Adapting the artificial intelligence model

That’s the question we asked after the explosion of interest in generative AI in the past two years. The possibilities created by the development of AI tools like ChatGPT have captured the attention of the world and attracted billions of dollars in investment.

"There’s a race underway to understand how these technologies will disrupt and transform many industries – and financial services is no different.”

There’s a race underway to understand how these technologies will disrupt and transform many industries – and financial services is no different. How will it change the way banks and other service providers operate and how will it transform the way workplaces function?

We set our engineers the task of developing a new chatbot powered by generative AI that we are calling Z-GPT. Think of it as ANZ’s own internal ChatGPT, and it’s already being tested with some of our team.

This new in-house tool has been enabled by OpenAI’s models via our existing relationship with Microsoft. We have also used some of Google’s AI tools as part of this effort.

Z-GPT functions as a so-called ‘private instance’, meaning its large language model (LLM) is isolated from the public domain and restricted to just ANZ. It’s similar to the public Chat-GPT, but it’s just for us. This means we can use it to help generate or understand internal information, helping us explore generative AI more safely.

In developing Z-GPT we were able to utilise our AI Centre of Experimentation in India. The centre was established in Bengaluru to help establish our initial explorations of AI and assist the bank best position itself for what is coming.

Our centre has some of the best engineers in the bank who have been critical in developing Z-GPT. They are spending their time extending the initial experiment to work across subsets of ANZ internal data to verify the prototype can make the business possibilities tangible.

Organising information

The LLM behind this tool gives us the power to understand and generate general human language. But as we connect to ANZ data stored in the bank’s information technology systems, we start to see really powerful outcomes

One example is around search. Trying to find information in a corporate network can be challenging. Information is often organised, but trying to find something is hard and traditional searches across multiple document stores are often uneven.

Z-GPT allows us to build highly targeted searches for types of ANZ information using natural language questions for our staff. This can be “what types of home loan documents accept an electronic signature?” or “can you summarise the features of the ANZ Plus Saver account?”  

We can also take this further, allowing us to ask questions on highly structured data. For example, one experiment allows staff to ask about payments processing outcomes such as, “How many payments have we completed in the last 24 hours?” Or “What is the growth in payments processed over the last 6 months?”

Not only can the question be asked in natural language, we can also show the results as a written report.

This is hugely efficient: generating these reports removes the need for the technical work to turn an English statement into a database query, get the results and then turn it back to English. We might end up replacing many reporting-based tools with simple natural language interfaces.

There is already a backlog of suggestions and formal experiments from across the bank, looking at efficiency and productivity gains. Everything from AI helping engineers write software code to penning copy for internal websites.

{CF_IMAGE}

An important point to make is at ANZ we only treat AI as an augmentation tool. The person using the AI still has the accountability to review the AI’s response and ultimately make the decision on whether to use the insight or answer.

We are still experimenting at tuning our AI tools with internal data to see if this can help our staff solve problems. The use will be expanded as our confidence grows that this can make teams more efficient and reliable. We’ll only be using this for our staff until we have significant data and confidence that might lead us to enable AI tools for our customers.

That point may not be as far off as many people might think as this technology is already having a significant impact across the industry. Creating and learning along the way with Z-GPT has shown how complex this emerging field is.

We have restricted our experiments in some cases to only generate responses that are specifically anchored to a set of ANZ documents. However, results always have some degree of variability – that’s the nature of human language.

Prescient accuracy

Getting LLMs to reliably produce precise answers is tough. But it is still remarkable to ask an LLM a highly informal and unstructured question and get an almost perfect answer: it can be prescient in its accuracy.

We also hope AI will help create a more nimble bank when it comes to data. Historically we’ve had to consolidate data before we can use it.

This technology will make it easier to connect disparate data sources because it can take one question against two different databases and consolidate the answers. We may no longer need to consolidate data to use it.

As the testing progresses we are experiencing what insiders label “hallucinations” – when the AI generates an incongruent outcome that is wildly off the mark. 

We are still moving carefully and we won’t apply generative AI on a large scale until we have more confidence on the strengths and weaknesses. There is no point investing in technologies until we more deeply understand the space. We have to get the answers right and manage any ambiguity with confidence.

At ANZ we believe this technology shows tremendous promise. That’s why we’re experimenting rapidly within small teams. As our understanding and confidence grows, we’ll widen its usage as appropriate.

Tim Hogarth is Chief Technology Officer at ANZ