🦄 Making great presentations more accessible.
This project enhances multilingual accessibility and discoverability while preserving the original content. Detailed transcriptions and keyframes capture the nuances and technical insights that convey the full value of each session.
Note: A comprehensive list of re:Invent 2025 transcribed articles is available in this Spreadsheet!
Overview
📖 AWS re:Invent 2025 - Reinventing credit origination process through AI (IND372)
In this video, Adrien, Global Director of AI Solutions at Cloudera, demonstrates how agentic AI modernizes credit risk applications for business banking. He explains that banking faces a $340 billion opportunity with generative AI, particularly in SME lending where manual processes dominate. Using Cloudera's private AI platform with partners like Neo4j, Dockling, and Galileo, he showcases a live demo of an automated system that creates AI agents from credit policies to evaluate loan applications against compliance rules, calculating metrics like loan-to-value ratios and down payment requirements. The solution is already deployed at a bank for audit compliance checking in credit origination processes.
; This article is entirely auto-generated while preserving the original presentation content as much as possible. Please note that there may be typos or inaccuracies.
Main Part
The $340 Billion Opportunity: Why Banking Needs Private AI and Cloudera's Hybrid Platform
Hey everyone, first of all, very happy to be here to talk about how we help our customers modernize credit risk applications and make the best out of solving concrete problems using agentic AI in this case. I will talk a lot about finance, why this is happening, and why we need AI agents in order to solve some of the credit risk challenges that banks are facing, particularly for business banking.
First, who am I? I'm Adrien. I'm a Global Director of AI Solutions for Cloudera. So what do I do? I help our customers around Cloudera AI to make use of the platform and to solve the right use cases with the right solutions for their organizations.
First, banking. Banking is the most impacted industry by generative AI. That's quite clear. I think according to McKinsey, we are looking at $340 billion in potential value creation thanks to generative AI. But it's not that we are looking at $340 billion that is going to happen overnight, right? There are still a lot of challenges that we see on this.
The first one is integration into existing processes. If you look at banks, there's a lot of legacy, a lot of systems that have been here for years, and you need to make sure that these systems can get connected to an AI system. When you look at training the people, entire processes were revolving around manual processes, human driven, and you have to rethink them on how you move this process to become AI driven with human control, which is a very different taxonomy. It's not just like sprinkling some AI on top of a process and hoping something would happen.
And you look at a common topic for banking: safety and guardrails. How do you ensure that the process is still safe and that you are not sending your credit data to the wrong person or the wrong actors, right?
But first, let's get back to Cloudera. So Cloudera, we are an AI and data platform. We claim that we can do everything from your data ingestion all the way to your AI on one platform with a single control plane and a single security layer. We are completely hybrid. Obviously our preferred partner is AWS. That's where most of our customers are using us when they are on cloud, but we can also do it on premise. So we try to really be a bridge between the two for customers who might want to do a zero code refactoring move to the cloud. They can obviously do it.
We provide a unified data fabric. What does it mean? It means that for all your applications, from the ingestion all the way to your AI, you are managed with a single security schema. You don't need to reproduce some security rule somewhere else and then realize that your access policies do not match, and then you get 25 audit findings and your bonus declines at the end. I was Head of AI for a bank, and I know that when you get too many audit findings, the bonus goes down. There was a rule like this. I discovered it the hard way at some point.
We provide the Open Data Lake House obviously as a foundation to AI applications, but also to reporting and any kind of more traditional things you want to do with your AI and with your lake house. But today we'll focus a lot more on the application layer. Why? Because obviously we have a big shift towards AI that banks and financial institutions, the largest particularly in the US, have been investing heavily on data lake houses, and now it's money time. You know, make this data useful and be able to build applications, refactor applications, and get some business impact leveraging AI.
We do this even more on the advantage of Cloudera, which is private AI. Private AI obviously is very convenient for us in the financial services industry, meaning we are able to provide this within your own VPC, whether it's on cloud or whether it's on premise. Everything remains within a single environment, so you can deploy models, whether it's an LLM or a credit risk model, in the same cloud estate if you want, which is really something that our clients appreciate.
We do value a lot of open source, so most of our stack is open source. Why? No vendor lock-in. I always say if you don't like us, you can take the code and go on open source if you want, which is, I think, a testimony to the value we generate. Clients are willing to pay for us not because they are locked in, but because we do provide them a lot of value. And the last part is hybrid, like I said: on cloud, on premise, you can use the estate as much as you want and deploy the code wherever you would like to have it.
Transforming SME Credit Origination: From Imperfect Information to AI-Powered Policy Compliance
Okay, let's get back to our applications. When I talk to customers about AI, they tell us that they see a lot of potential value. The potential value is 50% improvement in the process, particularly for business banking and commercial banking. Why is that? Because if you look at the landscape of credit origination for mortgages in most countries, it's heavily automated. Even in the US, you have a credit bureau, you get a score, and you have almost, not perfect but close to perfect, information. So automation is already there.
The real pain point for banks, and when I asked some of them how much time they spend more, they tell us SMEs is where they spend a lot more time. That's why the spreads for SMEs are very high, because operational costs are high and it's also difficult for them to get good insight on what's happening. Why? Because when you are an SME, sometimes you don't get audited. The audit report is very different from one company to another. They work in very small businesses, so you really have a tremendous opportunity there.
But in an ideal world, we don't need this. In a nice world, we have perfect information, perfect trust, and perfect decisions. What does it mean? It means that we have great data coming from great P&L statements that are uniform and everything. That's what you see on large corporations. That's not what you see on smaller enterprises. You have perfect trust when a company gives you a document and you know this document is trustworthy. Obviously that's not true, especially when you start looking at international lending and other less well-known enterprises. And you have perfect decisions, given the information that you have. You can make a decision where you know with certainty what is the risk you are taking on this loan, meaning your credit risk model should be working fine. That's obviously not the case.
We do have a lot more friction that is remaining when you look at a loan origination process. You go from an application, verification, underwriting at some point, and then maybe some approval if it's good. Here, what do you see as key blockers? You see quite a lot. Documents, many of them are going to be physical. Especially when it's an audited statement, you still find them with a stamp, and even if it's a DocuSign, it does not help much more. You have policies inside the bank, so the policy against which you're supposed to match the credit application. This policy, what is it? It was written by human for human. It's not written by an AI, it's not readily enforceable, and then later on, you'll see component monitoring and others.
So you have obviously a heavy human intervention in this particular process, whether it's application, collecting the document, then saying, "Hey, you're missing this file," making a decision. Yes, there are machine learning models that are making decisions, but if you look at the machine learning models for enterprise and small and medium enterprises, the human input and the expert inputs, as they call it, is still quite heavy. So that's why we do see that there's still quite a lot of opportunities, and the opportunities are numerous.
First of all, AI-powered origination. Can you make the document review more seamless? Can you review a credit memo and state against this policy, "This credit memo seems too much," or, "Hey, there are some deviations. Here are the deviations. Please go reach an approval for a certain by a certain authority." It could be a chief risk officer, it could be just a director, compliance, anyone. And the last one is obviously AI-powered loan management. We do see some banks now doing automated call collection. I was working at a bank before in Asia, and instead of having people calling for loan collection, they put a chatbot to it. And if the chatbot managed to reach out to the person, that is then passed to a loan processor. So maybe, and in the US, I know some of you might be receiving some calls that are automated from the banks in order to try to get you to pay back the loans.
And then later on, we can also have AI-powered loan management, risk review. And risk review is something that is very simple yet very complicated for enterprises. It's any new information that may lead to a risk review. Think about it like this. Tomorrow a board member of this enterprise is changing. What are the implications from a risk standpoint for this particular credit? Might be nothing, but if you have to do it manually, you have to research, "Oh, this person is on the board. What is the name screening for this person?"
What is the new risk scoring that we see, and so on and so forth, right? So it might not be straightforward. For today we're going to focus a lot more on, okay, I have a policy, I have a credit memo. How do we know that this policy and this credit memo are matching? And we do have some customers really actively investing a lot on this. Why? Because that's a very large bottleneck on it.
So we have a credit risk policy written by human for human. It might be a mortgage policy, it might be an enterprise policy. We have a risk memo, and basically the credit policy is going to state, okay, your loan to value must be 80%. Does this credit proposal fulfill this, or do we need an escalation process? That's a process that can take really a lot of time and leads to loans being postponed and not being approved in time. So we're going to have a demo where we create an agent per rules in the policy, and then this agent can be matched against the credit memo that can be seen as a credit application from an enterprise.
Before that, you know, the process that we have chosen for this is that we will have a creation for each agent by reading the policy, and an AI system can understand what are the right agents to extract and what are the right agents to create so that you get a list of, in this example, having a couple of hundred agents. Then we can select it. The selection in the example is still manual, can be automated. We do it manually because, you know, still some human intervention. You don't want to apply the wrong rule to the wrong memo, then you get obviously a wrong output. Then the execution and the monitoring, it's really going to look like this.
Live Demo: Building Agentic AI for Credit Risk Assessment with Cloudera, Neo4j, and Galileo
We're trying to bring together a bank process, a bank policy, and also the people in the bank as like human in the loop layer, and even in the future try to bring in the machine learning model. That's just the technical slide that I have to put. How did we build it? So I built it leveraging a couple of other technologies, some open source, some of our partners. Everything runs on Cloudera. Neo4j has a graph database in order to bring better, you know, relationship between various agents. I did find that it improves heavily the extraction of agents to have a graph database on it.
Also use Dockling as a document processor for this particular use case. There's other technologies that are possible, but overall, you know, we're going to extract, find the agent, assess what are the requirements, and then put it in a graph database. And when we want to evaluate the credit proposal, we retrieve these agents, assess whether the conditions are satisfied, and then get a final memo on it. This started when I need to tick. Alright, so this is the Cloudera, Cloudera AI Workbench where we have deployed an application that is meant exactly for this.
First step, you know, is we're going to select a credit policy, and the credit policy is one that I made up based on another bank's policy, but I can't use exactly their own policy, obviously. So it's a fairly standard one, where you have some threshold that needs to be met and a lot of rules in it. That's very typical of what you can find in a bank. We'll upload it, trigger the first analysis pipeline which is going to retrieve a set of agents. The set of agents are converted into some JSON file, but first we can obviously select them, so you will see.
One example is, you know, the credit report age limit here. It just states that the application must be as recent as 30 days, just a very basic rule. We can also have automated selection. I implemented it so that I don't need to bother thinking, oh, I need to trigger this one, this one, this one, get this agent to run, this agent to run, and AI will decide which agents are relevant for this particular credit application. Then I upload a mortgage application in this case with everything you expect from a mortgage application, income and other topics.
The policy was extracted and some graph, knowledge graph was built, and this does help a lot in bringing, let's say, a relationship between various rules so that you can get all the mortgage rules together.
We also provide observability through another partner called Galileo , and in this case it provides insight into what's going on with the agent that we are running, just a basic observability tool. If we go back to our application, it should be finishing, I hope. So it does give us an overall compliance score. In this case, the credit report age limit was selected and it says it is fine. The credit report is three days old, so most likely this particular agentic rule that we defined is actually okay.
The bank statement age limit is fine. It tells us that the minimum down payment for first-time buyers is actually correct. I think in this case the down payment is at nine percent and the credit policy was at three percent, so it's able to make some computation. This was not given in the application memo. In the application memo we stated this is the down payment, this is the loan amount, and therefore that's how it works.
We use our partner one more time, Galileo, in order to assess what's going on and whether everything ran properly. We're able to log every single step straight from Cloudera AI in order to look at what is the execution. In this case it's a workflow, so it goes from left to right in a fairly smooth manner, and we're able to debug if something happened. Was this the right agent to pick? Was it executed properly, or does it need to be rerun at some point? That's something quite useful in order to show there's some cost management in case we are using some other things.
That's all for me. I think I've been showing on Cloudera AI and leveraging some of our great partners on AWS how we can modernize credit applications. It's not just theoretical or a demo. I actually implemented this for one of our customers in order to double-check what humans were doing. Let's be realistic, agentic AI is still at the infancy, and there's still a lot of duplicate checks that need to be done. But there's actually a bank using this in order to check whether some potential audit issues are happening inside the credit risk origination process, so I think it's generating great value.
If you want to know more, please visit us at our booth 1175 right over there. We'll be very happy to discuss more about it. Thank you very much.
; This article is entirely auto-generated using Amazon Bedrock.
Top comments (0)