• Home
  • /
  • Storage
  • /
  • ITWeb TV: FNB to deploy in-app GenAI agents

ITWeb TV: FNB to deploy in-app GenAI agents

Admire Moyo
By Admire Moyo, ITWeb's news editor.
Johannesburg, 19 Apr 2024
In this wide-ranging interview, Dr Christoph Nieuwoudt, FNB chief data and analytics officer, speaks to ITWeb news editor Admire Moyo about how the big-four bank is innovating with generative AI. As the bank increasingly leverages generative AI, Dr Nieuwoudt also discusses FNB’s plans to recruit more data analysts from the bank’s graduate programme in a move aimed at building solutions from the emerging technology. #itwebtv #fnb #genai

First National Bank (FNB) is building a vector database to support generative artificial intelligence (AI) models that it is tapping into in a move aimed at improving the functionality of its mobile application.

Vector databases play a crucial role in supporting generative AI (GenAI) by providing efficient storage, retrieval and representation of data, as well as enabling transfer of learning and data augmentation techniques that improve the performance and robustness of generative models.

In order to build this vector database, the big-four bank is hiring graduates from across South African universities to train the system to efficiently respond to banking queries.

So says Dr Christoph Nieuwoudt, FNB chief data and analytics officer, in a wide-ranging interview with ITWeb TV. He also reveals the bank is looking to introduce GenAI-based agents to help answer various customer queries.

In 2022, Nieuwoudt told ITWeb that the bank was looking to build a 1 700-strong data analyst army.

“We are now sitting at over 2 000 full-time and contract data analysts. It has been an area that has been growing rapidly, probably from about 1 300 two to three years ago. It continues to evolve, with GenAI being the latest topic that is driving it.”

GenAI has the potential to revolutionise various aspects of banking operations, from customer service and risk management, to marketing and cyber security, enabling banks to enhance efficiency, mitigate risks and deliver more personalised services to their customers, he notes.

Moving targets

“We are not growing at the same rate that we thought, but we are always bringing in new graduates. One of the big reasons is that there is quite a big shift in the skills that are being required.”

Nieuwoudt points out that GenAI means there is a need for different skills. For example, he explains, programming in Python is necessary if wanting to use any of the large language models and GenAI libraries.

The reason for building a vector database, he says, is because the bank is not making use of public GenAI models, such as ChatGPT, which are not secure.

“If you think about ChatGPT, it’s a publicly-facing bot engine that people can use. But if you are using that [ChatGPT] or Bard or any other publicly-facing bot, it does not ensure privacy of your information,” he states.

“As an institution, we cannot allow our staff to use these tools for banking purposes. In a private virtual cloud, we’ve got systems that can use similar underlying large language models but that will be sitting in a ring-fenced environment where any data that we feed into it, or any questions that you ask, will be sitting in a private environment.”

Dr Christoph Nieuwoudt, FNB chief data and analytics officer. (Photograph by Lesley Moyo)
Dr Christoph Nieuwoudt, FNB chief data and analytics officer. (Photograph by Lesley Moyo)

This means FNB can completely protect the privacy of its clients’ data. “You can ask ChatGPT about any topic in the world, but that’s not necessarily useful for us. In the bank, when you ask it for a specific task, you will need to be able to guide it and prompt-engineer it.

“We are building a vector database – that’s our internal content that it references to in order to ensure the answers are accurate for our purposes.

“This is really why we need the new skills. Setting it [the vector database] up requires specialised skills so that the rest of our staff are able to use it and figure out limitations and learn how to phrase questions.

“Often, people tell me these systems [GenAI] are really dumb, saying: ‘I asked it this question and look at the answer that it gave me.’ But before you look at the answer, you need to look at the question. Did you give it the context to really understand what it needed to answer because it is not smart like a human being – it doesn’t understand who it’s talking to, what is their role, what will be of interest to them. It doesn’t understand any of these things, so you have to give it a lot more context.”

GenAI transforms CX

Nieuwoudt indicates that by using the data analysts, the bank is looking to improve personalised experiences, as well as the search and chat capability on the app.

“What you would find today is that the FNB app is incredibly rich in functionality. We often say internally that the app is almost too rich. There are more than 100 different applets that you can pick from − and the truth is many customers are not using even a fraction of that capability. We want to make it a lot easier for customers if they want to ask questions.

“Today it is using natural language to answer questions, but what it is not yet doing is giving you a nice answer that summarises the answer and then gives you the links.

“We already have a secure chat functionality where you talk to a banker, but the banker does not answer immediately. We are trying to get the GenAI agent to be able to give you an answer immediately, but enable it to route you to a human agent when that is necessary.”

He adds that the bank hires graduates from all of the universities across the country, although it works directly with four universities.

“We are looking at people that typically have a four-year degree; so, if you are doing a three-year degree, you must do an Honours as well. We are also interested in people with a Masters degree but not necessarily PhD because there are a few of those.

“In my field, we typically look at mathematics and statistics, as well as engineering, actuarial science or computer science.”

The teams are interdisciplinary, he says. “You don’t need five people with the same skill; you typically need some data scientists, you need data engineers, programme managers, etc.”