AI vs. Humans: What is the Future of Investment Banking?

When the personal computer became prevalent in the 90s and early 2000s with the advent of the internet, many low-skilled, low-paid jobs such as data entry, telephone switchboard, and basic bookkeeping were wiped out. But technological revolutions haven't always impacted jobs at the bottom of the ladder as the more recent example implies. The Luddites in Britain 200 years ago were skilled workers being wiped out by lower paid operators of new textile machines developed during the industrial revolution. The machine-smashing was necessitated by fears that they would end out of jobs and thus battled for skilled workers to be hired to handle the new machines.

ChatGPT was released just about two years ago to kick of the consumer phase of the generative AI technology wave. It's still early days but it's clear that this wave will be more industrial revolution than computer revolution in terms of types of jobs that will be impacted the most.

Software development, creative writing, image and video creation are all areas where several large language models excel. The biggest irony is silicon valley potentially decimating their own with AI tools they have developed. There is seemingly a race to the bottom with all sorts of dedicated code-writing tools being developed, including code agents that can develop software from beginning to end without human intervention.

What about investment banking? This is one sub-industry that remained relatively unscathed from the 80s to date despite the rise to prominence of the personal computer, introduction of the mobile phone and rise of the internet. Will the outcome be the same this time with the industry protected or will things be different? Let's take a look at what today's generative AI tools are capable of. Can they perform the job of an entry-level, $85,000 a year analyst?

What do Junior Investment Bankers Do?

Analysts are workhorses of the investment banking industry. They carry out manual, time-consuming grunt work which is essential to the big bucks their teams make. They churn out complex financial models, fancy pitch decks and conduct due diligence and research for transactions they work on, often at very short notice and while juggling multiple tasks. As highly paid as these entry level roles are, they aren't universally loved due to the high level of stress involved. In the height of the pandemic, a group of Goldman Sachs analysts made the headlines when an internal survey documenting unfair working conditions leaked. The major complaint was about working greater than 100 hours a week, poor mental health and workplace abuse. With the high adoption of GenAI and rapid increase in their capabilities, could the LLM tools make life easier for analysts? Or are they tools going to be deployed as replacements?

How Good are GenAI Tools with Investment Banking Tasks?

Besides ChatGPT by OpenAI which is by far the most popular large language model (LLM), there are other tools which are as powerful or in some instances more powerful. Claude AI, Google Gemini, Perplexity AI and Microsoft Copilot to mention a few. For this my 2024 high-finance version of Deep Blue versus Garry Kasparov, I am entrusting the task of defeating $100k a year analysts to ChatGPT. Anthropic recently released Claude 3.5 Sonnet, their latest, most powerful model to date. Sonnet outperforms GPT 4, Llama and Gemini on a wide variety of measures including math skills, reasoning and problem solving.

However, Sonnet can't create spreadsheets which is an essential part of an investment banker's job. Let's see how good ChatGPT is as an investment banker. I am picking a simple capital structure advisory transaction to see how ChatGPT will perform across the basic analyst / associate tasks. I have downloaded financial statements of a public company and fed a few pages in with the prompt:

After confirming that it understood my instructions, I instantly received the first output. Two sets of financial statements from the screenshots provided. It also created an Excel file which I could download in a single click. The Excel file was decently structured with each statement on individual tabs, the numbers were also transcribed accurately.

Verdict on Task 1: the LLM performs the task of transcribing numbers from a pdf into an Excel file accurately in seconds. Far outstripping what the most experienced or best trained analyst can deliver. A trained analyst would also format slightly better which is important in the investment banking world.

Okay, so what about the slightly trickier task of analysing the financial statements and identifying issues with the capital structure? The output from ChatGPT accurately identifies major issues like high finance costs, negative equity, increasing debt levels, and liquidity concerns from the balance sheet.

It however missed some positives and the analysis could have been expanded to include some more items. Guess what, I didn't go through the documentation to do the comparison. I simply uploaded the snippet of the advice above + original docs into another LLM - Claude AI this time. Claude AI picked up a few items that ChatGPT had missed, giving me a much better and holistic picture of what's going on with the company's capital structure.

This approach aligns with prompt engineering, a learned skill on how to get the best from LLMs. In summary, you get best results when you are as specific as possible. The clearer the instructions you give an LLM, the better the output you are going to get. The review of task 2 also nails task 3 which is a suggestion on what comes next. Point 5 talks about debt refinancing options which the company had flagged, in subsequent prompts the LLM accurately nails the suggested next steps for the company with two specific points:

Refinance high-cost debt to lower interest rates where possible, taking advantage of the company's improved scale and diversification.

Consider extending debt maturities further to reduce short-term repayment pressure.

Verdict on Tasks 2 and 3: the tag team of ChatGPT and Claude AI nails the tasks exceptionally well with minimal additional prompts. I would say the quality of the output and advice is above what would be expected of junior members of most investment banking teams. With the first task, I would say the LLM transcribes numbers at the level of a graduate level intern who has no previous work experience of investment banking training. With tasks 2 and 3, I'd say the output is at a level of a 2nd or 3rd year associate who has had some experience with financial statements and of corporate finance transactions.

Conclusion

The capabilities of consumer LLMs as at July 2024 is at a level that should scare any well paid professional whose job it is to create content or manipulate large amounts of data. These tools can do the job with high levels of accuracy at a tiny fraction of the time and cost of humans. More worrisome is the fact that the LLMs will only get increasingly more powerful. At some point, managers and corporate leaders need to decide how to deploy genAI tools - to boost efficiency of their teams or embrace opportunity to reduce costs.