The next big wave of innovation

Over the last two years, the world has been transfixed by the advent of generally available AI systems.  

"The next wave of AI though will be the setting of goals for an AI system, granting it degrees of autonomy, and giving the system the ability to make decisions."

ChatGPT heralded a world of new applications, concepts, and scenarios where previously ‘dumb’ computing could be used with natural language with almost unnerving ‘human-ness’.  

Many of us have used AI at some point in the last two years to do tasks few humans ever experienced before 2023: we can draft messages or stories, search the web, engage in debates, research topics, generate art, summarise and analyse complex documents – all effortlessly using our first language.  

We can now talk to computers, and they talk back to us in our language, along with all the power and resources of the internet. 

Intensify the reach

The applications are still moving from experimental to scaled.  

At ANZ we use AI tools every day in software engineering; in addition, we increasingly use them in our teams to write emails or first drafts, to summarise lengthy documents and to do research.  

We are building solutions that help our teams use this in banking functions – from reviewing loan documentation through to helping answer customers’ questions online and in person.  

And we’re exploring other ideas to help our customers use AI to help them understand and manage their money. AI is a powerful ‘over-the-shoulder’ tool we can and do use to make ourselves, well, better. 

As the current work proceeds, we now look to the next wave.  

We’re still in control

You may have heard of Agentic AI – which, while a clumsy phrase, is going to dominate the next year.  

But what is it?  How does it differ from the chat bots we all use daily? 

The key word is agency.

Having agency means you have control – and today’s nascent AI tools don’t have a lot of control, they’re largely reactive.  

When you ask a chatbot something – or use a spellchecker, or autocorrect, or transcribe audio conversations, or write code, or summarise a document – the AI technologies you use are contingent upon you doing something.  

They’re like a car: enormously powerful but reliant upon the driver.  

As I can effortlessly propel a car to 100km/h with a tap of the accelerator, AI can summarise the world’s information – but only if I ask it to. 

Degrees of autonomy

The next wave of AI though will be the setting of goals for an AI system, granting it degrees of autonomy, and giving the system the ability to make decisions.  

This is markedly different from something like ChatGPT that has no goals, no autonomy and can only make recommendations, rather than any kind of decision. 

While that might sound like a scary leap, we’ve had autonomous computer systems that make decisions for decades.  

When you set cruise control on your car, you’re letting the car control the accelerator autonomously.  

In banking, every time you swipe your credit card, a computer system works out if the amount you’re spending is within acceptable limits and whether there’s any indication of fraud before deciding to either approve or reject the payment. Autonomous payment decisioning is essential for speedy checkouts.  

Autonomy is always constrained to a specific problem space.  

If you have a robotic vacuum cleaner, its scope is to clean a room without oversight – but that’s it. It can’t decide to paint a room, or climb stairs, or open a door. When you ask Siri to turn on the lights when you get home, the agency extends only to a few electrical switches:  it can’t set the lighting based on your mood or the weather outside. 

The autonomous computer systems of old have extremely narrow scope. While we rely upon them every day, most autonomous systems are pretty dumb. They can’t adapt, they can’t deal with ambiguity, and they can’t handle too much complexity.  

For example, early robotic vacuum cleaners could barely get themselves back to base without hiccups, and they couldn’t deal with electrical cords or other non-vacuumable mess without blindly creating havoc.  

An agentic system, however, can be given much broader parameters: “Clean my house, avoid the cords, use this schedule, don’t run when people are around, go over areas that are messy twice”. These systems can then make decisions to avoid spaces that are full of cables, or self-adjust if someone moves furniture, or if there are lots of people around.  

As we get more confident in the safety and reliability of AI, agentic AI will become more of a part of our lives. Cars that handle more than just cruise control; appliances that anticipate and manage challenges; truly useful personal assistants that understand all our vague requests. 

And this will come to banking.  

Sharing responsibility

As the first wave of assistive technology for searching, summarising, drafting and analysing documents becomes second nature, we’ll find ways to use agentic AIs to take on more responsibility. For example, every Monday it’ll action it to summarise my diary, clear any meetings with no agenda, send me the expense report exceptions and remind me of my overdue tasks.  

As we write more and more code with AI, we’ll have agentic AI run over our code base looking for errors, duplications and bugs and curate software before anyone reviews it.  

Our fraud control systems will get better and faster at spotting criminal behaviour and keeping our customers safe.  

And when our customers message us through our apps or online, we’ll get better at helping with more complex questions, lessening the frequency we’ll need to escalate to our customer service team. 

As with every single IT system in the bank – traditional, autonomous and agentic – everything gets checked, tested, proven, watched and monitored.  

Agentic AI at ANZ will still operate within tightly controlled ANZ-defined boundaries: everything we do is always anchored to our mission to shape a world where people and communities thrive. If it’s not safe, it doesn’t get used.  

Agentic AI – AI with some degree of agency – is the exciting next step for improving the sophistication and ease of banking.  

Tim Hogarth is Chief Technology Officer at ANZ