Building Smarter Financial Services: The Role of Semantic Technologies, Knowledge Graphs and Generative AI

11 min readApr 3, 2024

This is an abbreviated version of Ontotext’s Knowledge Graph Forum 2023 panel on Financial Services. The panel was moderated by Peio Popov, Sales Director Financial Services at Ontotext, with the participation of Aurelije Zovko, founder and CTO at Zenia Graph, Joseph Hilger, COO at Enterprise Knowledge, and Nimit Mehta, CEO at TopQuadrant.

In this post, we present you with insight gathered at the Knowledge Graph Forum during the panel on Financial Services. Read about the latest use cases and trends in the Financial Services industry and learn how Generative AI and LLMs complement with key capabilities of knowledge graphs.

The business why behind knowledge graph projects

Peio Popov: Welcome, everyone! I’m happy to introduce this panel with Aurelije Zovko, Joseph Hilger, and Nimit Mehta. We’ll talk about trends, leading use cases, and popular issues in the Financial Services industry. So, let’s start with our first question. We all know that for any initiative to succeed, we need the attention and the blessing of company executives. What are your thoughts about it?

Joseph Hilger: People are starting to understand that knowledge graphs are not just a tool for storing data and information. They are a contextual layer and we see this in a couple of use cases.

The first one is Enterprise 360 where organizations are saying, “The information about my most important assets is spread across many different systems. But it’s not enough just to have point-to-point integrations. I need something that defines what those entities are and can align them with the data.” A graph can do that.

The other use case where graphs are exploding is what Gartner calls a data fabric. The semantic layer or a contextual layer is a key piece of their fabric paradigm. It’s that layer that sits on top of all datasets and provides context to make the data both machine-readable and human-understandable. It also can fuel various AI applications and we know that now everyone is interested in AI.

Aurelije Zovko: For me, it is a digital transformation. When you try to automate business processes, semantics, and machine learning, knowledge graphs can bring a lot of value. That can help you get the attention of executives because changing business processes is not easy. There’s a lot of resistance and politics. So, this is a big driver for the outcome because when you are saving money for the business, you can measure it and see its value.

Nimit Mehta: I think that 2024 is going to be a buckle-down year, but, at the same time, we’ll see a rapid explosion of experimentation. This ties in with the point about digital transformation as well as with making sure data-centricity is at the root of our approach. It’s a great use case because semantics can both accelerate things and make them compliant. There are different types of modernization. There is infrastructure modernization. What Aurelije just said is about making sure your business workflows, data consistency, and schemas are perfectly mapped, and not lost in the transformation process. Architectural modernization, on the other hand, is what Joe was talking about. It’s about data-centricity and the semantic layer.

Business KPIs to focus on

Peio Popov: When you make a proposal based on the return of investment use case, what would be the key business KPIs that you would focus on?

Nimit Mehta: You are talking about the three big ones: cost, revenue, and risk. The pattern that we often see is that it depends on what part of the organization you are talking to. Early on, the use case is around saving cost, time, and other resources. As you move up through the business and start talking to product managers, project managers, and so on, it’s about revenue. And, when you get to the top, it’s about risks and existential threats to the business. The most forward-looking data leaders and business leaders say: “This business has to survive for many years. It’s not about just hitting the quarterly numbers. How do I make sure I can manage risk?”

Aurelije Zovko: For me, what’s also important is business productivity and one thing is to measure the business processes that bring revenue. Another measure is cost reduction because when you can process real-time transactions without human intervention, you have more time to do cognitive work and can process more transactions.

Joseph Hilger: I love the example about cost reduction. We are currently doing non-financial risk identification for a big Financial Services organization. So, we put together an ontology and some semantic processes to identify non-financial risk. In a 1-hour demo, we were able to show that we could do in hours the work that was going to take 10 external consultants months to page through and find things. This gets back to risk identification and also to automation and cost savings.

I also want to step back and revisit the data fabric idea. Today, people have all these separate data domains. They can’t see across them and say, “This problem in my supply chain will affect my sales.” So, it’s very important to create an environment where business users can ask a system some questions and get the answers they need. I think we are headed towards getting that knowledge and making it seamless, like talking to a data bot.

The value of business information architecture

Peio Popov: How about business information architecture? Is there anything that would bring such big value that you can target it on its own?

Joseph Hilger: We worked with a Financial Services firm and they had 10 different business units depending on how people approached them (through banking, investment services, credit cards, etc.). All of these units were asking the simple question, “When did Joe Hilger first start doing business with me?” So, they had different files all over the place, different terms, and so on. How do you solve that problem? You apply a semantic layer and once you can answer that first question, you can move to the next one.

Aurelije Zovko: I will start with a negative example. We worked with an investment bank and created a Raspberry Pi bot, with a huge green button, so traders could hit it when they wanted to trade. In the beginning, it was even voice-enabled and that was a disaster because anyone else could execute the same trade. But my favorite is actionable real-time insights. For example, when I want to insure some property and want to find out if the CEO has been involved in crime. Another example is to connect different events semantically and drive business decisions.

Nimit Mehta: Yes, semantics is a great solution when things are complex. Also, when you need to be comprehensive and can’t just address a part of your organization. That’s where semantics shines.

Also, I want to add one more thing here. The traditional technology approach to talking to the world is, “Our features are so cool. They are the best.” But what people need is the opposite. Which is, “Show me the value. Show me the outcome. Show me the ROI.” It’s interesting how that changes, once you prove the value. How this technology can expand the way we think, so the features become secondary.

This is the tip of the spear as opposed to the whole army. The knowledge graph is the army. It helps you do analytics that you can’t otherwise do. It allows you to generate hypotheses on your data. It enables you to look at common pathways and see what activities can change behavior, particularly in the marketing world. So, the graph becomes the biggest discovery engine for insights that you and your teams have never even thought about because of the limitations of the previous technology.

The promise of Generative AI and knowledge graphs

Peio Popov: Looking at the market, how do you feel about the promise of Generative AI and knowledge graphs? Do you feel this promise still stands?

Joseph Hilger: It’s interesting how you phrased that question. If we talk about the Gartner hype curve, I’d say it’s still going up. Based on people’s reactions, we have yet to hit the trough of disillusionment. And, I’ve always thought that part of the reason for this is that people get excited about new technologies, but they don’t truly understand how and where to apply them best. So, they throw them at everything. We have customers who have said, “For next year’s work, we need an LLM somewhere.” But when we asked them what they wanted to do with it, the answer would be, “I don’t know. We just have to have one.” Those aren’t business-driven requests but requests that try to meet someone’s excitement. So, we are still in that phase.

Having said that, I watched a demo from a vendor who was doing data fabric and they were asking their data bot complex questions and getting natural language answers. To me, that was amazing. So, the tools are in place and there is maturity. But we are still talking about what is a good solution versus a bad solution. That means we are still immature. The technology isn’t. We are.

Nimit Mehta: I just want to add that the technology also isn’t mature. If I had to invest in Generative AI, the smartest decision I would make is to put that money in my pocket. Everything is moving fast, so either you are all in or you are chasing rabbits and someone else will grab it before you.

Aurelije Zovko: My view is a bit different. I started AI and machine learning over 35 years ago and, at that time, it was also a huge hype. But, as we know, nothing came out of it. This time, it’s still a hype, but I believe it’s real. I’ve been waiting for a long time for this and maybe I’m very biased. But I’m providing such solutions together and I see good results. Maybe the technology is not production-ready, but I don’t think we need to be 90% accurate to go into production. Maybe for some businesses, 80% is good enough. Maybe even 70% is good enough because people do it manually and make a lot of errors. So, I believe it’s a hype but it will be changing the way we do everything very soon.

Key capabilities of LLM technology serving the AI hype

Peio Popov: But, based on your experience, when the dust settles, what key capabilities of LLM technology would survive the hype?

Nimit Mehta: I agree that some business processes could take 70%. But as you move up, the idea of putting in production something at 70% gets more resistance. I’ve been in countless CDO meetings where I’ve heard things like, “We won’t just throw ChatGPT in our organization, because it’s too unsafe. No way. It’s open. It’s unsafe.” So, maybe they are open to some testing and exploration, but when it’s prime time, larger organizations won’t run the risk of putting something untested. Especially now. Maybe later.

So, I see two models there: the language model, unbounded and trained on general human knowledge, and the one bounded by enterprise context. The best general language model will be created by whoever has the most data on training language like Google, Amazon, or whatever. But the enterprise context language model is what we are all doing right now with semantics. These are not statistical inferences. These are the facts that drive our organizations. What is a customer? What is our product? What is fraud? What is risk? How do we define that as a company? That’s our collective wisdom.

So, graphs will be the best training foundation. Because ChatGPT isn’t a black box. There are some ontologies involved, some taxonomic structures, and so on. And, you can plug it on top of your enterprise context. So, graphs enable something that doesn’t exist in any other technologies. We call it LLMs on rails. It’s the rails that govern the LLMs.

These are things like your governance model, the caching mechanism, the ability to view and discover things (where graphs will provide great value!), and the output-input validation. So, this will happen, but it’ll take time, and it’ll need to be done well to get it into the largest organizations in the world.

Aurelije Zovko: I will give a practical example. Building bots where you type in questions and you get some functionality is a long process. So, I think the ability to talk in English will survive. I hate to type and now I can speak and have a conversation in a natural language. I believe that AI bots will grow and probably will be everywhere very soon. So, I think natural language interactions and the capabilities to have AI assistants that will help you do your business or personal work will stay.

Joseph Hilger: I’ll steal from Nimit, because I liked what he said. The first thing we’ll all learn, like with any other technology, is that you need a foundation. Knowledge graphs and semantics are that foundation, the guiding rails we’ll follow. We’ll learn that you just can’t throw things at an LLM and expect it to work. You need to do the work upfront. But, ultimately, when the dust settles, we’ll find that there are repeatable processes we do every day where we can use LLMs. This applies to any industry. We’ll get good at picking what those repeatable processes are and we’ll apply LLM to them. That will definitely stick.

Key takeaways about knowledge graphs and AI

Peio Popov: What would be a short sentence that you want our audience to take away with them?

Joseph Hilger: Graphs offer the flexibility to solve problems that we couldn’t easily solve before. Do your research. Challenge your assumptions. Look for the hardest problems and have the experts in the space prove whether or not they can solve them. That’s what I’d say to you, and I think you’ll be surprised with what you’ll get from it.

Aurelije Zovko: Embrace AI and knowledge graphs. In your business and your personal life. Don’t fight the big changes that are coming and have fun. And, find some good use cases where it would be best to implement this technology.

Nimit Mehta: What I’d say is that maybe 15 years ago, semantics was something that was kept in a dark corner. But now it’s happening. It’s never been easier to get started on it. So, when you pick your first projects or pick the ways you want to change the world, have value in mind. Think about scale and flexibility. There are so many ways you can use semantics to power your organization. Pick the one that you can drive home and drive it all the way home. Don’t just create a cool endpoint for someone to do cool stuff. Do it all the way. Be very clear about what you are building and why. Even if it’s narrow. Don’t build the whole layer from the start. It’s not going to work. Build the thing and get it out the door.

Peio Popov: Thank you, gentlemen! This has been one of the best conversations I’ve ever had. Joe, Nimit, Aurelije, thank you very much for your participation! You were amazing.

Aurelije Zovko: Thank you!

Nimit Mehta: Thank you very much!

Joseph Hilger: Thank you! This was fun.