GPTreasury: A first dive into AI in Finance and Treasury

Introduction

“AI is the ‘new electricity’… just as electricity transformed many industries roughly one hundred years ago; AI will also now change every major industry”

– Andrew Ng (globally-recognised British-American computer scientist)

What is AI? Will it take your job? Will you be reporting to a chatbot this time next year? Will your credit analysis start coming from some AI in a server somewhere? Will the next Elon Musk be created in the labs of OpenAI? Stay tuned and find out (well, probably not, but stay tuned anyway!).

Ok, it’s time. You’ve been hearing GPT this, LLM that, maybe even some sprinklings of “Aaah the singularity is coming” for months now. It’s finally time to start trying to understand what it all means, how it’s going to affect the finance and treasury industries, and what it might mean in the future for all of you working in treasury or finance. Join me in this blog series where I’ll be exploring the brave new world of AI, trying to demystify it, give you some simple understandings of what the acronyms really mean, its inner-workings, and where it might be headed – all while keeping it (mostly) within the context of the Treasury and, more broadly, the Finance industries. In this blog, I’ll be laying the groundwork and giving the introduction needed for the series to follow. I’ll briefly explain what AI really is, give a simple explanation of how it works, and present some examples of how it might be, and is being, used in Treasury and Finance as well as a brief note on where it might be headed.

What is AI?

98% of institutions believe AI can bring improvements to how business is done, according to a Forrester survey, while 80% of banks are highly aware of its potential benefits according to Insider Intelligence’s “AI in Banking” report. So, what actually is it? I promise to keep this as non-technical as I can but I do think context for understanding is important. If you don’t, feel free to skip ahead!

Firstly, you have probably heard the terms Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) being thrown around quite interchangeably. So let’s briefly explain the differences. AI is an umbrella term for making computers mimic human-level intelligence, such as performing complex tasks and cognitive functions usually associated with humans. ML is a subfield of AI, or a technique for achieving it, which involves training algorithms to learn from data without explicit programming, for example regression or clustering. DL is a further specialised subset of ML which involves the use of artificial neural networks (ANNs), which are interconnected processing layers inspired by the human brain. For the purposes of  this blog series, I will be using AI to refer to the umbrella term, however it is worth keeping in mind that most modern AIs (especially the more recent generative AIs) are powered by deep learning, so you can think of it as equivalent here.

So now we have established that, you need to know what an artificial neural network (ANN) is and how it works. Do not worry, as I said, I will keep it non-technical. Simply put, an ANN is a conglomeration of artificial neurones, or nodes, using weighted connections between nodes, wherein weights are adjusted during training using gradient descent optimisation to reduce the gradient of loss with respect to a weight, where gradients are backpropagated through the network. Cool, let’s move on.

Don’t worry, that’s just a nerdy joke! It helps to start with a diagram of a neural network:

This is the general structure of a neural network; there is an input layer which takes in the AI inputs (in the above case, two inputs) – this might be words of a sentence, images that need classifying, or a time-series of data – and passes them on to each node in the first hidden layer. Each node then combines what was passed to it to create a single output, weighting them as defined by the model weights (which is what gets trained). How it does this combining is simply a pre-defined linear function of the node’s inputs as variables and their weights as coefficients. This output then gets passed to the nodes in the next layer. This happens in each layer. For each node. That’s a lot of numbers. Think of it as each node gets given multiple numbers and says “I need to combine these to make a single number” which gets fed to the next node along with the other nodes’ outputs. The training process is the process of each node considering “hmm if I change how I combine these numbers to be slightly more this way instead of that way, does it improve the result?”. The way it decides whether and how to change the combining of its inputs is quite complex, but essentially is done by measuring the effect of each weighting in the node on the overall model’s output, and thus changing them to reduce the error of the model. The combination of all nodes in all layers doing this is what creates the intelligent model, with layers and nodes capable of summarising and extracting important features in the data arising from the training process.

AI in Treasury and Finance

So, how can this technology be put to use in our industry? Well, let me start by saying that today’s potential uses may be eclipsed by those of tomorrow. The AI industry is advancing and progressing so fast that this blog will be outdated by next month. Having said that, what can AI be used for in our industry today? AI has been used in the industry for years, albeit behind the scenes, for its predictive capabilities and extracting useful information from data. The explosion of Large-Language Models (LLMs) in recent months and years (more on this in a later blog) has meant that we can now deploy AI in new intelligent ways (pun intended) to take over many of our text-, language-, or conversational-based tasks. This makes AI very powerful; the implementation of which could be the difference between success and being left behind. In fact, a 2020 survey from The Economist Intelligence Unit suggests that 77% of bankers agree, while a 2022 review by Insider Intelligence estimated the aggregate potential cost savings for banks to be $447 billion by 2023.

Without further ado, where can AI actually lend a helping hand today? One big potential for AI is powerful automation; not hardcoding rules and processes like traditional automation programs, but true just-give-it-a-goal automation. For example, imagine an AI trained on an entity’s operating procedures to read and interpret complicated compliance, legal, and regulatory information to provide succinct and simple goals to ensure compliance. Or how about automated financial research and portfolio management: having an LLM-supported AI read company prospectuses and information, search the internet for sentiment analysis, machine-learning-powered technical, fundamental, and quantitative analysis to find patterns, and insights from financial data and market trends for an automated full-stack and real-time portfolio AI manager. In fact, AI can bring remarkable automation to a vast range of repetitive, rule- and decision-based tasks: data entry and document processing for banks and financial institutions, or cash positioning, settlements and reconciliation for treasury for improved efficiency and reduced operational risk just to name a few.

Furthermore, as mentioned above, the recent innovation of LLMs in AI has enabled machines that are capable of understanding and generating language. This has big implications for user experience and customer service, allowing automated and fast customer response times while maintaining relatively high response quality and information accuracy – especially if the AI chatbot is trained on relevant and specific information. With 2022 Emplifi research finding that 86% of customers will leave a brand after only two poor customer experiences, the benefits of implementing LLM-powered, specifically trained AI chatbots could be vast, and maybe we will finally stop having to listen to terrible hold music for hours on end. These AI chatbots can be used as more than just a customer service chatbot; imagine having a 24/7 personalised financial advisor or a conversational onboarding / KYC bot. Removing the need for hardcoding questions and answers has opened up many possibilities. There is one big caveat here though: as much as one can try to train an AI to respond in predictable and desired ways, AI is still a probabilistic system which means one cannot guarantee what response it will give, and in a heavily regulated industry where giving incorrect information to a customer could mean legal issues, this is a big consideration that has to be made.
Lastly, AI can power data analytics to find patterns and anomalies in data which would otherwise require complex statistical analysis. This can be applied in many areas, from transaction and payment analysis for AML and fraud detection to analysis of network systems to detect anomalous and potentially malicious activity for enhanced cybersecurity.

Examples

I have given you a brief explanation of how AI works and conceptual examples for how they could be applied in our industry, but now you’re probably wondering whether there are any real-world examples of AI being implemented in Finance and Treasury today. Well there are. A lot. In fact, a 2021 McKinsey survey had 56% of respondents reporting AI usage in at least one function of their organisation. Here are some publicly-available examples:

Conclusion and the Future

AI has the power and capability to drastically change the landscape in all industries, but especially in Treasury and Finance where repetitive tasks, text interpretation, data analysis and general customer experience are all abundant and important. It is clear that those institutions that do not implement and adopt it, or are too slow to do so, will be left behind while new and exciting business opportunities open up. Hopefully you now have a general understanding of the powers of AI, and how it works.

There is one final note I want to make, and that is, where is this going in the future? This is a very difficult question to answer. As with anything, the future is unpredictable, but with AI this is extremely true. The AI landscape changes week by week, e.g. if someone had asked you this question two or three years ago – would you have ever predicted having the AI we have today?

Having said that, I think there are some comments one can make about potentially where it will go and the dangers it might bring. Firstly, I believe the next step in AI is multi-modal models – that is large foundational models that can interpret not just text, but images, audio and so on. Imagine instead of just LLMs, we have LAMs and LVMs (large audio / visual models), meaning foundational models that understand on a general level how to “hear” and how to “see” (much like LLMs generally understand how to “read” and “write”). This is already beginning to be the case, with GPT4 already being able to accept images. Secondly, there is definitely a serious conversation to be had about the risks of creating true general AI – but I’ll save this for a later blog. Lastly, I’ll leave you with a bit of a thought experiment.

If a lot of the financial industry is built around capitalising on inefficiencies, what happens if one day AI is so widespread in the industry that markets become perfectly efficient?

If you want to learn more, stay tuned for my next blog!

*P.S. They’re coming for the marketers as well – the header image in this blog was generated by DALL-E (OpenAI’s image model)!