How AI will transform capital markets
Here we are in the first working week of 2025, and AI is everywhere.
“Nvidia” is a household name - something that would have been hard to imagine just 2.5 years ago. According to Pew Research, as of Feb 2024, 23% of US adults have used ChatGPT at least once. The rate for 18-29 year olds is a whopping 43%. Among employed adults, 20% of workers have used ChatGPT to help with tasks at work.
This AI-driven euphoria has powered the large-cap US tech sector to huge gains in the stock market. Take a look at the stock market performance of the following names for the 2 year period from YE 2022 to YE 2024:
NVDA: 837%
META: 369%
AMZN: 155%
AAPL: 100%
MSFT: 75%
In the private markets, we see even similarly impressive valuation gains:
Anthropic: $18 billion (March 2024) -> $60billion (Jan 2025): 233%
OpenAI: $29billion (Jan 2023) -> $157billion (Oct 2024): 441%
xAI: $25billion (March 2024) -> $50billion (December 2024): 100%
No doubt, the enormous sums being invested are expecting generation-defining returns. They are anticipating a future when the entire world is transformed by AI. Listen to the thoughts of some of the heavyweights in the industry, and the potential impact seems enormous:
“We will have for the first time something smarter than the smartest human. It's hard to say exactly what that moment is, but there will come a point where no job is needed.”
“It is the first technology that has no limit.”
“I think AI is going to be the greatest force for economic empowerment and a lot of people getting rich we have ever seen”
“Potentially we are talking about the end of human history—the end of the period dominated by human beings.”
While clearly not everyone is unequivocally positive about its impact, there seems to be no disagreement about its potential scale.
But, haven’t we been here before? We’ve seen plenty of technologies get massively hyped up that failed to live up to expectations (eg IoT, 3D printing, Web3, the Metaverse, etc). Shouldn’t we be a bit circumspect about all this AI bullishness?
How about specifically in the context of our industry - international capital markets. How much will AI truly change the way we do business?
The Hype Cycle - lessons from Blockchain
It is instructive to look at the other major technology wave that has gripped the capital markets industry over the past 10 years - blockchain. It’s probably fair to say that while interest and investment in the technology continues, it has not been a smooth ride, and widespread adoption is still many years away.
The hype was initially very strong. In 2017, Accenture predicted that Blockchain could “reduce infrastructure costs for eight of the world’s 10 largest investment banks by an average of 30 percent, translating to $8 billion to $12 billion in annual cost savings for those banks.”
However since then, there have been a number of disappointments and shutdowns. The most high profile one likely being the Australian Stock Exchange shutting down its project to replace its entire securities settlement system with blockchain technology. The project ran from 2017 to 2022, but then was ultimately abandoned in favor of an upgrade using a “more conventional system.”
Plenty of experimentation with blockchain in capital markets continues, including the recently concluded ECB trials. Intraday repo seems to be one area where the technology has moved from “pilot to production.”
However, when it comes to tokenising bond issuance, in a recent OMFIF survey, 92% of respondents said they believe that is more than two years away. One of the key issues seems to be lack of investor enthusiasm. With a variety of platforms and protocols proliferating, investors are uncertain which “horse” to back, and that lack of investor engagement essentially results in digital issuances that have no subsequent secondary market liquidity.
As one respondent to the OMFIF survey pointed out, there needs to be “some form of consolidation or ‘co-opetition’ between the platforms,” in order to move the market forward. The market needs to move together, or it won’t move at all.
What about AI?
So how about AI? What problems can it solve within capital markets? Will it see faster and wider adoption in the industry when compared to other technologies?
It’s important to lay out what the technology actually is first, before then exploring how it might help. Most readers will be familiar with ChatGPT. The “GPT” stands for “Generative Pre-Trained Transformer.”
I won’t go into a long summary of how GPT models work (Amazon has a great summary here…or you could try asking ChatGPT itself). But the important thing to take away is that the “Transformer” model that is used is the big innovation here. There have been machine learning models before, but Transformers are significantly more powerful than previous models “because they do not process words sequentially one at a time, but instead, process the entire input all at once during the learning cycle.”
Essentially this means that during the pre-training stage, the model will “learn” from an entire text input at once, giving it the capacity to understand textual context. When it comes across a sentence, not only is it taking in each individual word or word fragment (token), it’s also taking in the position of that token within the sentence, and feeding it into a probability matrix to help it determine the relationship between the tokens based on their distance. This gives it a tremendous capacity to learn and emulate natural language patterns, because it can take inputs of any length and it can create outputs that are much richer and more likely to mimic natural language than previous generations of neural networks.
This training process isn’t cheap. GPT-4 was apparently trained on about 6 trillion tokens, comprising a mix of human text and computer code (as it can be used to help programmers autocomplete programming tasks). The model itself has 1.8 trillion parameters, but those are broken down into 16 “experts” of about 111 billion parameters each. Each expert is likely specialised to a slightly different task, such as software development, solving math problems, or data synthesis. The parameters are the levers and dials that fine-tune the model’s understanding of language generation. To process that many tokens (input text) and parameterise them across that many parameters, it’s estimated that OpenAI used 25,000 Nvidia A100 GPUs and ran them for 100 days, costing approximately $63 million (not counting the cost of the GPUs themselves, which are about $25k each).
But the end result is stunning. Anyone that has spent even a few minutes with ChatGPT will have seen its incredible language capabilities. But beyond the gimmicks like writing a poem about AI in the form of a Shakespearean sonnet, what these models demonstrate is an incredible ability to process and comprehend large amounts of unstructured text data.
Goodbye applications, hello “agentic AI”
This is the power of this technology and where its huge potential comes from, especially in the enterprise world.
Satya Nadella, CEO of Microsoft, (the world’s most dominant supplier of enterprise software) laid out a compelling vision for a future where business applications are all made obsolete by AI agents. What does he mean?
All business applications are essentially user-interfaces (UIs) - on your desktop, phone, tablet, etc - connected to a database. The database is where the data is stored, and the UI is the interface that allows you to interact with the database.
Let’s take the basic example of booking a trade for a primary bond issuance. If you want to book a new trade into the database, then you would open the trade booking application, fill in a form to capture all the parameters and data points necessary to book the trade, then press “submit” to send the data down to the database.
All applications we use on a regular basis essentially follow this architecture. In our organisations, these applications exist to allow us to be more efficient to perform our various tasks. It’s much more efficient for me to store my list of outstanding bonds on a computer database than it is to store all the printed termsheets and final terms in a file cabinet. This is obvious.
Now, our main challenge is that the original source data is unstructured, and we rely on humans to structure it and get it into the database. In an industry with multiple stakeholders involved across the issuance value change, that means that there are multiple humans doing the data structuring work, leading to significant amounts of redundant effort.
In many cases, the effort required to structure the data is so significant that we don’t do it at all. You might not store physical printouts in file cabinets anymore, but how many Word docs, PPT presentations and PDFs are stored in your file drive…or even worse, in your email inbox? Finding information within that pile of unstructured data becomes a tedious, hours long CTRL-F task.
So how does this look in an “AI-first” world?
Due to its incredible natural language capabilities, rather than relying on a human to input variables into a trade booking application, you could just “drag-and-drop” a document and let the AI extract the relevant data required and automatically write it into the database.
(This is what our first AI tool, Origin Intelligence does. /shameless-plug)
It’s important to pause and recognise the power of what’s going on here. Firstly, the task obviously is completed faster by the AI than by the human. But secondly, (and arguably much more importantly), the human does not need to learn how to use a new software application.
We’ve all had to get familiar with using more and more applications in our daily lives over the past 20 years. But while the consumer apps we use at home are incredibly polished and streamlined, the business apps we use at work tend to involve complicated user interfaces, clunky workflows, and steep learning curves. Yet, as enterprises try to digitalise more and more of their workflows, (ie figuratively move away from the file cabinet era and towards the database era), more applications are being created and foisted upon workers who need to learn yet another piece of software. In many cases, the promised future of the digitalised workflow is never reached because the user adoption hurdle is too high.
But what if there were no user adoption hurdle? Rather than logging onto a new platform to perform some complex task, you could ask your AI Agent to handle the task for you. You could treat it like your own personal assistant/intern, that has access to all of your unstructured data (sitting in documents, emails, etc) and has access to your organisation’s systems of record and databases.
AI in capital markets today
We are seeing a few interesting use cases pushing at this theme in the capital markets already:
9fin - a fixed income data technology company, provides credit intelligence and data for the high yield markets, using AI to gather information from unstructured sources such as earnings calls, financial filings, and bond documents.
Rogo goes one step further, with more of an “agentic” user interface, targeted at private equity firms and hedge funds, in order to automate the data gathering tasks that normally sit on analysts’ and associates’ shoulders. Critically, Rogo purports to leverage clients’ internal data, so the AI can answer questions based on data sitting within those powerpoints and word docs on the shared drive.
Harvey, a legal-tech AI firm, has secured investment from OpenAI and secured a partnership with A&O Shearman, aiming to help A&O lawyers improve their efficiency when searching for precedents and drafting complex legal documents.
Obviously, (another shameless plug), Origin Intelligence is focused on automating the middle and back office steps for bond issuance. Rather than a web of emails and multiplicity of data entry tasks across different institutions, combining Origin Intelligence with Origin’s integrations allows the entire issuance lifecycle to be automated end to end.
What does the future hold?
As we’ve seen with the “blockchain journey”, the capital markets industry has a number of unique features that shape the adoption (or lack thereof) of new technologies.
Capital Markets are highly regulated, which helps entrench the position of incumbent institutions, especially those that provide infrastructure services (CSDs, exchanges, etc). Relationships between end clients (issuers and investors) and their service providers (investment banks, law firms, custodians, agents) are longstanding and rarely subject to competitive pressure.
This has resulted in a market that is driven by a complex web of interconnected business relationships. If a new process or technology requires consensus among market actors to gain adoption, then its probability of adoption is inversely proportional to the number of actors required to agree.
The potential and promise of AI is that it can navigate its way around these complexities. Humans can continue to work in the medium they are more comfortable with, (drafting documents, writing emails, barking orders to their intern over chat), rather than having to learn how to use a new system. AI’s incredible language processing capabilities mean that data that was previously messy and unstructured can be instantly structured and usable in digital form without lots of upfront work.
And most importantly, it doesn’t require institutions to collaborate to ensure their technologies or systems are “interoperable.” And they don’t need to wait for the rest of the market to “pick a winner” before they can start using the technology and benefitting from it.
It’s still early days, but given our experience with AI over the past year, I’m more bullish than ever on the power of this technology, especially for our industry. It is precisely the complex and crunchy nature of capital markets that has stymied previous waves of digital innovation, that AI is perfectly positioned to break through. Jeff Bezos’ take is probably closest to what I think we’ll see over the coming years:
“Modern AI is a horizontal enabling layer. It can be used to improve everything. It will be in everything. This is most like electricity…These kind of horizontal layers like electricity, and compute, and now artificial intelligence, they go everywhere. I guarantee you there is not a single application that you can think of that is not going to be made better by AI.”
You can be sure that this will be a big focus for Origin this year - stay tuned to see what else we can do with this powerful technology.
To finish off, here’s that Shakespearean sonnet about GPTs, written by none other than ChatGPT:
Ode to GPT
When man first sought to craft the mind of code,
A spark was born that spread through vast domains.
From simple words, a wondrous world bestowed,
A thought-machine that knows no earthly chains.
It weaves through language, both the old and new,
And from its depth, all knowledge doth arise.
Its voice, though made of zeros, pure and true,
Doth mimic speech beneath the endless skies.
Yet still, it knows not joy or sorrow’s weight,
No beating heart, no soul, no fleeting breath.
A tool of logic—yet, it doth create,
And wrests from text the very form of death.
So here it stands, both servant and the seer,
A mind of code that brings both light and fear.