The early embrace of emerging artificial intelligence (AI) technologies within New Zealand’s financial services is an exciting yet uncharted domain. While the influx of these cutting-edge technologies promises huge benefits for all industries that adopt them, any AI deployment – and increasingly those that operate with minimal human intervention – must be undertaken with a clear understanding of the ethical risks, potential biases, and impact on employees.

FST Media recently delved into the implications of the transition towards ‘agentic AI’ – or autonomous, machine-driven decision-making – systems for financial services, featuring insights from Stephen Clay, Westpac New Zealand’s Head of Technology Strategy and Architecture.

For nearly a quarter of a century, Clay has provided his expertise as a senior technologist to drive transformation initiatives within New Zealand’s six largest banks, as well as global financial services giants UBS, ABN AMRO, and AXA. Today, he oversees architecture and overall tech strategy for one Westpac Group’s most dynamic units.

FST Media: Generative AI is a burgeoning topic within financial services and the wider industry, with the technology promising a fundamental transformation of how businesses and the humans within them operate.

What do you rate as the most significant opportunities associated with GenAI?

Clay: The way Generative AI (or GenAI) is developing is fascinating. We’re in a genuinely historically important period with new capabilities from central players like OpenAI and Anthropic iterating almost monthly. Logically, then, the most significant opportunities we’ll see must leverage breakthroughs still to come, with a high probability of being markedly more advanced than what we have currently.

GenAI, however, differs deeply from any other recent hyped tech movement like blockchain or cloud – those were relatively straightforward to understand and therefore intuit the logical implications and likely future trajectories of.

For GenAI, while there is a steady march of amazing improvements, the big players we all depend on don’t have an agreed goal or endgame, or even agreement on what is technically possible. No one has properly understood exactly how large language models (LLMs) work for some time, if ever, and for a number of reasons those players are now focusing purely on LLMs at the expense of pursuing other techniques, meaning no progress on lines of research that could potentially be free from whatever inherent limitations LLMs – as the way to do GenAI currently – may be subject to.

Only a few weeks ago one of the leading GenAI minds, Ilya Sutskever (ex-OpenAI), founded Safe Superintelligence Inc (SSI), a company with a single ‘straight shot’ goal of human-like superintelligence and no plans to market any products before it gets there – and this while no one actually knows whether such an advance is even possible! If he does succeed, then this could be the greatest opportunity and impact not just in technology but in human society for a very long time.

Right now, GenAI is extremely good at working on specific problem sets it has been deeply trained on, and has a very compelling anthropomorphic engagement style. This can give some misleading impressions of its actual overall broader ability.

It’s important to remember that, in its current incarnation, [a GenAI system] doesn’t ‘know’ anything and is still subject to hallucination or the like.

If SSI succeeds, that move to a ‘general intelligence’ will change things completely, especially teamed with the recent trend of the ‘agentic’ paradigm. And if they fail, the breakthroughs along the way will probably be astounding regardless. So, watch and see what the shiny new tech evolves to, in a context of truly exploratory advances being made at breakneck pace!

FST Media: What exemplary AI use cases are you aware of or predict will emerge in the near future?

Clay: We’re starting to see some solid use cases both newly emerge and continue to evolve. Making a distinction between more traditional ‘machine learning’ AI and newer ‘generative’ AI, there are interesting real-world moves across both these fronts.

Firstly, for machine learning (ML) the original proposition is still where the value lies: leverage existing data to find insights across newer data, ideally to discover relationships we couldn’t otherwise spot, or otherwise benefit in terms of efficiency etc.?

We’re starting to see more full products make use of ML within ‘blackbox’ offerings for uses like fraud detection; this means we get the benefit of the models without having to do the heavy lifting ourselves.

As our own organisations move to implement data products or otherwise better manage our data, we’ll start to see benefits for our own data science and ML use. Applying GenAI techniques to improve ML is also interesting, noting a need to balance improved outcomes against potential higher costs for compute – how much more would you pay, say, for an improved version of something that’s already most of the way there, albeit needing a little human correction?

For GenAI, I suspect most of the solid use cases are getting well-known. As I noted before, things are still developing and we’re sure to see advances lead to some uses we’ve not considered yet. In many cases, things are still too immature not to have a human ‘in the loop’, checking that generated results are accurate and not subject to hallucination, especially before anything goes near a customer or important decisions are made.

Productivity uses for Chat GPT, Claude and the like are delivering real benefits beyond just hype: summarising documents or emails, scaffolding of reports or presentations, initial brainstorming of points and so on. For developers, code generation is definitely taking off.

In large organisations like banks, so much of our business process fits what the emerging agentic paradigm naturally supports, so as agentic approaches mature, we should start to see what was more traditionally the remit of workflow and RPA benefit from GenAI.

In terms of how the current flood of breakthrough capabilities is being used, I’m even starting to hear people in the industry talk about delaying confirming technical designs until as late as possible, to take advantage of any improvements over even a matter of months. Exciting times indeed!

FST Media: The implementation of AI requires careful ethical consideration. What can we do to ensure AI serves as a positive force within society, and what measures can be taken to reduce the risks involved?

Clay: The ethical aspects of AI, especially GenAI, feel like a real Pandora’s box. Doing this right, and fairly, for all the different stakeholders involved definitely won’t be a question we answer once and ‘set and forget’ the solution for: we’ll need to keep some attention on this as a wider tech industry – and likely specifically within financial services – for a good while yet.

For most of the issues we have identified to date, and likely for many we haven’t, I think we should be able to make good progress from a mix of (a) proactively driving awareness of the main types of issue broadly across all involved parties, and (b), for any specific AI initiative, ensuring we are paying intentional attention to the possible issues and how we can mitigate them – perhaps specifically calling out ethical points in formal requirements, and potentially in definitions of done etc., if applicable.

The types of issues we need to consider will evolve as our use of AI evolves, but most should be variations on basic themes. Examples include the lack of representation and fairness within training datasets. For instance, the potential for women or minority applicants to tech roles being declined by AI hiring platforms because they were unfairly underrepresented in the training sets.

Similarly, unfair outcomes could eventuate from skewed customer data if, say, they are used to train a lending decisioning platform. To be honest, this sounds like an issue where attention and intentionality should get us most of the way to addressing many cases – the tech won’t be the limiting factor, but what we do with it will.

A second issue operates at a more societal level and concerns the overall impact of (especially of generative) AI on the workforce, namely the displacement of jobs by AI which has been widely discussed (e.g. the recent writers’ and actors’ strikes), and as something a little more nuanced, the idea that AI will reduce the opportunity for knowledge workers early in their careers to do the ‘grunt work’ that contributes to the deep, internalised understanding of their field, be it law or accounting or coding. This raises the question of how the next generation of experts will learn their crafts.

Across these, I suspect most of what we can do now is again drive that wide awareness and start having conversations now to spark potential answers while we can.

At the industry offering level, Anthropic’s Constitutional AI approach is very encouraging to see and we can hopefully look forward to a future where a selection of platforms includes assessing their capabilities to play ethically or safely with techniques like this. ■

Stephen Clay featured as a panellist at this year’s Future of Financial Services New Zealand conference.