ANZ debuts chatbot to experiment with AI

“One of the things we think it will be able to do is answering natural language questions, such as, ‘How many payments have we processed in the last 24 hours?’,” says ANZ’s chief technology officer Tim Hogarth.

“For generating reports, you won’t need to do technical work to turn an English statement into a database query, get the results, and turn it back to English. You will end up with natural language interfaces for traditionally more reporting-based tools.”

ANZ is also conducting AI experiments to create copy for internal websites, compare documents, and help engineers write software code.

The work is being led by a new AI Centre of Experimentation, based in India, where ANZ has huge operations. “The centre has some of the best engineers in the bank spending some of their time rapidly prototyping these experiments,” Mr Hogarth said.

National Australia Bank and Commonwealth Bank are also looking to the subcontinent for AI engineering expertise.

ANZ’s large language models (LLMs) have been developed with OpenAI’s ChatGPT system, under an existing partnership with Microsoft. It also tapped Google’s AI tools.

One problem with the public LLMs is they use data which is frozen in time when the model was trained. ANZ has adapted this to live data and focused the models on its own policies and documents to improve accuracy.

“Getting large language models to get precise answers is tough, but when you take the LLM and tune it with a subset of documents, it can be prescient in its accuracy,” Mr Hogarth said.

AI will also help banks find data kept across various IT systems. “Historically, banks have had to consolidate data. But I believe this technology will make it easier to connect disparate data sources,” he said. “You don’t have to consolidate it to query it.”

Challenges include ensuring answers are accurate “and none of these hallucinations sneak in”. Hallucination in AI refers to it making things up from what the technology discerns. Westpac chief technology officer David Walker has also warned about the risks of generative AI, including its capacity to hallucinate.

“We are not ready to apply generative AI on scale until we have more confidence on where the strengths and weaknesses are. Sometimes the results are really impressive and then something strange happens,” Mr Hogarth said.

“There is no point investing in technologies unless we understand the space. We are a bank: we have to get the answers right; we can’t deal with ambiguity. But we have built a backlog of suggestions and formal experiments, looking at efficiency and productivity gains.”

When asked about the impact on jobs, especially among technology staff, Mr Hogarth said ANZ’s “demand for engineering exceeds the supply, and we always need to do more”. “This is essentially an augmentation tool. The engineer still has to have oversight, and those who can work with this technology will be invaluable for the industry.”

ANZ said rolling out AI technology to customers to conduct self-service from a banking app could be around the corner. “We are training AI with internal data our staff can use when talking about staff problems, and then will look to expand it out to make teams more efficient,” he said.

“After that, you could get the confidence of using it in front of customers. I don’t think that is a long way away. I think this technology is already starting to have an impact across the industry.”