From Slow to Smart: Itaú's Epic Bank Statement Transformation!

Hello everyone. Good afternoon. Good afternoon. My name is Eduardo Rodriguez and I'm a principal solutions architect at AWS. I'm super excited to be here today and talk more about this modernization journey on how Itau modernized their bank statements and how they are creating new experiences for their customers such as the intelligent search. Joining me today, I have Alison. >> Hello everyone. I'm Alison Alison Carneiro, an IT executive at Itau and today I have the pleasure to be here share with you our journey with a modernization banking in statements. >> Okay. So, in terms of agenda, we're going to start with the context. Alison is going to share some numbers from Itau with you and the challenges that they were facing before the modernization. Then, I'm going to dive deep into the new architecture and I'm going to explain the main flows and also the technical decisions that we took when we created the new platform. After that, Alison is going to share the modernization and outcomes so far and he is going to tell you how they are improving the quality of the service for their customers. And finally, I'm going to show you the intelligent search and this is how Itau is transforming their customer experience. And we're going to close with some lessons learned and the next steps. So, Alison, could you please share with us the context before the modernization? >> Sure. Before we start, let me present who is

Itau. We are one of the the largest private bank in Latin America with 100 years of history. We have you over 93,000 employees and 17,000 only technology. I'll just employ this to support over 17 million customers. We are presently in 18 countries and our market value is over 93 billion dollars. So, after understanding who after understanding who we are, let me explain why we had to change. New technologies have changed our customer habits. Today, they demand speed, they demand speed, personalization, and seamless personalization, and seamless experience. experience. To respond to this, we need we had to recognize our service end-to-end, always putting the client at the center. We focus on four key pillars. four key pillars. design, technology, design, technology, products, and data. products, and data. This combination allow us to deliver a better experience with you more quality, speed, and efficiency. That's is already happened with Itau intelligence, where we use AI to improve and personalize interactions. This is the base of our transformation. customer centric and AI power.

So, let's look at the challenge we were facing facing in the past. When we started to look at the experience, we identify some key challenge. One of them was latency. Our architecture was legacy and monolithic. And it did it is carry very well. As demand increased, performance is started to degrade and response time became higher, which affected the overall experience in our app. We also faced some limitations in data access. Our customers could only see around 3 months of historical data. Yes, 3 months of the data in the online channel. We also faced some limitations in business. Our customers couldn't do a long period of search data and couldn't do a complex filter. When we compare individual and corporate account, the gap became very high. For corporate account, we didn't offer data aggregation to reconcile the transactions or balances. And to And to to focus on this challenge, we group in in three main areas. First, customer centric. Second, platform, and the last one is time to market, which means increasing agility and experience for our teams. To talk about the platform, I would like to invite Eduardo to deep dive into. >> Thank you, Edison. So, now let's talk about the evolution of the bank statement on AWS.

But, before I dive into the architecture, I'd like to share with you the vision that we created when we designed the new platform. So, the vision was clear. We needed to create a new unified platform to serve all the channels and to have both individual and corporate accounts in the same place. >> [clears throat] >> [clears throat] >> And for that we defined five technical drivers. The first one was resilience. So this is a mission-critical service, right? So you cannot be unavailable for any reason. So we designed a new cell-based architecture to support this to support this high availability requirement. high availability requirement. And then we have the scalability. As you already know, they have over 70 million customers daily and they are growing daily on a daily basis. And for that we need to have a platform that can scale as needed. And we also need to have the costs under control. So this is kind of a mental model that we use it when we are we were designing this new architecture. And we cannot modernize things and expanding even more when you create a new things. So the cost optimization is always in mind, right? And because this is a highly regulated industry, the security, governance, and compliance needs to be in place. So you need to have the least privilege access in place. You need to have the fine-grained access in place. You need to have zero trust between between the services in place. And this is job zero at any financial services customer, right? And all of that to create a strong foundational platform to have the new features such as AI-powered ones and be capable to create new experiences for the customers. Taking a look about the high-level

architecture of the bank statement. Here we have some some some main layers that we created. So the first one is the explosion and the channels layer. The In this layer, the channels can call the APIs from the platform and to request [clears throat] the bank statements, right? And then we have the processing layer. At the processing layer, is what we do all the orchestration between the services and for instance, we uh identify the customer, we do the complex filters that we need, we do the aggregations that we have and so on. And then we have the most important layer in this architecture, which is the data layer. As you can see, we are not using a relational database. We are not even using a no SQL database. As you can see, we are using Amazon OpenSearch service as the data layer. And Amazon OpenSearch allows us to have the complex uh features and the complex search patterns that we need to do uh this in this platform. >> [clears throat] >> So, uh Amazon OpenSearch is the case that we are using the right tool for the right job in here. And because of the um distributed nature of the service, this allows us to scale as we need. Uh as I'm going to show you, we are applying some advanced resilience patterns in this architecture and I'm going to explain it later. So, we also have the ingestion layer. And because we decouple this platform from the financial ledger, we have the ingestion layer to have the transactions being ingested. So, this is the ingestion workflow that we have.

First, we have many data sources and the transactions are happening all the time and in the online transactions, we are using streaming to capture to capture all the transactions and then we do we do uh uh the processing on these transactions. And here is important because we're not only consuming these transactions, we are also enriching them, creating um better descriptions for the transactions, categorizing them. And also, remember that I said that we needed to have both uh individual and corporate accounts in the same platform. Here, we split the clusters from OpenSearch into the types of accounts. This is the first thing. And then, we also have the reconciliation batch ingestion. They are using a big data pipeline backed by AWS Glue to ingest all the batch from the transactions that are reconciliation transactions during the day at the new platform. Right. So, we also have the search workflow. In the search workflow, we start with

the channels channels calling calling uh uh the APIs from the platform. And then, we have [clears throat] here in the processing layer, I'd like to highlight two advanced resilience key techniques that we implemented in this um platform. The first one is the shuffle sharding. So, we are splitting the data from the customer into shards. But not only that, we are also doing shuffle sharding. How many of you already uh heard about the shuffle sharding resilience better before? Okay. I think no one, but uh we are doing uh the we are not only splitting the data between shards, but we are replicating the data from the same customer for more than one shard. And this allow us to have a reduced scope of impact in case anything goes wrong because I as a customer doesn't share I don't do not share the same shards as Edison do. So let's say that I I live at the shard one and three and Edison in shard two and four for instance. And that case if any for any reason one of the shards goes down another one will be available to respond to the request. So this is how we are providing the high availability in the platform. And this allow us to have another technique. And another technique technique in the resilience pattern that we are implementing here is called it hedging. The hedging means that we are doing multiple requests to this to the shards that we have. Let's say that I have two shards and I will do two requests and I pick the first one that comes available. And for any reason one of the shards are not available to respond to the request I have another one to get the data that I needed. And this not only also not only help with the high availability but also with the response time. Because we pick the first response that comes available and we do the filters aggregations and we respond the bank statement to the customers. Cool, huh? Cool, huh? Uh Uh anyone anyone seeing here that we have also Amazon Bedrock in this architecture. Right? So I'm going to explain the role of the Amazon Bedrock in this architecture later. But for now I'd like to invite Edison to share with you the modernization outcomes so far. >> So thank you Eduardo. Now the best part

of the presentation is outcomes results. Let's see before and after modernization some numbers. To start, do you remember our main three pillars? First one, customer-centric. Second one is platform that mentioned by Eduardo. And the third one is agility. Let's first talk about the agility. Our feature lead time before the modernization, we were deliver about 90 days to deliver lead time. And now, after modernization, we are able to do only in 30 days. Only 30 days. When we talk about the

data access, before I said we had some only 3 months of historical data. Now, we can provide over 5 years in the online channel. 5 years to do a complex filter, search a long periods of the data, and then do aggregation for our corporate accounts. To demonstrate the engagement, we can see here before, we are able to do only 2 million requests per second monthly. And now, after modernization, we are able to do over 8 million requests. It's demonstrating gain of the search for our customers. Let's see a concrete example of the modernization before and after. Before the modernization, uh around 40% our feedback satisfaction was related about unclear description transactions. It because our architect was legacy and had some limitations, fixed position any number of characters. After modernization, After modernization, as you can see, we prioritize the key information to demonstrate the descriptions of transactions. The first line we choose action, for example, send money to, for example. The second line is about for who send money to John Smith, for example. And the last line is about amount dollar. Send money to 1 $1,000. With these improvements, we can see an increased over 5% of satisfaction. And we can see a decreased over 5% too of our detractors. But the real impact we can see in a corporate account.

Do you remember I said before data aggregation? Before, in general, all companies do the aggregation or reconcile or reconcile the transaction and balances with Swift file. But they need to wait until the day and the end of the day to reconcile their transactions. Now, with modernization, we can provide one different way. The API bank statements. With this API, they can do the reconciliation transaction and balances in near real time. With this, we can see here increased over 5% of satisfaction. We can see a share of funds over 4% and this is so important because that API the customers bring their activities from other banks to Itau. And we increased over 30% of product banking. Which means we can offer make offer a a product that Nerd needs. Okay. So, the second pillar is about

platform. As you can see, October we were in both platform legacy and modern modern. In 2023 November, we made the decision to move our clothes to the cloud. And November why is a special month? You know in November we have the Black Friday. And I don't know why, but in Black Friday after someone pays something online, I don't know why, it's a behavior, they going to check their balances or bank statements. And because that we have a huge increase of your requests. In our legacy architecture, as I said before, the grade all experience in response time. And before we our response time around 2 seconds. So, huge response time. And now with modernization, you can see here only 200 milliseconds. Over all 94 billion requests in 2025. When we talk about the MIPS consumption, let's see. let's see. In 2022, In 2022, we had around we had around 20,000 MIPS on the mainframe. MIPS means millions instructions per second. Could be very expensive. As I said before, modernization is a journey. In 2025, we had only 3,000 mips, which means is a reduction of consumption mips over 9%. When you talk about the our third pillar, pillar, we talk about the agility. Before the modernization, we were able to do around 200 changes, technical changes on the productions. And now, after the modernization, we can see we can do over 300 changes on the productions. And it is means increased over 53% faster than before. >> [snorts] >> [snorts] >> When When we talk about business feature, we can see an increase over 45%. So, next I will show for you the real big numbers after 3 years in our journey. Let's talk on Let's talk on storage. storage. In just 3 years, we have over 216 petabytes of stored data in just 3 years. In Brazil, we have the strong regulation that we must provide over 10 years of stored data after an account an account closing. Could you imagine multiply 10 years of storage by 206 petabytes? Could be expensive too. But, with Amazon OpenSearch combined to with S3, we use the life cycle policy. We can split in different tiers each type of the data, for example, hot, ultra warm, and cold tiers to control the costs. And to demo- to demonstrate engage, we can see over 80 million search rates on the cloud, and to isolate each type of customers, we can we have over third clusters to support our requests. So, what more we can do with modernization? Could imagine it? AI intelligent or AI AI AI. So, now let's talk about ditch Eduardo

and you deep dive in AI intelligent. >> Thank you, Alison. Wow, amazing numbers, right? So, um to talk about the future, we're going to share uh how it always creating a new experience for their clients. Uh Uh and we called it the intelligent search. So, to have like an example here, let's say that I want to know how many Pixes I did last month. By the way, Pix is our uh instant payment system in Brazil. Or let's say that I want to know if I I don't know paid my gym membership last last month. Or let's say if I want to know if my salary was deposited last month. I just type in the search bar at the bank statement, and we have here a natural language query processor that will route this request to a NLP layer that will understand the query in um the natural language, will uh identify the context, will identify the intent, the type of the transactions that the user are trying to search, and using Amazon Bedrock, we're going to understand all of the context and translate that into a new optimized OpenSearch query. And then we are going to process and run against the data layer to filter and to aggregate the results to present to user. Pretty simple. Um this is a simple feature as they shouldn't do here. They they shouldn't be. Right. Uh in terms of uh technical architecture it's simple as well. We have let's say that they have the channel calling uh the the platform APIs and doing a semantic search. So, we do that and we have the routing layer uh routing the request to the NLP processing layer and then we have a many uh layers of treatment and identifying the context and uh what the the user want to see and what they really want to uh filter in terms of bank statements and transactions. And then we go to um we invoke the LLM uh through the Amazon Bedrock to translate that into a open search query. And then we run against the data layer. And as you can see, we are not creating nothing especially here. We are just understanding the context and we are just understanding in in in the natural language what the user wants to see. And then we filter these results and we deliver to the customer. Just by doing that, it's already uh has uh 80% less errors in the channel when the the user is trying to search for a for a any um bank statement or filtering transactions. So, this is a super good result that they have as far. So, moving forward moving forward they started with the traditional search, right? And by doing the modernization, they already achieved um much more complex and much more uh deep filtering capabilities using open search. And now they are doing the intelligent search using the natural language processing to translate the natural language query to a uh database query.

database query. Right? In the future, they are building the conversational bank. So, as Edison mentioned before, the habits of the customers have changed, right? So, Itaú knows that and they are doing that and they are building on top of Itaú intelligence to put this kind of experience for their customers. So, this is This is right now, they are building right now, right? In the future, the the natural dialogue and device commands will be the new normal for them. for them. So, So, to close this session, I'd like to share with you some lessons learned and the next steps.

First of them, uh modernization is a journey. Edison already said that, right? This is not one-time event. You need to keep evolving the platform every day. You need to optimize performance. You need to improve the scalability. You need to strengthen the security. You need to create new features for the business and so on. and so on. And all of that, uh you have to have uh a strong foundation on architecture in place, as we did in this platform. We created a cell-based architecture and we applied and implemented uh design patterns for resilience as shuffle sharding and hedging. And by doing that, we created a strong foundational platform for big statements. And that's why we can create simple and innovative features as the intelligent search. Right? And Right? And to in lastly to in lastly I have we have the cross cross-functional governance. This was crucial for the success of this project. And to have the alignment between the stakeholders from business to architecture team and the engineer engineering team, this was crucial to have everyone at the same page for the success of this modernization. Right? As the next steps, they are rolling out the intelligent search for all the customers. And they are building the new experience by doing conversational conversation dialogue in the bank statements. And they will be help ready to move out from main family from from mainframe completely by the next year. Guys, that's it. Uh I'd like to thank you all for your time. I hope that that was helpful for you. And if you have any questions, Alessandro and me are going to stay here. Please let us know. We're going to be happy to talk about it. Thank you so much. >> That's