A year ago, the mood surrounding Generative AI’s application in investment banking could be described as rosy optimism. We gauged this across prospect meetings, gung-ho announcements, and countless conferences and industry panels. We even recall a banking COO telling us that they were considering changing their summer IB analyst hiring for 2024 due to the potential of Gen AI.
For context, in late 2023, cutting edge research hinted at the promise of directly applying existing general-purpose Generative AI tools such as ChatGPT toward enterprise workflows. The headline statistics suggested easy double-digit gains in productivity. This not only catalyzed significant spending on infrastructure to bring Gen AI within the regulated perimeter of banks; it also reshaped the AI landscape between early adopters and fast followers. Increasingly, we saw organizations with the technical know-how aggressively pursuit lofty builds, while others preferred to see more materialization before making a headlong investment. This set off the era of Investment Banking 3.0.
Fast forward to today, and many banks have made progress, learning a lot along the way. But the challenges around an accelerated roll out of general-purpose Gen AI have become manifold, perhaps lending credence to the views of researchers on the need for diffusion to take place before general-purpose technology takes root.
There are five broad areas that have made adoption less seamless than optimists would have hoped: hallucinations, false confidence, entitlements, data security, and governance. These issues loom large within regulated institutions such as banks. Outside of engineering where coding copilots have hit the ground running, there are fewer role-specific productivity gains than initially projected.
When some of the issues surrounding data control and hallucination became during the Gen AI rollout, CIOs pivoted. Beyond coding copilots (a resounding success), most banks have focused their initial efforts on a core use case: improved enterprise search. While not as lofty as the dreams of a bionic banker, this is nothing to sneeze at. Try searching within a bank’s portal for a document or topic, and it quickly becomes apparent just how challenging it is to find a relevant match. Many banks employ thousands of middle-office employees whose job is to tackle the administration of such documents.
Improved search functionality is being powered by Retrieval-Augmented Generation. RAG is a panacea to some pressing problems for foundational models. RAG allows for an LLM’s ability to understand contextual relevance to be traceably applied to subject-matter specific proprietary information. That helps alleviate many (but not all) issues surrounding hallucination and false confidence by empowering the user to verify the source. As Marco Argenti, Chief Information Officer at Goldman Sachs, recently noted:
You have the RAG, which is the retrieval-augmented generation, which is actually interesting because you tell the model that, instead of using its own internal knowledge in order to give you an answer, which sometimes, as I said, is like a representation of reality, but it's often not accurate, you point them to the right sections of the document that actually is more likely to answer your question, okay? And that's the key. It needs to point to the right sections, and then you get the citations back. So that took a lot of effort, but we're using that in many, many cases because then we expanded the use-case from purely, like, banker assistant in a way to more, okay, document management, you know, we process millions of documents. Think of credit documents, loan documents.
Because RAG points the LLM toward information within the existing entitlements architecture, powered by say an on-prem deployed open source model, appropriate data security and governance oversight can be implemented. Quite the neat solution!
This brings great benefit towards the management of troves of information embedded in legalesque documents. Managing workflows across these documents traditionally required large amounts of human oversight to ensure compliance, accuracy, and applicability. As Shadman Zafar, co-Chief Information Officer at Citi put it, “Many operations roles require repetitive, rote tasks, like sifting through documents, validating data, and summarizing information, which are prime targets that AI can automate.”
We are seeing further indications that the problems of general-purpose applicability are being digested by banks as they roll out Gen AI. JP Morgan has recently rolled out its solution, dubbed LLM Suite. As a memo recently sent to employees noted, “Think of LLM Suite as a research analyst that can offer information, solutions and advice on a topic.”
In many ways, this is JP Morgan’s attempt to provide all of the “enhanced search” benefits of a general-purpose LLM such as ChatGPT (which it has banned internally) within the framework of its safety and compliance architecture. Because it is general purpose (and thus doesn’t address many of the drawbacks highlighted earlier), the bank stressed it should be viewed as an assistant with ideation, instead of a copilot within the nitty-gritty workflow. And so, it is tempering the expectations around its applicability to role-specific work.
Morgan Stanley has addressed these problems by focusing their Gen AI effort on private wealth, something we noted in our most recent investment banking earnings recap. They have equipped nearly 16,000 advisors with a chatbot called AI@MS, with what appears to be a RAG architecture. The intent was to improve upon an existing knowledge base service called FAST.
Private wealth almost certainly contains more use cases where RAG-enhanced retrieval and LLM-powered synthesis can be productive time savers, with less downside risk. Compare the risk of false confidence with an advisor sending a client a market summary, versus a banker building a valuation model.
So, it has become clearer in the last year that CIOs have appropriately identified the strengths and weaknesses of LLMs. By mostly working around those weaknesses, they have rolled out functionality by channeling the technology through lower risk pockets of the bank and by layering in RAG for more tailored use cases. The result is a partial roll out that will take time to expand its footprint.
What is conspicuously absent is the promise of role-specific benefits, geared towards role-specific tasks. Take for example the intricacies of what a DCM banking managing director does daily vs a junior analyst. Is there a Gen AI solution to this? Well, no, beyond making an email more concise. This is probably due to the list of limitations that we highlighted and the need for prompting to break a user out of their existing workflow. It is our view that there still is a need for more use cases and paradigms to be discovered that drive role-specific productivity.
At ModuleQ, we aim to deliver role-specific productivity to investment bankers, full stop. It’s a BHAG. But like all great aspirations, delivering on what’s hard is often what’s truly valuable. It is how you “win the future.”
The paradigm we’re using to deliver role-specific productivity into investment banking is our Unprompted AI. Unprompted AI works differently than today’s Gen AI in a variety of meaningful ways. We believe it addresses many of the weaknesses discussed above, while also unlocking the role-specificity that the industry so desperately needs.
Unprompted AI delivers salient insights directly to bankers. By tailoring those insights to their habits, persona and role, we are enriching their specific workflows with salient and timely information. By delivering it directly to them within their communications hub (Microsoft Teams), we are respecting the need for them to focus on their existing workflows instead of context switching to new ones. Investment bankers are fundamentally information brokers. By equipping them with better information across markets, internal research, as well as clients and prospects, we aim to make them better bankers. Early indications are of meaningful productivity increases in their specific workflows.
Because Unprompted AI focuses on seamless delivery rather than the lofty aspiration of general-purpose omniscience, we can avoid the messy downside of hallucination and false confidence. By configuring our recommendations within the bank’s perimeter through trusted partners, we address AI governance issues.
This means we can focus on what drives investment banking productivity. ModuleQ’s Unprompted AI delivers ROI in the form of greater prospect generation and better client communication from the get-go, without challenging implementation hurdles.
If the last 12 months were a realization of the challenges in applying general-purpose technology to investment banking workflows, the next 12 months will be about finding role-specific productivity gains to justify the time, effort, and expense of deploying AI within the bank.
With all the AI investment made by banks, and the pace at which change is accelerating, the future is now. We are convinced that role-specific productivity will unlock the next generation of winners, and we think that novel paradigms such as Unprompted AI will be the solutions that win the future.