How generative AI is reshaping finance

Garreth Hanley:
This is INTHEBLACK, a leadership, strategy, and business podcast brought to you by CPA Australia.

Aidan Ormond:
Welcome to CPA Australia's INTHEBLACK podcast. I'm Aidan Ormond. Today we're exploring the evolving role of generative AI in insurance and finance with Patrick Leung. Patrick is a data science tech lead. His expertise spans natural language processing, text analytics and AI strategy, and he's currently focused on implementing generative AI to automate insurance workflows and advising executives on emerging AI trends.

Now, his career has taken him across industries from applying sports analytics to developing banking apps and shaping data strategies at Standard Charted as well as experience in startups. Today we'll discuss how Patrick and his team are using generative AI to transform insurance and finance. Welcome to INTHEBLACK, Patrick.

Patrick Leung:
Hello. Hello everyone. Honour to be here.

Aidan Ormond:
Look, it's great to have you along. So many questions I want to ask you. So why don't we just jump straight into it.

Patrick, there is a lot of talk, let's face it, about AI and we're hearing terms like traditional, generative and agentic AI, to name just a few. So can you first set the scene and briefly explain what generative AI is, what it isn't, and how it differs from traditional AI or machine learning?

Patrick Leung:
Yeah. So maybe I'll start with thinking from the shoes of the audience and think about what are the day-to-day tool that the audience would have been using and that's generative AI related. First thing coming to my mind would be the ChatGPT app, which is a chat board that you throw in whatever question you have, and it most likely will answer you almost like a human is answering it.

So that's a text-based generative AI. You probably have also looked at tools like Midjourney or Stability AI. Probably when you scroll on Instagram and you saw a highly realistic photo, but then it's generated by AI. So that's another line of image generation that is in the line of generative AI as well.

Aidan Ormond:
I mean, it's an exciting world, isn't it, Patrick, that we're currently living in. And you've touched on a few of the technologies that we're looking at and we're using. In your opinion, what are some of the most exciting or, say, impactful applications of generative AI across industries today?

Patrick Leung:
If I need to say one thing, I would say it's the AI agent right now. So if I take a step back, most people would think, okay, generative AI is just ChatGPT, which is a side tool that is fun to chat with, but it's not super useful. But what's really groundbreaking right now, it's the agentic AI. And what's agentic AI? It's a real tool that can bring job from end to end... from the start to the end. So it's a bit like the colleague that work with you, but it doesn't need your help in between.

Let me give you an example. Let's say I want to build a website for a podcast, and this website has to have recording and playing functionality. Traditionally, what you would do is you would write down this requirement and give you to a programmer. And the programmer will break it down and write the code and submit the website. Now, this Copilot thing, what it can do is to not only digest your requirement, but then it can build a strategy and to write the plan. It will write the plan like, oh, step 1, I'm going to build the UI. Step 2, I'm going to build a backend, and step 3, I'm going to do the test cases. So essentially, it just replaced the programmer from end to end.

So that's something exciting for the coding industry. Maybe another thing that's worth mentioning, it's on the creative industry. So recently, I saw a news that the film director, Jamie Cameron, was collaborating with Stability AI, which is the parent company that create Stable Diffusion.

So the reason behind this collaboration is for Jamie Cameron when he need to create one seconds of frame, it actually take 100 hours to render the image in a very high end resolution. But with a collaboration with Stable Diffusion, it actually turn 100s of hour into a few seconds. Yeah. And that's how powerful AI could be.

Aidan Ormond:
I mean, you just mentioned agentic AI. I guess for a lot of our listeners, they may not have those coding backgrounds, but also there are low-code or no-code platforms which you can build AI agents with. Is that not correct?

Patrick Leung:
Yeah. I think nowadays, it's really going to the no code environment and really make us tech people really, really nervous because coding became not the core skill set anymore, but how to describe the requirement, how to assess if the code is good quality became more important.

Aidan Ormond:
So that really opens up AI to people who don't have a coding background, which I think is a real positive, isn't it? I've seen lots of platforms that have the ability for you to build agents, and teams of agents even, underneath an agent, without having a huge amount of coding, or in fact no coding background at all. So that, I think, brings a wider audience. Does it not?

Patrick Leung:
Yeah, definitely. And I think we're going to touch on it later, but I think our audience, maybe some of you guys are non-technical, but one thing I want to emphasise is AI literacy. So don't be too afraid of AI, but now is the time to really learn about it because the barrier of entry is really low right now.

Aidan Ormond:
I couldn't agree with you any more. I think everyone should be studying AI, but that's just me. And I guess that's everyone on this show today. But a lot of our listeners are actually in finance, in accounting. So Patrick, how is generative AI transforming traditional financial analysis practises, and what are some real world examples of this shift?

Patrick Leung:
Yeah. So interesting question. And maybe I'll break it down a bit, because when I think about financial industry, it is really wide. It's go from maybe predictive modelling to business advisory to auditing or even to workflow automation. So maybe I can start with something I'm more familiar with, which is a predictive model. So when you talk about predictive model, you talk about the few things.

First, you collect the data, you build a model, and then you look at the result and try to communicate it. I think generative AI helps in every steps. So the first step in collecting data, what's super interesting right now it's, previously, when you do data collection is actually very hard. You probably might not have access to all the data.

The data can be unclean, the data might contain sensitive information, and then you cannot build your model. But right now, with generative AI, the coolest thing right now is synthetic data. So it don't really need real data to train your model, but then it create synthetic data that remove all the sensitive information for you to train it.

So that's one thing that's very fancy. And once you get the data, the model development part is also very simple. So we have a GitHub Copilot, and we have a lot of coding assistant that almost you don't need to code anymore. You just describe it in natural language and it's going to do all the hard work for you. And even at the end, for reporting and for summarising the insight for the executive, that's exactly what generative AI are good at.

Aidan Ormond:
You actually mentioned the word synthetic before, Patrick, and that's an interesting word, because synthetic media is something that's really taking off. For example, platforms like Synthesia and similar platforms where you can actually text to video, produce videos. What are your thoughts on that aspect of AI, synthetic media for example?

Patrick Leung:
First of all, I think a lot of creative people are worried that their job might be taken, because now to make a video or to make a painting is just a prompt or a sentence. I think right now, I'm not too worried. I have friends working in the creative industry and they said, "Well, it's worrying, but then the quality of the end product is not that high yet." So I think we're not at a point to replace creative people, but on the other hand, I see it create another value, which is all the image classifier.

For example, let's say, in medical industry, one thing they struggle is to get high quality photo for a negative and positive sample. So imagine I need to get 100 photo of X-ray with cancer, it is really hard to get, because of all the privacy issue, how scattered the data are across different hospital. Now, what you can do is to use generative AI to generate cancer related X-ray. And then you can easily train your cancer detection model. So it's something really cool these days.

Aidan Ormond:
Yeah. That is so cool, and so, so valuable in this world. You mentioned GitHub CoPilot earlier, Patrick, and I'm interested to know, there's also a Bloomberg GPT software out there. Can you explain a little bit more about that?

Patrick Leung:
Yeah. Well, I think it's a model that released 2023, and I think it's one of the first one that get released in the financial industry. I think it's interesting because Bloomberg is widely used by investment professional.

And from what I understand, it's like a ChatGPT embedded into the Bloomberg terminal, and then it can do things like summarise, breaking news, extract key event, and do some service risk factors, analysis for portfolio managers. So I think it's really an example of tool that blend into the day-to-day workflow and to help tiny bit improve efficiency on certain part of your day-to-day work.

Aidan Ormond:
And I guess that speaks to the possibility of many different businesses and organisations building their own proprietary software in this space?

Patrick Leung:
Yes. So I think it's a super interesting topic as well. What's different for the BloombergGPT and the ProperGPT is just specialisation in industry. So imagine, ChatGPT is something generic. So it's not specific to one industry. But what the company are right now doing is to do context-specific model, fine-tune it. So it tailor to the industry or even tailor to a specific company.

Aidan Ormond:
That sort of feels like where that space is moving. I mean, that is a good segue into the next question, which is what are the main challenges that businesses face when adopting AI solutions such as generative AI at scale? I mean, what do you think are the main challenges?

Patrick Leung:
Again, super interesting question, and it's a very real question. I think a lot of company wants to step into AI, but it really link to how mature they are technically. So imagine you have, in the extreme case, a company that still have all the data written in paper. Then it's impossible for them to tap into this field, because gen AI need data to flow. You need to fit in company data to gen AI for gen AI to make decision for you. It also link to the infrastructure and organisation structure.

So if your company don't really have a good IT team, it's going to be very hard for you to connect your data to the outside world in secure and reliable fashion. And I think for gen AI to truly work, the entry point is also important. Is your gen AI a customer-facing tool? If yes, then you need an entry point for customer to interact with your gen AI. Or if it's an internal-facing tool, then you again need an entry point for your internal staff to interact with your gen AI.

Aidan Ormond:
And of course you do need talent as well. You need people who are ML and engineers, for example. And you work with a lot of staff. I mean, what's the situation with talent at the moment?

Patrick Leung:
Well, AI talent are always expensive and I think... Well, I think it's an interesting situation right now. There's tonnes of entry-level talent, but then a lot of them are lacking experience. And right now, I think the hottest position are the middle-level AI talent. But then I think, globally, there's a shortage ML engineer, AI or data scientist.

I think what's interesting though is there a new job title that is appearing. For example, prompt engineering, it's something new that you need specialised people to write prompt for you. And also, I think data annotator is another interesting example, because you need human in the loop to verify if the gen AI is making the correct decisions.

Pre-Ad Announcement:
We'll be right back with more INTHEBLACK after this short message.

CPA Australia Ad:
Learn how to use data to make good business decisions by enrolling in CPA Australia’s ‘Data-informed decision making’ short course. You’ll learn how to apply data-analytics to real business problems, along with how to maintain data integrity, navigate ethical dilemmas, and manage issues such as data quality. This interactive course uses videos, activities and simulations to help you learn. For more information, find the link in the show notes.

Post-Ad Announcement:
And now for more INTHEBLACK

Aidan Ormond:
You just mentioned there, Patrick, prompt engineering, and maybe some of our listeners wouldn't understand that. I know a lot of us reverse engineer a lot of our prompts to get the best possible data in, the best possible data out. Just talk to us in broad terms about prompt engineering.

Patrick Leung:
So if you interact with ChatGPT, you'll realise if you change your command slightly... Let's say, you change your wording from a very simplistic sentence to a more descriptive sentence. For example, let's say, I want to generate a logo for CPA. You can say, generate a logo for CPA, or you can describe it in detail like, generate it in yellow in certain format for what purpose. Then it will... going to give you better results. So this process is something a prompt engineer would be very good at.

Aidan Ormond:
So I guess what you're saying is good data in means good data out.

Patrick Leung:
Yeah, definitely. Always the same for garbage in garbage out.

Aidan Ormond:
Yeah, exactly. Exactly. And I guess you and I could have a whole podcast on prompt engineering and how you can place some of these AI tools as experts to give you a reversed engineered plan or strategy. There are so many different ways that you can utilise the power of prompt engineering.

Patrick Leung:
Yeah. Actually, if I go a bit technical, there are now fancy way to do automatic prompt engineering, that is with a human in the loop to say, "I like this photo. I don't like this photo." There are AI to judge the AI output and to improve the AI. Yeah. It's something in the field of recursive self-improvement, so it's something quite scary, actually.

Aidan Ormond:
Well, if you are saying it's scary, Patrick, it must be scary, because you know more than anyone here. But fair enough. I guess a lot of our listeners also will be listening to this podcast and thinking to themselves, "Well, what new technical skills do I need to learn to stay relevant in finance?"

They may be finance professionals, they may be younger, they may be in the mid-point of their career or even at the later-point of their career. So what do you think are the technical skills that these people should be looking at?

Patrick Leung:
Yeah. I would start off with saying, I think a minimal understanding in coding and data, it's essential, because I think we just talk about... The entry barrier right now is very low. You don't really need to learn coding anymore. But that said, you still need to be able to understand roughly how coding work and how to execute a code, and also understand the basic concept of input and output and where to run your model.

Yeah. So I think this is one of the essential skills. The other one, which is quite interesting, I think a lot of people from now on will transform from an individual contributor to a manager. And the reason of that it's because of AI agent. So I would imagine in a future, and that feature won't be that long into the future, each of us will be managing a couple of AI agent. So we will be responsible for the output of this AI, and we are responsible to guiding this AI to do the right thing. Yeah. Interestingly is management skill would be something everyone need to learn.

Aidan Ormond:
That is interesting. And what you're saying essentially is that an AI agent or a team of agents become your teammates at work alongside your human teammates. So it's like a way to manage your own teams. Some are AI teams or AI agent teams and some are human teams.

Patrick Leung:
Yeah, definitely. And I think the reason is, as you might have observed, ChatGPT or gen AI will make mistakes. So they're not 100% foolproof and they need human in the loop to guide them to do the right thing. And I would imagine that something a future professional will be doing, they will no longer be doing the groundwork. Let's say, they won't be building a model, but then they will be guiding the AI agent. What does a good model means, and what does a good report means?

Aidan Ormond:
Yeah. And for example, ChatGPT says, "When you put a prompt in, this could make mistakes." I mean, it is actually there in front of you, isn't it? So, it is almost like a disclaimer. So I think what might be important is to always check whatever the data you get back, as best as you can, because it should not be taken as read, I guess, that it's completely correct. Hallucinations is an example. Could you explain what a hallucination is?

Patrick Leung:
Yeah. Exactly. Yeah. Well, hallucination, it's... How should I put it? So ChatGPT sometimes will come up with a sentence that looks like fact, but then they completely made that up. For example, I can ask for, you know, "Give me a paper that talks about generative AI in Australia." Well, they can easily come up with a paper that is not existing, but then it looks super realistics. So these are one example of hallucination.

Aidan Ormond:
That's a word that I think people will learn about as they learn more about AI. But Patrick, I have to ask you. One question I get asked a lot is ethics and privacy within AI or the parameters of AI. And it's crucial in finance. Let's face it. In your opinion, how do organisations ensure responsible and ethical use of generative AI with sensitive financial data?

Patrick Leung:
I think that's a really serious question, and I think there are a few angle to look into it. First thing about the input data, I think organisations really need to start classifying their data into sensitive, non-sensitive or public data set, for example, and have different measure. For example, if you put in sensitive data into generative AI, and if that became part of the training set, what would happen is the generative AI could disclose customer secret in the next chat.

For example, if they have seen people's password in their training data, maybe the next user can say, "Oh, what is the password of Aidan Google account?" And then the generative AI might just spit it out. So that's one thing that is pretty sensitive. And secondly, it would be controlling what the generative AI can say and cannot say. So right now, the industry have some hot keyword called a guardrail. So meaning, a set of rule that the generative AI has to follow.

Aidan Ormond:
Exactly. I mean, is there anything that people can do in terms of their settings when they're using AI tools to ensure that the security is at least a little bit stronger or more robust?

Patrick Leung:
Yeah. So I don't think it's a solved problem, but there are many company that tried to tackle this. Well, the current practise is they do a lot of filtering. So every time a generative AI output an answer, actually, before it goes in front of the user, there will be another AI as judge to say, "Is this appropriate? Is it sensitive?" And then it's going to filter out the inappropriate response.

Aidan Ormond:
Look, it's an amazing time to be alive, I think, in this world in terms of AI. We are recording this in June 2025. So what are the top three trends that you're seeing, Patrick, with GenAI, and what do accounting and finance professionals need to do to prepare? And what is the best way for our listeners to stay across trends when it comes to AI?

Patrick Leung:
Well, I think really, it really links to AI literacy. So try to get up to date with the news. Don't turn off when you hear technical things. Try to understand what's going on. Recently, I've seen, I think the CEO of Anthropic saying AI is going to wipe out, I don't know, 20% of the white collar job. And a lot of my friend thinks it's just hype.

But for me, as a professional who works for 10 years in the AI industry, I've starting to see this become a reality, because very simple workflow can be automated right now. Yeah. People either became managers of AI or you get eliminated. I think that would be the picture.

Aidan Ormond:
So what you're saying is that managers who deploy AI to make their jobs more productive will be more in demand than managers who don't.

Patrick Leung:
Yes, exactly.

Aidan Ormond:
What a world we live in. Just to finish off, Patrick, I mean in 10 years time, we're talking about key trends here, but in 10 years time, what do you think the work world will look like in terms of AI and how we do our work?

Patrick Leung:
Yeah. That's a very broad topic. I'll try to answer that. If you think about it, there are few limitations with AI. I think one is energy. So mostly, the cost with AI is to energy, the electricity that need to run AI instance. So one limitation is the electricity that need to run it. And then maybe the other thing that I can think of it's, think about area that AI can do recursive self-improvement.

To give you an example, the coding world, why a lot of programmers are being replaced right now is because coding itself is self-contained. So you can generate unlimited scenario to rerun your code, rerun your code, write another code, and then compare the result.

And then the model can self-improve. But for scenario, like a financial market, because you cannot generate a new client, so it's not a scenario that the model can keep iterate and keep creating. That's why I think financial industry is relatively safer.

Aidan Ormond:
Well, that will make a lot of our listeners very happy to hear that, Patrick. And I'm pretty sure you and I could chat about this all day long, but I think we're going to have to call a close to this particular episode. But thank you so much for sharing these really compelling insights. And for our listeners eager to learn more, please check out the show notes for links to additional resources from CPA Australia.

And don't forget to subscribe to INTHEBLACK and share this episode with your colleagues and friends in the business community. Until next time, thanks for listening, and thank you, Patrick.

Patrick Leung:
Thank you.

Garreth Hanley:
To find out more about our other podcasts, check out the show notes for this episode, and we hope you can join us again next time for another episode of INTHEBLACK.