The Fast Track to Value With Generative AI in Financial Services

“You must automate everything and use Generative AI in all your systems right now.”

Every boardroom in every company, 2024

Yes, we all understand the mission, but how exactly do we carry out this “Generative AI in Financial Services” task? The industry is full of legacy software and ever-changing compliance issues, not to mention challenges with scaling resources and limited budgets.

You would be forgiven for taking an ‘if it isn’t broken, don’t fix it’ stance.

Probably every company would love to spend tens of millions of dollars creating modern platforms from the ground up that have flawless generative AI in finance built into every feature, but practically speaking, that’s just not going to happen — we live here in reality.

So what do you do when you live in reality? Create realistic solutions that help address immediate problems, and make well sure that you’re not destroying existing software systems or processes while you do it.

That said, there are many ways to apply this new technology. The use cases of generative AI in financial services are nearly endless.

You could use GenAI in financial services to automatically create and process documents like loan applications and regulatory filings by having GenAI read unstructured data sources, create summaries, and ensure that the documents comply with banking regulations and standards. You could train a model on your knowledge base and use it for automated customer support or personalized financial advice.

The list of ideas is long, and they’re not particularly hard to come by. The difficulty lies in deciding the right ideas for GenAI in financial services that will directly help your organization.

This article will look at a use case for adding generative AI to KYC processes. While this example could be helpful for some banks, the main value is the approach rather than the specific solution.

By applying the principles outlined in this article, any organization could identify and prototype the use cases that will have them cheering in the boardroom.

Step 1: What’s the need for GenAI in Financial Services?

The first step for our POC was to brainstorm some of the biggest challenges that GenAI could solve in Financial Services with our in-house bankers and technologists. The one that came to the surface was customer screening during the KYC process. It’s a perfect fit for a number of different reasons outlined below.

KYC has a massive volume of data

Many banks use Lexis Nexis or Thomson Reuters to search the vast amounts of data on a potential customer like Mr. John Smith, who we’ll use as an example. As of today, 67% of companies conduct manual, or mostly manual, adverse media screenings. This means that lots of people are going through lots of information on Mr. Smith.

In short, this is very inefficient.

Why have people process all that data when LLMs can read it much faster?

Current systems generate many false positives

Automated screening systems like Thomson Reuters usually generate false positives, which means that even when you’re able to narrow down a list of media sources, you’re going to end up with a lot of materials that are irrelevant.

These false positives arise because the screening searches generally match your subject (John Smith) with negative keywords like ‘terrorist’, ‘robbery’, or ‘parole’.

Was Smith robbing the store? Was he the owner of the store? Was that person even the John Smith that we are searching for, or a completely different John Smith?

At most banks, humans answer these questions right now because they are able to put Mr. Smith into context.

This is the perfect job for GenAI in financial services.

Why have people go through each false positive when a robot can help them save lots of time by giving a summary of the context?

There are more languages than just English

English is a great language, but what if Mr. Smith spent some time robbing banks in Paris?

It’s as unfair as it is impractical to expect all of the people working on adverse media screening to know French, Arabic, and Portuguese in order to understand exactly what Smith got up to in his latest tour around the Mediterranean.

The good news is that LLMs mastered the art of translation long ago – a skill that can be applied to chase Mr. Smith’s misdeeds all over the globe. There’s no need to continuously reinvent the wheel when it comes to GenAI in financial services.

Step 2: What will the prototype include?

Now that we understand the need, it’s time to figure out what this looks like as an application, tool, add-on, or other solution. Our wishlist for features included the following – most of which directly address the needs outlined above.

Manage the volume of false positives

This feature has GenAI distinguish the context in which the person has been flagged.

To give a real-life example: I searched my own name in Lexis Nexis and found that I came up in a negative context because of an article I wrote years ago about AI that somebody I don’t even know had cited. It only took a brief read to understand that I had no association with any nefarious behavior – yet there I was as a false positive, causing a potential KYC officer to use precious time reading about me.

To implement this feature, we had GenAI read the text in its entirety and infer from the context whether the individual in question was associated with wrongdoing according to a given set of criteria. After that, the model makes a brief summary of why it made this determination so that a human can browse the summaries rather than digging through the original text.

The criteria we use to understand the context can be customized for each bank, requirement, or process, depending on the needs of the organization. GenAI in financial services doesn’t have a to be one-size-fits-all.

The context matters

While false positive filtering helps reduce the actual number of articles that a KYC officer needs to go through, role definition, explanation, and summarization speed up the process of going through the remaining flagged articles.

It’s important to note that the PoC can’t eliminate all false positives, but it certainly cuts down the number, which lightens the load on a human – exactly the aim of GenAI in financial services.

Here, the model uses context to understand when the individual in question is actually doing something wrong, gives a brief explanation of why, and flags the person as a risk so that a human can make a determination. As the human KYC officer is going through the summary of the person’s vile ventures, they can make a judgment call as to the actual risk tolerance of the organization.

If the summary provided by the bot isn’t enough and they need more information, the full text is just a click away.

Bundle up the bad events

When Mr. Smith mugged an old lady in Spain and every news source within a 100-kilometer radius wrote about it, Lexis Nexis ended up with a laundry list of hits.

Usually, these would be sprinkled throughout the results, and as a KYC officer is going through them, they’ll mentally have to note if a given publication is talking about that same unfortunate event in the back alleys of Madrid.

Event-based aggregation takes all those negative articles, groups them together, and gives a summary of the event in question.

This means that the KYC officers can make a determination about the event as a whole, regardless of the number of publications reporting on it. This saves an incredible amount of work trawling through all the individual articles – GenAI in financial services at its best.

Find the right man

According to some sources there are something like 45,000 people in the United States with the name John Smith. As you can imagine, when he applies for a home mortgage Lexis Nexis returns a heck of a lot of hits.

One way to get around this is by filtering based on personal information extracted from the text, such as Smith’s birth date.

This can greatly reduce the number of false positives and help to hone in on the exact John Smith who would like to buy a house or invest in our latest project.

Step 3: How do you productize and integrate the PoC?

The next step is to assess the value of this PoC. Does it address the fundamental needs described above? If not, what should we do? At this phase we’d either abandon the idea, which could make sense as little time and money have been put behind it, or refine the idea so that it better addresses the need.

If it does prove out the requirements the next question is: how do we take that concept and actually use it at scale within a bank or other financial institution? The idea isn’t just to get any GenAI into financial services; it’s to get the right tools that work seamlessly with the organization.

There are two main options here: we can integrate it into the existing system or use it as a standalone application or tool.

There are pros and cons to each one.

An integrated solution is a fantastic choice for organizations that have a modern enough core that they can essentially use the same features but put the raw power of GenAI in finance behind them. A litmus test for this might be something like: if there are still people alive who understand how your banking system works, you might be able to update it with GenAI. If not, you’ll likely need option two.

A standalone application is great because it doesn’t disrupt anything that is already happening. The drawback is that it’s obviously not integrated, and it will require some additional training and a change in process in order to start using it. On the positive side, you can build more features into the application without worrying that it will break critical functions from your existing system. For some, this could be the best option for getting GenAI into financial services.

The right decision on this topic should come out of a conversation between technical, business, and compliance leadership.

Once a decision has been made, the team is well-positioned to plan the solution architecture and then build it out at scale.

Step 4: How do you get a GenAI financial solution that works extremely well?

Once the scaled application (or feature) is released, we’d go through an iterative cycle of running it in parallel with the original methods and measuring the results to ensure compliance and accuracy, and mitigate risk. If it falls short in any categories, we’d need to gather requirements and then refine the tool.

A process like this would isolate individual aspects of functionality and make sure that each one meets the requirements. For example, we would start by marking all false positives but make sure that a person reviews each article. After the system hits a particular benchmark that all stakeholders are comfortable with, for example, the system is 99% sure that a given input source is a false positive, we can start to expand the use of the tool.

Even at this stage, we’d still send a small percentage to humans to continuously validate that the tool is working. Afterwards, we can slowly reduce the percentage that is double-checked as we gain more trust in the accuracy of the system.

The same method could be applied to other cases like article summaries. At first, humans would read the articles in full. Thereafter, they’d be shown a summary, and they would need to validate it. After, we could reverse the validation process so that a person would need to see the summary first and then validate it by reading the article.

By going through this process, we both refine the GenAI for financial services tool and also start to train the workforce at the same time. The best way to do this is for people to genuinely buy into the new process and see how it benefits them and makes their jobs easier.

After a few iterations and a sign-off from key stakeholders like the compliance dept., we can safely move away from the old system.

The best time to plant a GenAI for financial services tree wasn’t 20 years ago

The best time to get started with GenAI in finance is right now.

It’s still early days with GenAI in financial services; however, we’re already seeing the incredible, high-impact, real-world difference that it’s capable of. And best of all, you can explore and prove out these ideas very quickly.

The first iteration of the tool described above only took our team about a month. That was from concept and requirements all the way to basic functionality and extremely lackluster UI. And again, that’s just an exploration of generative AI use cases in finance.

We are not talking about ripping up the financial software that you depend on every day – we’re talking about supplementing it with powerful GenAI for financial services tools or building features into it that save you time and money and differentiate you from the competition.

The first financial institutions to act in this way will definitely be the main beneficiaries.

Not every GenAI for financial services project will be a smashing success, we know that, but the approach of continuously examining our legacy systems and seeing where this emerging technology can make a difference is extremely valuable. Making that mindset and approach a common practice in your institution will position you very differently from many of your peers.

Simply put, the more run the business operations that you can pawn off to robots, the more time, resources, and money you’ll have to grow your business. And there are quite a few GenAI use cases that could help you grow, too.