In this video, we demonstrate a real world example of how LLMs can be used in the financial services industry.
We're going to be using the example of a financial crime analyst who works for a international payments business and they are responsible for monitoring these payments for fraud and compliance breaches.
We will demonstrate technical approaches such as retrievel augmentation over a real time analytical database.
Video Transcript
In this video I wanted to walk through a real world example as to how large language models and generative AI can be used to make a employee much more efficient and effective.
We're going to be using the example of a financial crime analyst who works for a international payments business and they are responsible for monitoring these payments for fraud and compliance breaches.
And we can see here we have a list of transactions and they are going to different countries.
Each transaction has different amounts and they are going to different beneficiaries.
Now what we're going to do is look at a few examples of using large language models and generative AI to make this agent more efficient and make their business overall more effective.
Firstly we are looking at a technique here called RAG or Retrieval Augmented Generation where I'm asking questions about the data which is being captured in real time.
So for instance a simple question how many transactions have taken place in Israel and we can see that there were 89 transactions that took place in that country.
Here I'm going to say what was the largest transaction in Israel and the answer is was $200,000 approximately.
So here what we're doing is the analyst is kind of looking at their data through a new lens and digging through it to find situations which they need to be concerned about.
Here a similar example we're asking how many transactions have took place where the person is listed on a politically exposed person database and we can see that there are 148 transactions which have taken place and this is something which a compliance officer would definitely worry about.
So basically this is giving like a financial crime analyst a new tool to understand the state of their world and identify situations which they need to dig into.
So next what you could also do is say something like which beneficiary has sent the most money to find like the largest transaction today and maybe that is something which I need to go away and investigate.
And we can see here that it was a customer called Simphia Smith who has sent the most money, a $385,000 so a very large foreign exchange transaction.
So next what I'm going to do is go and ask a different foundational model, a foundational model, Claude which we are using via AWP Bedrock, to produce a letter to Simphia Smith asking her about the source of her wealth.
This is something which regulated financial organisations have to do all the time.
You know people are breaching limits so we have to reach out to them to understand you know that there's no money laundering risk.
And here the generative AI model has created that letter for us and we can do things like change the terminology and the style to get it into the format that we need.
What I can also do here is say add an item to my calendar to remind me to follow up with Simphia Smith in one week and using a feature called agents we can go away and we can interact with a calendar or any other API and it will schedule that follow-up.
So this will you know help to contribute to remove risk if we don't forget these follow-ups.
And again you know there's a number of ways to achieve this but this is a tool which could potentially make organisations more efficient, secure, controlled, compliant.
The next technique I'm going to show is something called knowledge bases and this is where we have documents like this which are proprietary to our business and here we're looking at the government guidance for FX businesses and if we pick a line at random we can say that it shows that if we are going to display rates to our customers, imagine in an FX Bureau, how should they be presented and the document says things like they have to be clear and easy to understand and complete and things like that.
So as an example I'm a compliance officer I don't know what the regulation is so I've asked if I give an exchange rate indication how should it be presented you know in my physical store.
Now what the LLM has done is gone and configured that it's gone an interrogated document and it has come back with an answer explaining that it has to be clear, comprehensive etc etc and imagine how useful that would be if you had thousands of pages of regular regulatory information.
There's also a sentence here that there is no requirement for all of the information to be displayed on the same board.
So imagine I'm designing my store and I'd like to say do I have to put all of the FX rates on the same board or could I have one for Europe and one for the USA for instance.
What I'm going to do is just ask that question in plain conversational English do I have to display all of the efforts on the same board.
We've gone away and we've interrogated the document and it has responded and says no there's no requirement to put them on the same board.
Finally there's a section in the regulation about receipts and it says what do you have to include on a receipt if a customer makes a transaction with you and I'm going to ask for knowledge base again if I provide a receipt what do I need to include on it.
It's going to interrogate the document and the answer is that you have to include things like the exchange rate, the net total amount, the date of the transaction and the name and address of the trader.
Now what I can do having interrogated the knowledge base is I'm going to copy the things which I have to include on my receipt and then I can flip back to a generative AI model and I'm going to say create a receipt which contains the following information and then I'm going to paste in that list so it needs the exchange rate, the net total amount, the name and address of the trader and here what we're doing is using generative AI to create that receipt for me directly from the regulation.
So that's the city example but what I'm trying to show is how we can use things like RAG to understand the state of the world.
We can interrogate knowledge bases and then we can flip into like generative AI modes where we are saying can you send an email, can you schedule something on my calendar, can you generate a document for me and then maybe I could go back into an asking questions mode and it's very you know