ventures, the companies that we are building, the first wave seems to be a whole lot of efficiency-related stuff and productivity-related stuff. What would you put in the effectiveness category?

Hyperpersonalization & external-facing AI deployments

Andreas

The effectiveness category is obviously any solution or use case that generates new revenues for the bank. That comes with a new set of complexities. The internal employee efficiency use cases were all very contained in terms of addressing only internal employees, and the number of internal customers it's being exposed to is a controlled environment. Now, opening up these capabilities to the customers of the bank—external customers—raises a whole new set of concerns. Is it secure? Is it compliant with regulations? What about hallucinations and other issues that banks have experienced with early versions of this new technology? Overcoming this kind of mental hurdle in a lot of banks is quite important to move into this new category of use cases.

The use cases that we see here is first and foremost, I would say, hyperpersonalization. Hyperpersonalization has obviously been around for many years, and there have been a lot of technologies in the market that use machine learning models to collect customer data, mostly transactional data, and then predict certain behavior based on pattern analysis. This created a set of personalization options but didn't go into that level of hyperpersonalization that we think is now possible with large language models.

Large language models have the ability to process any type of data, whether it's structured data in the form of text input, documents, or unstructured data like audio files, information gathered from social media, or video information. Consolidating all these different types of information and then using predictive AI and generative AI to interact with the customer and recommend financial services or things to do within the bank that the user might not know at the time the information is being received—I think that is the next big thing that has to be implemented in banking.

The better you understand your customer, the better the services are being offered, and the higher the success rate in terms of conversion, activation, or cross-sell will be. That improves the effectiveness of the service being offered.

Wealth management, trade finance, and vertical AI use cases

Amit

That's a great way to put it. Actually, there's no use case discussion in financial services GenAI which is complete without talking about the segments and the subsegments. Could you throw some light on—we're talking here about lending, accounting, wealth management, and so on. From that particular angle, where do you see the most common use cases today?

Andreas

I think there are specifically vertical use cases and then there are horizontal use cases. The horizontal use case that we can see is specifically around intelligent virtual chatbots that can be used to interact with your customers. Regardless of the business function you operate in—retail banking, corporate banking, private banking, wealth management—there is a need to offer 24/7 service to your customers that is consistent and personalized so that it speaks the language of the customer and provides information based on the underlying data available from that customer.

If we go into more of the vertical use cases, they are very specialized to the individual line of business. A few examples: In wealth management, some sort of financial planning and advice agent that can offer personalized investment advice on your wealth platform, continuously monitor and analyze your stock portfolio, and recommend you change the allocation based on information from media and sentiment analysis.

We also see specific use cases to address certain issues in corporate banking. Banking today is still an industry which is very manual and very paper-based, specifically corporate banking. If we think about trade finance or corporate lending, there is a plethora of documents that need to be filled out in different qualities and formats. The time to process this information requires a lot of resources from the banks in terms of human resources and time. There are specific use cases on how you extract the information via OCR, entity tagging, and then comparing it to certain internal databases to reduce the time for these processes significantly.

DBS Case Study: $750M revenue via AI

Amit

That's super interesting. I'm going to ask you something more specific. I know it may be difficult from a Google perspective, but there may be some stuff that you have put out in the public domain, especially around a client story or a case study. I wanted to know if you can walk us through a client case study. What we are trying to understand is if there's a case where generative AI radically changed the equation, and even you were surprised by the outcome.

Andreas

Yeah, I think one of the customers that is using AI very successfully is a large Singaporean bank, DBS. DBS has publicly announced to be using Google Cloud technology for a large part of their AI transformation projects. When we started working with DBS, it was quite interesting because one thing they applied from the very early stages of their advancements into AI is to implement a very clear framework on how they are going to measure success on the outcome of the use cases. These measurements include KPIs of how much revenue can be generated, how much efficiency can be implemented, and other things.

Just recently, I think in April this year, DBS announced that in 2024, with all these different AI initiatives that they have launched using the platform, they generated more than $750 million in revenue. That is quite a staggering number. We are still living in a world where a lot of management consulting companies, banks, and CIOs are running around on the assumption that their competitors are implementing GenAI use cases primarily to experiment and run POCs. But seeing DBS, as the leading digital bank in Asia Pacific, create such a strong outcome is, I think, market leading. A lot of banks are looking into those numbers and wondering what DBS is doing differently and how they can follow the lead to generate a similar outcome.

Amit

We will try to put up some of this case study, since it's available in the public domain, for the audience later on. It's very interesting. DBS was one of our customers back in my previous startup, and I love the level of detail they go to benchmark best-in-class products and how they work towards it. So it's no surprise they are early into this as well.

From pilots to production: how banks are scaling AI

Amit

What is very interesting is in some of the other geographies, we have also seen some of the largest companies, which were usually slow to adopt new technology, today going ahead and doing generative AI pilots and even implementations way ahead of some of the small to mid-size players. So something very different is happening with GenAI.

Since we are talking about bigger companies, this whole thing around pilots and POCs—we see them across the world. But I think there's this big question: when do we move from pilots and POCs to mass scale adoption? Or maybe you are already seeing it. What is your read of the market currently?

Andreas

I think the market has kind of transitioned now from experimenting with AI to implementing AI. There is a clear trend towards a top-down approach in many banks that we are talking to, where the top leadership is basically mandating a mind-shift change within their organizations to implement AI at every possible business process within the bank. These are really transformational changes that are currently happening, which obviously raises a lot of concerns and questions on how to implement and execute such a change.

One thing that we've also noticed now is that while banks have experimented on use cases, in 2024, with the advancement of agents, there is finally a technology in place that allows them to combine these single use cases into an end-to-end user journey and process to potentially augment or complement existing workflows within the bank.

Rise of multi-agent systems & end-to-end AI workflows

I think this is starting to become more prevalent now. This technology is allowing banks to implement AI in a large-scale motion. Before that, you had specific single use cases like report generation and chatbots, but the technology didn't allow you to have this combination of different tasks and keep the memory of those tasks so that you could process it through different steps of a workflow. With the capability that is now available, we see a lot of banks rushing to identify processes within the bank where this technology can be implemented as a multi-agent type of implementation.

Amit

It's great to hear and very reassuring that things are going from pilots and POCs to production. But then, of course, there are many financial institutions who may still have various types of issues. It could be people mindset to change, policy issues, compliance. How is this different from what used to happen during the fintech days? Do you see a lot of this still happening in terms of roadblocks to implementation of GenAI use cases?

Compliance, governance & ethical guardrails in BFSI AI

Andreas

I think nothing is changing because fintech has always operated in a highly regulated environment. It is about customer data and customer information, and it needs to be highly protected and secured so there can be no leakages or incidents. When we talk about these roadblocks and issues, the first question always is security and compliance. How do you implement this new technology in a way that follows regulatory requirements and the compliance rules and policies the banks have set to protect their customer information?

Banks are approaching this in different ways, but we have seen the most successful customers of Google Cloud implement a platform approach. A platform approach where they follow an AI governance platform that decouples the use case implementation from the underlying infrastructure and technology solution. How do we achieve that? By creating a set of capabilities where we templatize the services that Google Cloud offers, for example, access to Gemini and the prompts. These templates can be audited, verified by security, and approved by compliance. So whenever there is a new use case, the banks don't have to go through the same cycle of security approval, questionnaires, and compliance officers. This dramatically accelerates the launch of new use cases in production and helps mitigate the concerns and risk that come with implementing new technologies.

The second concern is a lot around the uncertainty of whether the output the AI model is generating actually makes sense. Transparency and explainability in AI is important. You need to create a robust framework with guardrails and security to continuously validate and measure the output and ensure it aligns with the risk framework within the banks.

Last but not least, within that context, we have ethical issues. Addressing potential biases in the training data and making sure there is no discriminatory outcome, specifically for information produced for external customers, is important. Putting guardrails and quality controls in place around this is key.

Legacy infrastructure & data silos: the real bottleneck

If we then think about the second part, it's also something not to neglect: many banks in Asia Pacific, but also globally, are built on technology that was launched many, many years ago—decades ago. Banks have never completely rebuilt the architecture; they have always built on top of what already exists to implement new features and functionalities. A lot of banks, as a result, are living in a technology nightmare where there is a spaghetti architecture with millions of different types of solutions, different levels of technological maturity, and fragmented and siloed data buckets across the enterprise.

AI is a very data-hungry application. It requires a lot of information, and accessing that information by integrating into all of these legacy systems can become quite cumbersome and difficult to achieve.

Amit

This actually reminds me—I was at a bank many years ago, and they showed me this big printout they had with all the various applications they had implemented over the years. I think it was somewhere like 150-plus things. Every five years they would have a few more, bolted on to a previous platform or a new platform, and it is messy. I can relate to what you're saying.

Andreas

I have a fun story to that as well. I was once working with a bank in Thailand, and we were talking about digital transformation of some of their backend systems. These backend systems were decades old. The technical teams there asked us for recommendations on how they could sniff the network because they couldn't find out anymore all the integrations that go into that system. They were looking for all the source and destination systems going in and out of this application, and there was no network map, no architecture map available. The people that implemented the system had long gone into retirement or moved on to other banks.

Google Cloud’s AI model stack: Vertex AI, Gemini, model garden

Amit

I'm assuming from a Google Cloud perspective, what is the approach? Are you bundling or offering your foundational models along with the cloud?

Andreas

We offer our models in different types of bundles. We have our Flash model, we have Pro model, and different other models. As a customer of Google Cloud, we will work with you to identify what is the most appropriate model for your use case. We have our platform technology called Vertex AI that has access to all of the large language models that Google offers, but also access to a model garden of more than 200 additional models. It also offers direct integration into Hugging Face, for example, where you can download open-source models and run them on Google Cloud technology.

For all the models that Google Cloud offers, there is documentation available, like a model map, that gives you information on what the underlying training data has been and the use cases it is most applicable for, so you as an end-user can make informed decisions.

Why data is the new heart of financial services

Amit

I'm changing tracks now, going more to financial services specific. Most of the AI stack today feels off-the-shelf, but I have spent 12-13 years in fintech, and what we know about financial services is it's deeply verticalized with very specific workflows. One of the companies we are building, Fastracker AI, focuses on workflows that are so nuanced. The onboarding process is so different from lending or a banking account, and the same with processes around meeting notes and CRMs. What's the hard part about building infrastructure and solutions that adapt across wealth, credit scoring, retail fintech, and across geographies? How should Google think about it, and how should startups think about it, given the need is for very specialized, verticalized applications for specific workflows, but you also want to build it in a way that it can expand later on?

Andreas

That's a crucial question. What is the common denominator for AI use cases? I mentioned it before: AI, generative AI as well as predictive AI, is very data-hungry. It can only produce meaningful output if the data feeding into those models is meaningful at the same time. The data verticalization and the contextual nuances of that is the most critical part.

When I talk to banks, I like to say that for many years, banks have followed the approach that at the heart of the bank is the core banking application that processes all the information and provides the output to omni-channel solutions. But with the advancement of AI, data has become the new gold. If we think about the heart of the bank today, we have to think of it as the data lake that consolidates all the information available for the customer and builds that holistic customer 360 profile.

The data verticalization, consolidation, and provisioning becomes very crucial. Implementing strategies to avoid silos of data, but create domain-specific data products within the bank so they can be consumed easily, is key. Data in wealth management is completely different from data used in credit scoring or trade finance. Having that productization of data is important so whoever consumes it can make the most meaningful use of it.

Also, we are living in a world that is highly digitized. The emergence of digital banks is proof of that. This leads to a great variety of unstructured data that can be consumed, such as news, video, and voice. More recently, in the context of hyperpersonalization, it's about how you collect behavioral data of your customers. How do they interact with the mobile device? How fast do they click? Are they uploading certain documents? Do they revert to previous steps of the workflow? Consuming that information and providing it to the processing platforms is quite important.

Amit

Super interesting. I sold our company to Prove (which acquired UnifyID), a leader in behavioral biometrics, and spent a bunch of time with that team when we were trying to launch it in Asia. What you said about data is so important. A lot of entrepreneurs building GenAI companies don't realize that. I had a discussion with a pretty senior guy at Clara, and they said it took them more than a year to figure out the data lake situation: how to get data streamed to one place so that real-time enterprise searches could happen. People don't realize that the actual applications of AI didn't take much time; it's the data lake and how you manage the data where you should spend a whole lot of time.

Andreas

I think a lot of fintech applications as well have that issue of creating copies of the data. In conversations with banks, if we propose a new solution, one of the first questions is always, "Do we need to create another copy of our data and put it into a new database? How do we ensure consistency and replication?" That creates a lot of concerns because it moves into the issue that has always existed in banks: operating data in a lot of silos that do not interact with each other. While the fintech applications might not consider that their problem, they are one of the creators of these issues because banks are forced to create all these copies of data and continue to keep those silos in place.

Model strategy: banks using multiple foundation models

Amit

I'm going to switch gears quickly. When you're talking to banks and financial institutions, are they choosing one foundational model, like Google's generative foundational models, or are they trying to use all? What's the situation right now?

Andreas

I think there has been a lot of experimentation which has led to a lot of banks implementing an approach where they select models for specific use cases. Traditionally, Anthropic has been very strong on code-type use cases, whereas Llama has been very strong on certain other use cases. Most recently, when DeepSeek came out, a lot of customers were interested in exploring it for use cases that don't require a lot of guarantee, because the processing was very cheap.

That has been the situation since the beginning of the year. What we are now seeing is models become more mature. At the moment, Gemini is leading the pack with a lot of industry benchmarks. But one KPI that is becoming more important to customers is: how much intelligence can you actually generate out of an API call by using one prompt? If you look at the latest benchmarks, you would clearly see that Gemini here is leading as well. It generates the most intelligence out of a prompt, which is quite important for banks to realize the KPIs and targets they have applied to their use cases.

Are banks building or buying AI/GenAI applications?

Amit

Very interesting. Related question: when it comes to generative AI applications, have you seen any bank which is actually trying to do this in-house as opposed to buying from outside?

Andreas

It's a quite interesting conversation. I just recently had it within the office. The question was if any banks are basically training their own model. The reality is, so far, most of the banks I've spoken to would prefer to use out-of-the-box, pre-trained models and then use something like RAG to inject the data they require for inference. The reason is that for the last 18 months, training a model has been just too expensive. If you try to create your own foundational model and train it, as a bank you probably don't have access to the vast amount of data required to train the model. Secondly, training and running it on infrastructure has been too expensive, so banks have not really been exploring that option to a large extent.

Amit

And what about the application layer? Let's say a bank needs a document processing engine. Have you seen any bank or FI trying to build this in-house?

Andreas

Yes. The most prominent example is that a lot of banks always have these conversations around build versus buy for their core banking applications. Specifically banks that operate in very complex environments with fragmented sets of markets and different regulatory requirements would like to have a certain level of control over their applications and services. So they have followed the route of build versus buy, and we recently had a number of these conversations with our customers.

But the problem I see is that with the advancement of cloud and now AI, the pace of innovation is so significant, and the amount of effort, resources, cost, and time involved to create a new solution—does that really benefit the bank's overall objective of creating more revenue, reducing operational cost, and increasing time to market? Based on my experience, most of the time, a measured approach of buying an off-the-shelf solution that offers a platform approach with flexibility to configure and build on top is usually more beneficial than building everything from scratch. There is a reason why fintechs exist: they spend hundreds of millions of dollars and hundreds of thousands of developer hours to build the solutions you deploy. For the majority of customers we talk to, it's not something I would advise to build in-house.

The evolution of Agentic AI in banking

Amit

Just one last question from my side. This discussion would not be complete if we don't discuss the whole Agentic AI evolution. I want to know what the opportunities are, what the bottlenecks are today. We are talking about multi-agent systems. What have you seen? What are you seeing in the evolution of agentic AI?

Andreas

Agentic AI is obviously the next big thing. Finally, we have something available that can be implemented into the existing business processes and workflows of the bank. What I would be very cautious about, however, is what I'm seeing with many customers at the moment. They are trying to do a "rip and replace" of the existing business process with new technology, but that is not going to realize the benefit and value you can generate out of such groundbreaking technology.

When it comes to agentic AI, you obviously need to completely rethink the entire business process and reimagine how agentic AI could remodel it to become something that is more intuitive, more customer-centric, and more hyperpersonalized, rather than just having an agent that processes data from A to B with the ability to connect data sources in a more efficient way.

We are at the beginning of that evolution. Working in a regulated industry, we also have to accept that banks typically take these kinds of changes very slowly. Implementing this on a small scale first to see how the technology works in the long run, and then moving into a larger-scale business transformation, is something that will happen over the next 5 years.

Conclusion

Will AI disrupt traditional financial paradigms?

Amit

Maybe I'll ask a final, final bonus question. From 2005 till now, there have been so many waves. We spoke about core banking solutions, centralized credit bureaus, traditional audit firms. When fintech came, when blockchain came, there was a discussion: are we going to disrupt these paradigms and have newer forms of credit bureaus, a new core banking system, or newer audit firms? With GenAI, are we finally there? Is something going to happen to some of these things?

Andreas

Certainly, there are things that are going to happen, but I don't think it will be as dramatic as we think at the moment. I think we have to think about AI and GenAI as a technology that will augment our capabilities to take away all the low-level work of collecting, preparing, and processing information, to the point where we can then finally make the decision on the information that has been prepared. AI will greatly help with all these activities so that you, as a credit underwriter, as an officer in a credit bureau, or as a lawyer in an audit firm, have information at hand that is better prepared and more accurate, so you can make the decision faster and more informed.

Amit

Great. This was super fun. Thank you so much, Andreas. I also know that you were unwell over the weekend, so thanks for agreeing to do this while you were recovering. Thank you for your time and all the insights.

Andreas

Thank you so much, Amit. I really appreciated the questions. It was really fun. Looking forward to the next one.