# Agentic AI in Financial Services: What’s Next | DC Fintech Week 2025

- Channel: [DC Fintech Week](https://streameth.org/dc-fintech-week)
- Date: 2025-12-21
- Watch: https://streameth.org/watch/yt-VgZcq6fRVVk
- YouTube: https://www.youtube.com/watch?v=VgZcq6fRVVk

## Transcript

Good morning. In 1997 when Deep Blue beat Garry Kasprov at chess, we called it a step change in artificial intelligence. Then in 2016 when AlphaGo beat Lee Sedol at Go, we called it a step change in machine learning. What we are witnessing in 2025 is unprecedented and fundamentally different. We're seeing AI systems not just crawl through corpuses of data and reason and identify fraud. We're seeing these systems actually help broker deals, draft complex financial instruments, do real-time trading, and displaying reasoning that surprises even their creators. These systems are displaying agency. The ability to act independently, goal-driven, seek information, pivot when that information changes, and make autonomous decisions. So, the question is not whether agentic AI is going to reshape financial services industry. It's happening right now. The question to ask is whether you will be architecting this transformation for your organization. Or are you going to allow yourself to be disrupted? I'm Mickey Iqbal with AWS and in the next 30 minutes, I will demonstrate to you how agentic AI is reshaping the future of financial services today, right now. Most important, you will walk out of this room with a mental model on how you can optimize and tune this technology to be the architects of transformation for your organization. So, let's go ahead and get started. The way agentic AI is fundamentally different from what I call passive AI, which does reasoning, answers questions, is that it has the ability to have a goal-driven independent behavior and it can independently seek and solve objectives. It gathers information and based on that information, it can create different reasoning-based tasks. And then, based on the information that it gathers back from these tasks, it can make autonomous decisions. So, the secret sauce is in agentic AI's ability to formulate plans, take these plans and divide them into various tasks that need to be accomplished. Identify the set of tools that these tasks would be accomplished by, run those tasks, gather more information, do more reasoning if necessary, and then make a decision autonomously and with minimal human intervention. So, enough of the talk. Let's see how all of this works. So, we're going to go through uh a demo of agentic loan underwriting. So, this is a process which I'm sure many of you are very, very familiar with. So, before we go into the demo, let's talk a little bit about what it takes to do this manually. So, imagine you have a loan application that comes in, you assign it to a supervisor who would look at that loan application and they would have to accomplish multiple tasks. For example, they would have to verify the income of the uh person who's applying for that loan. They would also have to verify what kind of expenses that person already has, whether they can qualify for this loan or not. They'd have to look at policy and regulation and many other things. And they also would need to have a credit check done. Now, all of these activities are tasks that require seeking information, gathering information, verifying that information, making sure that it adheres to policies and standards of the organization, and then reasoning through it to make a decision. The decision could be to either approve or reject a loan or to ask for more information. So, what we're going to see next is the same process now being done through agentic AI. It's a short demo and I'll pause a couple of times in the middle just to kind of provide some explanations. So, we launch an automated demo and um this is going to mimic the process that I just talked about. If you could pause, please. So, in this particular case, we have a supervisor agent, which supervises three different agents, and I'll talk a little bit about those. But, the task in hand is they've been given a loan application. They've been asked to make a determination of, you know, whether this loan should be granted or not. Go ahead, please. So, once the supervisor agent starts looking at the application, they look at the various documents that have come with the application. And it reasons and figures out that there are three discrete tasks that need to be performed. Stop, please. Pause, please. In this particular case, they have to verify the income. So, there's a lot of income-related documentation that's been gathered. They have to also verify the expenses. So, there's a lot of expense-related documentation that's been gathered. And then, they have to also run some credit checks. Let's go ahead, please. So, the way it works is the supervisor agent then gives these tasks out to three sub agents: the investment agent, the expense agent, and the credit agent. Each one of these agents would then go work in parallel on a different set of tasks to try to get to the goal that's been assigned to them: verify income, verify expense, and get the credit reports. Pause, please. So, as this information is being collected, each one of these agents can has the ability to act independently and autonomously to provide a set of recommendations, which will then be passed on back to the supervisor agent. Go ahead, please. And as this processing is going on, you have to consider that the supervisor agent is now getting additional context from all of these agents. Let's stop, please. So, at this point in time, what you will see is that there's a different set of recommendations. You can go ahead. That have been uh provided by these different agents. And these are now being considered by the Pause, please. Uh by the supervisor agent. And one of the things that scrolled very fast that was identified was that the loan application had an income of about $120,000. But as the invest as the income verification agent gathered the bank documents or the pay slips, it verified that that did not match what was being deposited into the bank on a monthly basis. So, there was a discrep- discrepancy between the amount that was shown in the loan application and the amount that the bank was actually getting from the employer of that person. So, this basically flagged it and the supervisor agent decided that there's additional information that's required. And at this point in time, you can go ahead, please. Uh the super- you can get a human into the loop, which is extremely important in these cases, and uh you can get additional information. So, how are we seeing multiple organizations gain value from agentic AI today? In the financial services industry, there's really, you know, I could put this into three different buckets. So, the first bucket would be in workplace productivity. In this particular case, many of the financial services uh industry uh players are using uh agentic AI to power the work that their knowledge workers do. So, for example, research analysts, traders, uh folks that are working on risk assessments have the ability to use agentic AI into their workflows and be more productive, be more efficient, and optimize the cost of the work that they're doing. We're also seeing in terms of workplace productivity a lot of great work being done by the developers who are using agentic AI tools. And we've seen major transformations happening in this area. So, what used to be take years and months in terms of coding and development or even migrations and transformations is now being done in a much much shorter time frame. Secondly, we're seeing major business workflows. I kind of demonstrated one for you for underwriting and loan application processing, but there are others as well that are being transformed across the board using these technologies in the financial services industry. Now, these use cases could be all the way from intelligent document processing to fraud detection to creating hyper-personalized personal experiences for customer for customers calling in as well as different type of analysis and research that's required for for trading or financial services. And finally, we're seeing a lot of financial services organizations utilize agentic AI to run massive simulations at scale and running experiments to kind of figure out what are the opportunities that are ahead of them as well as what are some of the trade-offs that they need to make to avoid the pitfalls that might be coming in in their future. So, initially, I talked about hey, you're going to be able to walk out of here with a mental model for how you can be the architects of this transformation for your organization and how you can optimize this technology to really become a competitive advantage for your organization. So, the mental model that we've seen successful customers use when they've succeeded with implementing agentic AI for their business is kind of captured on this screen. So, first of all, we've seen that it's very, very important to work backwards from a real business problem that you're trying to solve. Trying to implement AI or generative AI or genetic AI for the sake of technology is never a good idea. Uh, we've seen organizations become very successful when they have identified a specific problem that they want to solve within a specific workflow. So, an example could be they they want to reduce the time it takes to process a loan, um, they want to have consider considerably more success when they're trying to identify fraudulent transactions. So, those are the kind of workflows that you can then apply genetic AI principles to and technology to to solve and and get good results. Secondly, these organizations that have been successful across the board, and there many of those that are customers, have created their data as their biggest differentiator. Originally, many of these companies have data that are that's in silos. There's very little communication, or I would say meaningful interaction or communication between this data. So, what these organizations have done as a best practice is they've applied modern data architecture principles, which allows data to be exchanged in a meaningful manner across these silos, but while it's properly governed and with responsible AI frameworks on top of them. And this data exchange has made it possible for them to take their most valuable asset, which is their data, and use it for insights driven, value driven work. Trust is non-negotiable in the financial services industry. All of the organizations that we have worked with, which have been extremely successful in deploying generative AI and getting value out of it, have clearly identified and implemented a responsible AI framework for their organization. They've also put in the right compliance, security, and regulatory controls, and built guardrails for their eyes AI systems that operate within the confines of these controls. And that's been very, very important. You also need to demand the best price performance. At AWS, we believe in model choice. Our Bedrock service has 30-plus different model providers that provide various number of models, large language models for helping with generative AI as well as agentic AI. And we believe that no one single model can fit all use cases. So, customers should be able to look at various models, run a comparative analysis based on the use case, and we provide these kind of comparative analysis tools, and then decide which model they want to use, and they have to be conscious about price as well as performance while they do that. And finally, as you saw in the demo, it is in order to uh accomplish complex uh tasks or use agentic AI in workflows that involve uh very complex tasks, it's really recommended to use a multi-agent approach, which is what I demonstrated to you. So, with that, I'd like to conclude this presentation and invite Chris over um to the stage so we can have a Q&A session. >> Much. So, that was a super, super, super interesting display. I guess I'll just uh we'll take a seat, I I suppose. Sure. Um so, I guess the first the first question is I think that with AI sometimes it can be a little bit overwhelming. So, you'll you'll see, "Hey, it could be used here. It could be used here." Uh I are there any use cases when you focus and you sort of double-click on early adoption? Uh any particular use cases in financial services that you are seeing a real pickup in where you're seeing enthusiasm um and really a push uh on your end to develop? >> Yeah, we're we're seeing a lot of different use cases. I think the earliest adoption we saw was in two different use cases. One was intelligent document processing >> Mhm. >> and the other one was uh chatbots for questions and answers. So >> Can you define that just for us, like particularly the first category, intelligent document processing? >> So, intelligent document processing, look, I mean, uh all of our financial services organizations have a lot of different documents. Some of these are uh in structured formats, so they could have data that's stored in databases. Some of them are simply in PDF files or unstructured format. Uh and then some of these could be metadata that's associated with different code or application that sits in multiple repositories as well. Uh they also may have uh software as a service solutions um where it's like ServiceNow or Jira or they have documents in Microsoft Office 365 or in SharePoint. So, there's a huge wealth of documentation across the organization. And there are different types of workflows that are associated with, first of all, identifying where these documents are, then, you know, collecting these documents, then analyzing these documents, and then, you know, coming up with insights or decisions, right? Uh so what we have seen is a lot of financial services organizations are trying to do this with generative AI. And, you know, we recently announced our um Amazon QuickSight, which provides all of this a virtual team member that actually goes out and indexes all of these different document sources within your organization. In addition to that, it also provides intelligent research based on all the documents that have been gathered. And then, it also provides visual insights, you know, graphically or otherwise about this. And so, we see this particular use case being deployed in the financial services industry. What I mean by chatbots, it's, you know, simply a question and answer type of service. So, again, based on all of these different documents, an analyst or a researcher within the financial services organization can chat with an intelligent agent and ask different types of questions, you know, about organizational policies, about compliance, or about, you know, previous research that's been done in that area. Similarly, what we're seeing also is a use case around improving customer experience. And this is where you may have customers calling in for different types of services. And rather than, you know, just simply responding to these by a person, there's chatbots that can actually interact with them and provide this information. So, an example I would give, one of our customers is the Rocket Companies. And Rocket Companies, you know, it's a they own a lot of different firms like Rocket Mortgage, Redfin, and several others as well. And you know, they're using the the the AI agentic technology and generative AI from AWS in developing chatbots. And what they have experienced, and you know, the CTO was on stage at our re:Invent conference talking about this. They've been able to reduce the time it takes to process a request that they get from the customer by 68%. And during these interactions that they're having with folks that are kind of calling in through this chatbot and asking different questions around mortgage processing, they've done 70% of their interactions are solved by agentic AI. And one final number I'd give you is that they've published a statistic that they are three x more likely to close on a mortgage loan once they've gone through this process. >> So, as a dumb lawyer, the first thing that I found super interesting in your answer, particularly on the on the documents sort of intelligent document processing it, is that your answer was an engineer's answer. It was like it was the answer of well, we're not even going to get into the sort of use case first. We're going to get into the functionality. And that there's interest in the functionality first, and then sort of we can find these cases. That is the functionality of being able to organize, identify, locate those documents, etc. That was your answer on my question, which I was thinking naturally about okay, is this, you know, trading purposes, customer-facing purposes, is this something that's being used internally for your compliance? Your first answer was first of all, let's just have a basic functionality of just organizing our our internal data and records. >> Yeah, because I think it's fundamental to have good data. Otherwise, you know, garbage in, garbage out. So, you have to, you know, the best practice where we've seen a lot of these organizations succeed is when they're putting the right fundamentals, the foundations beneath. But then in terms of the business use cases, I mean, our own finance team is using intelligent document processing to process invoices, you know, tens of thousands of them on a monthly basis, and identify issues with them as well. Uh we've seen a lot of organizations kind of actually uh you know, take, you know, years and years of historical data, scan those into electronic repositories, and then use those to kind of do different type of document analysis such as for mortgage and lending and underwriting processes as well. And finally, you know, we have stock exchanges that are using AI to kind of do smart research with different documents and analyzing risk. And then coming up with different measures to kind of mitigate these risks. >> So So when your customers come to you, and I'm even in the back of my mind I'm thinking a little bit about government, but just in general, do they know what they want all the time? I mean, there's a certain kind of a question that I I suppose you need your customers to upskill in terms of their understanding of gen AI in order to even kind of come to you to be able to ask questions in the first place or to make demands. I mean, like how What does that relationship even even look like? >> Yeah, first of all, you know, we have really smart customers uh and and they've invested a lot in and and they're continuing to invest a lot in, you know, internal education, enablement, as well as working with partners like ourselves to kind of upskill uh not just the um the staff that's, you know, doing the operations where gen tech AI would be applied, but also executive leadership as well as developers. Um our experience is that, you know, we always try to work backwards from the customer. So a lot of times Sometimes I would say a customer would come and they would describe a problem, but it's a it's a symptom of something. It's not really why that problem is occurring. So we have this working backward process that we'll work with them to really identify what is the root cause that's creating that symptom, and then we help them focus on that and and come up with a solution that's going to be scalable. >> I'm sorry. So the AI AWS the AI doctor, you know, to identify the different symptoms. I mean, you put a lot of emphasis on on sort of governance, and I think that that's important both in terms of the quality of the, you know, garbage in, garbage out question and the trust issue that you emphasize. Can Can talk a little bit more more practically from your vantage point? Like what what does governance sort of mean to you? I mean in terms of uh security management and and other issues? >> Yeah, I mean governance is is you know, to us it's it's of primary importance. Security we call it our job zero at AWS. So, everybody starts everything starts with security. Uh so, when we talk about governance from a security perspective, you got to have multi-layer security. So, it's fine-grained security all the way uh from identity to perimeter. Uh in in terms of governance, it's really about identifying, you know, who owns the data, what are the policies uh of that data in terms of who it can be shared with, what kind of access can they have, over what time can they have that access, and then really being able to ensure that you have a set of policies and a framework that you've developed that allows you to work with that data, right? Um so, we've actually published a AWS framework for responsible AI. Uh many organizations that we work with have their own uh versions of the framework as well. And uh to us it's, you know, having the right guardrails because you've probably heard and seen a lot that, you know, LLMs can create hallucination. Uh they can generate bad information. Uh it may not be factual. Uh and there's also the concept of bias and toxicity and precision and accuracy uh that really needs to be addressed when you're working with any kind of an AI application. So, for us, governance also means having the right guardrails in place so that you can avoid all of these problems or minimize them. And so, we've built technology that allows us to do automated reasoning checks as well as implement governance in a managed service environment with our bedrock service using guardrails that have different dials and knobs for you. So, you can tweak these dials and knobs based on your organization's responsible AI framework to get the right kind of results. So, all of these things together kind of help provide the governance structure. And of course, the last thing I would say is you got to have the right top-down stakeholder alignment and management sponsorship as well to make sure that this governance really works. >> So, generative AI is always developing. I'm I'm sure, you know, every single year we we've for like I guess I think this is the third year where we've had a demo and I'm like, "Wow, that's pretty neat." And more often than not, what's being shown didn't exist a couple months earlier. And, you know, when you when you were when we in our earlier panels we're talking about the velocity of of of innovation. I mean, how exactly do you and and does does does or would you suggest and what do you see over at AWS in terms of supporting the infrastructure for even launching the new AI products in in in the first place? I mean, when you think about the deployment of agentic AI systems, you know, what are those key infrastructure capabilities that you need and how does that relate to just even the speed of deployment for the needs of customers? >> Sure. You know, at AWS our approach to the development and deployment of new technology is based on having a comprehensive stack which is enterprise-grade, but it has the flexibility where if you're a startup or if you're a large enterprise, you're able to use it. And then, you need to have the right guardrails which I mentioned before around that. So, really, you know, we have these primitive that we work with to kind of develop and deploy any kind of technology uh that you see as becoming available as an AWS service. So, as an example, uh when we work with deploying uh or working with generative AI or large language model, our Bedrock service includes 30 plus different model providers. They are model providers ranging from uh Amazon itself to Anthropic, Cohere, Meta, OpenAI now, uh as well as many others as well. So, the idea is to ensure that this development and deployment provides the customers with choice, as well as flexibility. Uh we also are very much uh you know, deploying systems that use open source as well. So, uh our customers can bring in their own custom large language models and deploy it within this framework as well. Uh similarly, as we've uh you know, been working with agents, what we realize is that it's easy to kind of, you know, create an agent and deploy it in a demo environment, but you know, it takes a lot of hard work to go scale it up to a large organization, to tens of thousands. >> Well, you know, I I I want to go there for some maybe for our last question. So, we we had a conversation with um Karen Kimbrough, who's the chief economist for LinkedIn, talking about sort of the changes in the labor market, the needs of sort of new workers in they're looking for jobs, and the rise of uh of sort of skilling when it comes to uh uh uh almost any kind of financial services job, particularly in AI. We'll have soon Adrienne Harris from Department of Financial Services over in New York, who has dramatically increased the size of of her workforce and and and the fingerprint um for for AI. Um what do you see as the techniques, the strategies that the successful integrators of AI What are they doing? How are they taking the digital transformation process? Does it even look different from sort of earlier phases of of of digital transformation? Is there something new there? >> Sure. Since we got about a minute left, I'll try to give a quick answer. >> So We we look we're working with uh different customers at three different levels, right? In terms of education, enablement, and making sure that they can keep up with all the new advancements. So, one is at the executive leadership executive leadership level. So, you know, if you do a internet search on uh agentic AI for leaders, uh you know, a guidebook for executives, you'll find that publication from AWS, and that's specifically tailored towards those leaders. Uh we've got specific skill-based specialized skill-based education that we're providing in terms of AWS certifications. And then finally, the new approach we've taken is we've gamified this learning. So, you can actually go and search and look for uh you know, generative AI practitioner, uh which is a you know, based on a game that AWS has created so that we can gamify that learning and make it very very easy for, you know, the broad audience to get really familiar with generative AI and get >> So, so uh by the way, a huge huge proponent, you know, people talk about gamification. I mean, gamification is probably the one of the fastest ways to onboard people to new technologies. It's great for customer and consumer protection. Um big fan of that. But, uh thank you so so very much. Our time is up. Uh but, uh you know, just seeing how agentic AI is being deployed in financial services, you're at the cutting edge. So, thank you so so very much for your time. >> Thank you, Chris. >> Yeah. Thank you. >> Yeah, you're welcome. >> we're sure.