Learning Library

← Back to Library

Multi-Method Agentic AI in Banking

Key Points

  • Large language models (LLMs) are powerful but have known limitations, so solving complex problems requires a “multi‑method agentic AI” that integrates LLMs with other automation tools such as workflows, state management, business rules, and analytics.
  • Combining LLMs with proven automation technologies makes AI systems more adaptable, transparent, and better able to withstand regulatory scrutiny.
  • In a banking loan‑approval scenario, a conversational LLM‑driven chat agent can capture a customer’s intent (e.g., “borrow money for a boat”) and translate it into structured requests for the bank’s broader agentic AI framework.
  • The overall solution must orchestrate the chat agent’s output with machine‑learning models, decision engines, and compliance checks to evaluate borrower suitability and automate the lending decision.
  • This hybrid approach demonstrates how LLMs serve as a valuable component—not the sole tool—in building robust, end‑to‑end AI systems for complex, real‑world use cases.

Full Transcript

# Multi-Method Agentic AI in Banking **Source:** [https://www.youtube.com/watch?v=-mldKsBR0UM](https://www.youtube.com/watch?v=-mldKsBR0UM) **Duration:** 00:17:32 ## Summary - Large language models (LLMs) are powerful but have known limitations, so solving complex problems requires a “multi‑method agentic AI” that integrates LLMs with other automation tools such as workflows, state management, business rules, and analytics. - Combining LLMs with proven automation technologies makes AI systems more adaptable, transparent, and better able to withstand regulatory scrutiny. - In a banking loan‑approval scenario, a conversational LLM‑driven chat agent can capture a customer’s intent (e.g., “borrow money for a boat”) and translate it into structured requests for the bank’s broader agentic AI framework. - The overall solution must orchestrate the chat agent’s output with machine‑learning models, decision engines, and compliance checks to evaluate borrower suitability and automate the lending decision. - This hybrid approach demonstrates how LLMs serve as a valuable component—not the sole tool—in building robust, end‑to‑end AI systems for complex, real‑world use cases. ## Sections - [00:00:00](https://www.youtube.com/watch?v=-mldKsBR0UM&t=0s) **Multi-Method Agentic AI for Lending** - The speaker argues that while large language models are powerful, solving complex tasks such as bank loan decisions requires a multi‑method agentic AI that integrates LLMs with workflows, business rules, and analytics to ensure adaptability, transparency, and regulatory compliance. - [00:07:29](https://www.youtube.com/watch?v=-mldKsBR0UM&t=449s) **Workflow-Powered Loan Application Agent** - The speaker describes a system that tracks each customer's progress in a database, uses workflow technology exposed via the Model Contact Protocol as an agent to orchestrate loan applications, and initiates eligibility checks without relying on a large language model. ## Full Transcript
0:00Agentic AI using large language models or generative AI is a powerful architecture, no 0:06question, uh, but large language models have well- known issues and ... and constraints. And so if you want 0:12to solve complex problems, you're going to want to adopt what's called multi-method agentic AI, 0:18which combines large language models with other kinds of proven automation technologies so that 0:24you can build more adaptable, more transparent systems that are much more likely to survive 0:29regulatory scrutiny. So, as we said, large language models are a great tool right there. They're a 0:34wonderful tool in your toolbox, but they need to be not the only tool in your toolbox. You need to 0:39be able to combine them with things like workflow, to manage state or uh, decisions and business rules 0:46so that you can give explicit instructions to uh, to your agents. You need to be able to provide 0:52machine learning and other kinds of analytics so that you can build a robust solution that solves 0:59complex problems inside your agentic framework. So to illustrate this, I'm going to pick a complex 1:06problem, in this case, how does a bank decide to lend you money. So I have a bank, and it's got 1:12money that it wants to lend people. How does it decide that a particular person, that someone I'm 1:17going to have borrow money, is n ... a suitable person to lend the money to? So let's think about that 1:24in an agentic AI framework. So, the customer now doesn't want to use traditional systems anymore. They 1:30don't want to sit in front of a long form and forget all these details. They want to have a 1:34conversation. So first and foremost, I'm going to need some kind of chat agent that they can talk 1:38to. Now, our chat agent is a classic use case for large language models. They work great. They're 1:44really good at understanding what you're saying and ... and your intent and the voice you use and speak 1:48different languages and so on. Now, these are going to be there for large language model agents. So as 1:53chat agents are going to be a large language model agent. And in this particular case, I'm going 1:57to assume that I'm going to configure this agent not to try and have a whole conversation with you, 2:01but just to try and figure out what it is you're asking. And generally, you're going to either ask a 2:06question or you're going to ask ... tell the bank you want to ... to do something. These are the two things 2:10you're likely to do. So this chat agent is going to be prompted, configured so that it attempts to 2:15take your inputs and turn them into something that it can, uh, you know, pass on to the bank's 2:22agentic AI framework to say, this is what this person wants us to do, or this is the question 2:26this person is trying to answer. So, uh, let's take this particular person now. Let's say they want to 2:32borrow money, but they want to borrow money for a boat. So, what they want to do is they want to buy 2:36themselves a nice new boat. And so they have a question, which is can they borrow money for this 2:42boat? So this is a question. So now the chat agent takes their text which might be misspelled or 2:47mistyped or long-winded and says, oh, this person's asking the question about loan policy. What will 2:52we lend money for? And so it's going to pass that on into the system. Now, the next agent in this 2:59network needs to decide what to do. And so generally what that means is it's going to look 3:04to find another agent that can act for you. So these are often called orchestration agents. So we 3:10call this an orchestration agent. It's also going to use a large language model. All the large 3:14language models are in green. And that large language model is going to go look in some kind 3:18of registry where it's got a list of all the available agents uh, defined so that it can say, okay, 3:25go look in the registry, see if you can find an agent that seems to deal with whatever this 3:29question is. Uh, and then that a ... registry will also have some kind of standard definition for how to talk 3:35to these other agents. So now my orchestration agent can take the request about loan policy and 3:40say, okay, I need to look in the registry and find an agent that can deal with loan policy. So, in 3:45this case it's going to go say, okay, look, there's a ... there is in fact a loan policy agent. Now this 3:51is another large language model. And this one's using uh, this is sort of going to be to ... to take the 3:55bank's documents, and tell you what they say in some more intelligible fashion, rather than 4:00reading the whole set of documents. And these are generally going to be using what's called 4:04retrieval augmented generation, RAG. So I've got lots of documents and these might be product 4:08descriptions, they might be uh, risk policies, they might be uh, marketing materials, all sorts of 4:14different things. They've got this large volume of ... of uh, material, all these documents. And I'm going to use it 4:20to power an answer to you that's more helpful than just giving you the document itself. Now, 4:25these documents in most large systems are going to be stored in some kind of file management 4:29system ... And that file management system enables me to take different collections of ... of documents, 4:36because I'm going to have huge numbers of these documents potentially, different collections of 4:40these documents, and vectorize them, put them in a vector database and deliver them to different 4:45agents that you're going to use RAG to answer different questions. So I've got one set of 4:48documents designed to answer the ... the loan policy question, but I might have several other agents 4:54answering different kinds of policy questions that use different subsets of my documents. And 4:59all my documents are going to be kept up to date using the sort of normal file management kind of 5:03stuff. Documents come in, they get added to the repository, they get updated, they get removed from 5:08the repository. And so, I ... I've got an automated process that keeps all my policy agents up to 5:13date, revitalizes the database, resubmits them, retrains them so that at any given time, I've got a 5:19collection of policy agents that can answer policy questions. So now the orchestration agent 5:23says I've got this loan policy question, I've got a loan policy agent, I'll pass the question on, and 5:28it gets back an answer which explains in natural language with references to these documents how uh, 5:35the bank reviews lending money for powerboats. And so now I can pass that on to all the way back up 5:40my chain and tell ... tell the customer. So now I said to the customer, great, um, yes, this is our 5:46policy, this is when we lend money, this is the kind of boats we lend money on, how much we land, 5:50all those kind of questions. So this is all so far, so good. Now, so far, I've only needed to use large 5:55language models because I'm really just interacting with the customer, chatting with the 5:58customer. But now the customer says, okay, how do I ... I want to apply for this loan. How do I apply for a 6:04loan? So now they want to do something. They want us to not just uh, tell them about it, and they want 6:10us to actually sell them a loan. They want ... yeah ... and we want our agents to get to handle this. We don't 6:15want the agents to have to say stop at this point and say, oh, you want us to do something, you have 6:19to talk to somebody. We want our agents to be able to sell on our behalf. We want them to be able to 6:23act, do something concrete, in this case, fulfill a loan. Okay, so now this ... this question goes through 6:29our chat agent and again goes back to the orchestration agent. And the orchestration agent says, okay, I've 6:34got a loan application agent. I bet that's what I need. So it goes and finds this loan application 6:38agent. Now, the loan app is a more complex agent. So think about loan 6:45applications. There's lots of ... lots of steps, lots of data I have to collect. The likelihood is 6:49you're not going to complete it in one sitting. Yeah, you might have to go look for a document, you 6:53might have to go pick up the kids from school. Any number of reasons you might get interrupted. So, a 6:58loan application agent has got to remember state. It's got to remember how far you've got, what the 7:02next step is. It's got to understand how all this works. And LLMs generally are not very good at 7:08this kind of stuff. So what we're going to use is ... is an agent based on a workflow technology, a 7:13workflow platform. Now, workflow platforms are software infrastructure that manages, you know, 7:18processes, workflows. They typically have a definition of the flow in a visual model like 7:23business process model and notation BPMN that lays out the steps in the process. And then they g ... 7:29have a database where they create an instance every time a customer starts a process, so they 7:34can keep track of how far each customer has got through the process from start to completion or 7:39abandonment. And so they manage this state for us and give us a way, therefore, to know exactly how 7:44far you've got and let you restart and ... and reengage the process. And then they can throw out ones that 7:49get too old and all those good things. So now I have an agent, but it's going to be built on a 7:54workflow technology. Uh, and then almost certainly we're going to use something like MCP, the model 7:58contact protocol, to expose that workflow technology as an agent so that I can consume it 8:03as an agent. So now, the orchestration says, okay, um, what you need to do is ... is start this loan 8:09application agent. So it's going to go ahead and start it, and then it's going to respond to 8:13whatever the loan application tells it to do next, and pass it on to the customer and interact with the 8:17customer. Now ... so this is our first agent that doesn't need a large language model; this needs a 8:21workflow technology. Now this workflow is going to have a set of steps in it. And one of the first 8:26things it's going to want to do is determine if this customer is eligible for a loan. Now we have 8:31obviously, um, you know, we've told the customer what the policy is, but what we haven't done 8:38is establish that this customer is actually eligible for a loan. We haven't applied the policy 8:44to this customer. And this is uh, typically a decision agent. Now, a decision agent uh, is also generally 8:50not a good candidate for large language models, because if you're in any kind of uh, serious 8:55organization, you want to always make decisions consistently, right? You want to apply the same 9:01logic to each customer. You don't have to make the same ... give each customer the same answer, but you 9:05have to follow the same logic when you come up with an answer for each customer. You've often got 9:09to be transparent about this, you've got to be able to explain it to regulators, to auditors, to 9:14people who run the business. And large language models are not good at either of these things. So 9:19generally, we're going to use a business rules management system or a decision platform uh, that 9:23enables us to uh, manage all the logic behind one of these and again, deploy it at services that we can 9:29map in MCP and make available to our agent. So this eligibili ... eligibility agent is going to 9:35take a set of data from the workflow and say given that set of data, is this person eligible or 9:41not. So the workflow agent is going to have to access data. Now it might get that data from the 9:47chat conversation you've had with the chat agent and so on. But it may almost, almost certainly have 9:52to look up additional data. So it's going to go get customer data, uh, data from your internal systems. 9:56And again uh, that's going to use data technology exposed through MCP. So you have a data agent. So 10:02now the loan app goes and gets that data, passes it to the eligibility and gets a response back. So 10:06now I've got a sense that you are in fact eligible for this loan. Good news, you're eligible, chatbot 10:11is having a nice conversation with you. And it feels like a natural conversation to the 10:15customer. But behind the scenes, it's being managed in a very concrete, reliable way. So now, you're 10:22eligible and you say, great, um, you know, then you keep going through the process and the loan 10:27application process, you know, it might be talking to the orchestration agent, but more likely it's 10:31got a concrete definition for all these pieces. And the next piece in the process is going to be 10:37to actually decide if it's going to loan you the money or not, the ... the ... the origination decision as 10:42it's called. So, you know, you're eligible for a loan. Now you have to tell me exactly what you 10:48want the loan for. And I'll tell you if I'll lend you money and on what terms I'm going to lend you 10:53money. Now, that loan decision, what are the critical things it needs to know is ... uh, you know, 11:00your credit bureau data. It needs to know what asset you want and so on. So there are other data 11:04elements. So we might get another data agent to go get your credit bureau data, for instance. Um, we're 11:09going to pass all this data to the loan. But we need to know the asset. What are you borrowing 11:14money for? What is it you want to buy? And so, uh, the orchestration agent uh, needs to find that out. And so 11:21it's going to ask the customer do you have an ... an asset you want to buy. And the customer says, yes, I 11:25have this boat. But the orchestration agent's now got two choices. They can either ask you to fill 11:29in a whole bunch of forms or chat in ... interminably about this boat, but it recognizes that it's got 11:35another kind of agent—it's got an ingestion agent. And the document ingestion 11:42agent's job is to take a document and turn it into data we can process. And so it says, well, do 11:47you have a document describing this boat. And in fact, I do. I have a brochure and this brochure has a 11:53... has a little picture of the boat. Yes, it's got a little picture of the boat and it's got a bunch 11:57of information about the boat— how old the boat is, how big it is, what it weighs. And uh, it's got a sort 12:01of handwritten uh, number on it, which is the price the guy at the dealership gave me. And perhaps his 12:05business card is stapled to it. Um, so it's got all the information I need. And so the agent says, 12:11great, go ahead and take a photo of that, scan it or whatever and pass it to my document ingestion 12:16agent. And now what the document ingestion agent is going to do is it's going to take the requirement— 12:21What does the loan application agent need to know about an asset? And then it's going to look at the 12:27document and see how much of that information it can find. Now large language models are 12:32tremendously good at this. It doesn't matter if the brochure was printed on a cheap printer and I 12:36p ... pulled it out too fast so the back page is a bit blurry, and then I've stapled a business card to 12:41it and he's handwritten the number to it. It will work through all of that complexity and extract 12:47data from it. They're remarkably good at this. So this ingestion agent sucks all the data out of 12:51the brochure and says, great, you know, here's what I found. And it passes that on to the ... to the rest ... back to 12:57the loan application. And that says, okay, that's the complete definition of the asset. I know what 13:03kind of boat it is, how old it is, how much it costs and which it weighs, what size it is. And 13:07that was the set of data I needed about the boat. So now I'm good. I've got the credit bureau data 13:12from the credit bureau, I've got your uh, bank data from you as a customer, and I've got this brochure 13:17about the asset. Now I can go ahead and make a decision for you. So, uh, I'm using large language 13:23models to find out what you want, I'm using large language models to help me get data in very 13:28quickly, but I'm relying on workflow and decision agents and data agents to do the sort of heavy 13:32lifting of the decision in the background. There's another use case for large language models. So 13:38when I start getting into uh, human interaction again. So let's say that uh, the answer comes back from the 13:44loan decision that we're not 100% sure we want to lend you the money. Uh, lots of decision agents are 13:51like this. They have a yes, no, maybe kind of cadence. Yes, we will; no, we won't; maybe we will. But 13:56we have to clarify. So the loan application comes back and says, hey, customer, we're going to need 14:01you to talk to a customer service rep, uh, someone in our call center and resolve some 14:06inconsistencies or some issues that are in your application. And the customer says, you know, I'm 14:10too busy, I gotta, I gotta go, um, I'll check back in. And they leave. They leave the chat, they leave the 14:16conversation. Now this is okay, right? Because remember we're using a stateful loan application 14:21so we know how far they got. We know what the next step is. And so when they come back later and say, 14:26hey, I'm back, uh, can I restart my loan application, the chat app, the chat can say, okay, they want to 14:32restart a loan application. orchestration agent knows, it goes, talk to the loan application 14:37and the loan application agent says, yep, I got a loan application process in flight for that 14:41customer. Uh, once you've authenticated them, I'll tell you where we got to. And where we got to is 14:45they have to talk to the call center. So they say, fine, I'm ready this time. So now we have a call 14:49center rep. Now they work for the bank, and their job is to resolve whatever this issue was. And 14:56two more use cases for large language models occur at this point. The first is what you might 15:01call a ... a companion, um, agent. You know, we used to call these copilots, but obviously Microsoft made 15:08that a bit more complicated. Uh, so we have these companion agents, and their job is to help this 15:12person. And the reason they need one is that by this point, a lot of information about you—there's 15:18all your bank records, there's your credit bureau records, there's the application you find, there's 15:22the information about the boat ... There's all this information about you and about your application, and 15:26this, um, agent needs to be able to quickly answer questions about that information, access 15:33information. So it's a classic large language model looking at this collection of information 15:38about you, some in documents, some in databases, and making sure it knows how to answer. That may also 15:43refer to some of the policy documents, uh, access the ... the policy agent. So it ... it's got access to the rest of 15:50the sort of corpus that you have as a bank. And one of the other large language models is going 15:54to immediately talk to is what we call an explainer agent. So decision agents, one of the 16:00reasons we use decision technology to build decision agents is that they are good at 16:06explaining, documenting how they made the decision. So unlike a large language model, which can be a 16:10bit black box, uh, a decision agent is going to be built on a decision platform that very explicitly 16:16records exactly how it came up with the answer it came up with So in this case, I have a log of 16:22exactly why we said maybe to this customer, but it's an internal log. It's designed f ... for the 16:28bank to understand why you got to. Maybe it's not necessarily going to mean anything to the 16:33customer. You might talk about a loan policy that you're in breach of or something, but it's going 16:37to use internal terminology for this stuff. So what an explainer does is basically turn that 16:42into a natural language explanation for the call center rep. So here's what that log means, so that 16:49when you talk to the customer, you can explain aah, the problem here is that your credit bureau 16:53says this. But the W-2 you sent us says that, right. And so they ... they ... they can get a natural language 16:58explanation of what the issue is. So they ... they then work with the customer, they're adding new data to 17:04the system, they're working with their companion agent, they eventually resolve all of this. And 17:09then, uh, they don't approve the loan; what they do is they resolve the issue. Once they resolve the 17:15issue, they tell the loan application agent I resolved the issue, and it retries the decision 17:20again. That's how the process is set up. You resolve the issue, I resubmit it to the loan 17:25application. This time it says yes. And lo and behold, now you get the money so you can go off 17:31and buy your boat.