To Build or Buy: Investment Banks Weigh Options for Large Language Models

The emergence of Large Language Models (LLMs) is starting to reshape traditional industries, and investment banks are no exception. The models’ remarkable ability to understand natural language, and process vast amounts of data, could revolutionize the way investment banks approach risk management, manage information flows, build their investment strategies, or even tailor hyper personalized account management.

Nevertheless, industry leaders are treading carefully - they know that information is akin to currency, and that feeding a Large Language Model with copious amounts of potentially sensitive data is not without consequences. Above all, bankers want to establish that models won't break rules (from ensuring data privacy and protection to maintaining ethical standards in data collection and usage) and create unacceptable levels of regulatory, and reputational risk.

While banks, like many other sectors, are upbeat about the prospects of integrating AI into their workflows, the environment they operate in presents unique constraints that warrants a prudent and thoughtful approach to large language models.

While banks, like many other sectors, are upbeat about the prospects of integrating AI into their workflows, the environment they operate in presents unique constraints that warrants a prudent and thoughtful approach to large language models. Investment banks are assessing various factors such as compliance with existing or prospective regulation and resources . Indeed while many investment banks could fund research, they still lack focus and specialized skills, and often a strategic approach. To fully understand how AI can disrupt traditional workflows, let’s first explore the potential of utilizing Large Language Models in investment banking, from streamlining operations, to improving decision-making and driving a competitive advantage.

How Large Language Models could be reshaping Investment Banking

And yet, despite the potential for significant progress, the investment banking industry faces several hurdles, with compliance being perhaps the most crucial of all. The challenge lies in the fact that, as is often the case, innovation is far outpacing regulators' ability to keep up. Although discussions are underway regarding how to manage this rapidly evolving field, a robust framework is still several months away. Investment banks cannot afford to wait that long, and therefore must navigate this uncharted territory autonomously. Fortunately, a few key initiatives, such as the EU Artificial Intelligence Act and the effort spearheaded by US Senator Chuck Schumer, offer some valuable guidance on the critical issues banks must keep in mind so that, when AI regulations finally come into force, they are better enabled to transition into the full observance of emerging regulation.

Both proposals include provisions on transparency, explainability, data protection, and human oversight , all of which are relevant to investment banks using large language models.

The challenge lies in the fact that, as is often the case, innovation is far outpacing regulators' ability to keep up. Although discussions are underway regarding how to manage this rapidly evolving field, a robust framework is still several months away. Investment banks cannot afford to wait that long, and therefore must navigate this uncharted territory autonomously.

Compliance considerations for investment banks using Large Language Models

The preexisting regulatory burden limits the type of data that can be fed into models, and constrains access to the outcome of its data processing. If a financial institution uses research to produce investment recommendations in any form, the link between the advice given and the source of the research should be traceable. More: depending on the jurisdiction, it might need to be materialized by a declared research relationship. Add a language model, and the process becomes meaningfully more complicated.

With one eye at the ongoing regulatory debate and another one at the competition, banks also face another dilemma when it comes to LLMs:

Buy or build?

Building a large language model from scratch is not a simple task. It requires a significant investment in time and resources, including access to large amounts of compliant financial data, computing power, and a team of experts to train and develop the model. Additionally, training the model correctly plays a pivotal role, requiring months of data preparations and overcoming numerous unforeseen issues with what Machine Learning experts are calling ETL (the process of extracting, transforming and loading the data into a model).

Given these challenges, many banks are asking themselves whether it is sensible to build their language models, or take advantage of the external resources at hand: from existing Large Language Models like GPT that can be trained in-house, to data partners that can fuel models with accurate, actionable external data and API solutions that are able to solve most pain points of in-house Machine Learning teams.

While building a LLM in-house is complex and time-consuming, it offers full control of the roadmap. At the same, using existing solutions saves resources and puts you ahead of the competition, but your model will suffer in terms of flexibility and customization. There’s no right or wrong answer, in the end it all comes down to prioritizing resources and goals.

As always, it pays to keep an eye on what the market leaders are doing.

The JPMorgan Chase approach

JPMorgan Chase, one of the largest and most prominent banking institutions in the world, has been at the forefront of adopting artificial intelligence (AI) technologies to enhance its operations and services.

The bank has allocated hundreds of millions of dollars annually towards AI initiatives and has brought in top talent from the tech industry, including renowned roboticist Manuela Veloso , who was appointed to head an in-house R&D lab modeled after some of the leading AI research organizations in the world.

One of the bank's primary goals has been to develop AI systems that can help streamline and automate various aspects of its business, from fraud detection and customer service to trading and risk management. To this end, JPMorgan Chase has been heavily investing in its data and cloud infrastructure, ensuring that it has the necessary resources to support machine learning algorithms. One of its flagship projects in this area is a software service called IndexGPT that uses AI to analyze and select securities based on customer requirements.

Given these challenges, many banks are asking themselves whether it is sensible to build their language models, or take advantage of the external resources at hand.

The Morgan Stanley approach

Morgan Stanley has partnered with OpenAI and got early access to their latest text generation AI, GPT-4, the latest and most advanced language model developed by OpenAI. The plan is to create an internal chatbot that will generate answers exclusively from Morgan Stanley's proprietary content library, allowing advisors to ask questions and receive answers with links to source documents, eliminating the need to search through PDF files hosted on internal websites.

The partnership between Morgan Stanley and OpenAI marks a significant shift in the way financial institutions are looking to streamline their operations by utilizing cutting-edge AI technology.

This approach is particularly note-worthy because, by securing early access to GPT-4, Morgan Stanley gained a distinct advantage over their competitors. Moreover, it provides a compelling example of how banks can develop their own Language AI applications without having to invest in expensive development projects. The language model can be adapted and fine-tuned to meet the specific needs of the bank, making it a more efficient and cost-effective solution than building a model from scratch, while insulating their sensitive data and complying with the very strict regulations in the industry.

Key takeaways:

• Banks adopting large language models is not a matter of if but when, notwithstanding compliance and resource challenges;
• While developing an in-house language model has its benefits, it is not always the most practical option;
• With no obvious deployment pathway, LLMs are ultimately a question of prioritizing resources and goals, with JP Morgan Chase and Morgan Stanley offering compelling examples of different avenues;
• Banks need to take a hard look at their resources and determine the best approach to leverage the power of large language models if they want to stay ahead of the competition and make the most of the AI revolution.