Central Bank tightening created a challenging environment in 2023 for many corners of the global financial services industry. Bank balance sheets were stressed and key business lines like mortgages and corporate deal making slumped. Asset management firms also weathered a tough year, with many funds outside of pod funds and private credit underperforming benchmarks. Insurance companies had a more stable year, but there were pockets of tension, including rapid repricing of flood and home insurance in some net migration areas.
Firms responded to this overall stress with a renewed focus on efficiency, spending parsimony, and tens of thousands in headcount reductions. Yet even amid this austerity, firms continued to prioritize a strategic investment in data and AI technology, seen as mission critical. Goldman Sachs perhaps best personified this dynamic. The bank kicked off a big round of layoffs and announced a major reconsideration of its consumer banking ambitions, including a wind down of its Apple credit card partnership. Yet even in this streamlining environment, Goldman increased its 2023 tech spending by 6% year over year, on top of the 15% increase from the prior year.
According to Gartner, "Worldwide baking and investment services IT spending is forecast to total $652.1 billion in 2023, an increase of $8.1% from 2022. Spending on software will see the largest growth with an increase of $13.5% in 2023."* On the leading edge, JPMorgan sent shockwaves throughout the industry by announcing plans during its investor day to increase its 2023 tech spending to a staggering $15.3 billion, including $4 billion on platform and products, and $3.2 billion earmarked specifically to “unlock the power of data” and “modernize technology and software development excellence.”
Just as in 2022, investments in the industry spanned the tech stack from databases to ETL tools, machine learning, and MLOps solutions. However, modernization efforts in 2023 often meant more consolidation to best of breed tools and moving from legacy technology solutions to newer and cloud enabled vendors where possible.
There was also major new excitement for developing Generative AI or Large Language Model (“LLM”) technology, but in a prudent, risk conscious way. On one hand, many banks quickly banned unmonitored employee use of ChatGPT on work computers and resisted the urge to rush any new tools into live production. But on the other hand, such prudence didn’t preclude firms from simultaneously moving ahead with controlled Generative AI experimentation.
As we now look ahead to the new year, we make our top predictions for AI (including Generative AI) in financial services for 2024.
Generative AI Technology Is Promising, but Risks Must Be Considered
Generative AI has huge near-term potential to add value across FSI. Deloitte predicts that the top 14 global investment banks can boost their front-office productivity by as much as 27%–35% by using Generative AI. JPMorgan is reportedly already working with regulators, walking them through how their first batch of models are constructed, monitored, and controlled.
Purely as a technology, there are many use cases where Generative AI can enhance existing tools and processes right away. However, the models will not be perfect, and importantly, they will err in novel ways compared to humans or well understood existing machine learning. As such, it is crucial for firms to develop new business processes, risk, and regulatory frameworks surrounding Generative AI. For example, a LLM chatbot might be able to retrieve and deliver financial information to clients 24/7 and faster than any human advisor can, making its immediate deployment very tempting.
But, what if a client suffers a loss when relying on information provided by a LLM chatbot in error? What if the chatbot is rude or discriminatory? What are the indemnification rules when a human employee isn’t responsible for the interaction? Will the reputational damage be magnified? These business and risk questions will be top of mind for FSI firms as they consider which applications of Generative AI can be safely implemented in 2024.
Global Markets Pace Ahead for Generative AI Production
Some of the more promising pilot use cases across FSI are focused on research and global markets. The theme is to leverage LLMs’ strength with natural language to summarize news, earnings, and central bank meeting transcripts for traders and investors who don’t have capacity to manually monitor everything. Historically NLP models like BERT have been used for these use cases, but the new generation of LLMs can generate richer analysis, and be asked to adjust their summaries in a back-and-forth manner that is more accessible for users.
LLMs can also be used to scan news and financial reports to perform “factor” tagging for stocks. LLMs can pick up on meaningful changes in sentiment, association with a topic, product, political issue, or virality on social media tied to a given company.
Generative AI can also be very helpful for a specific type of factor analysis that has been criticized recently: ESG scoring. ESG KPIs are often primarily identifiable via management statements, news, and public sentiment. LLMs can assist in identifying when companies alter their operations and messaging or become more or less aligned with ESG topics. We anticipate Generative AI to be extensively leveraged for ESG analysis to improve ESG usefulness or replace it with an alternative type of “corporate responsibility” analysis.
Using Generative AI to provide summaries and factors for traders and analysts to consider, keeps a “human in the loop.” As such, it can be integrated into existing processes without extensive business and risk re-engineering, and we therefore anticipate adoption quickly in 2024. More ambitious integration of LLM factor tags directly into trading systems creates significantly higher risk, and will be developed much more slowly and carefully.
Customer-Facing LLM Chatbots Need More Development Time
In the medium term, we expect to see improved customer service chatbots across consumer and corporate banking, wealth management, and insurance. Currently, retail and corporate customers seeking information have to manually speak to an employee during business hours or leverage clunky scripted chatbots. LLM-powered service bots represent a generational leap forward in their conversational fluency and flexibility in retrieving information and fielding unique questions. They have the potential to significantly cut customer support costs and improve customer satisfaction across FSI.
However, any chatbot intended for direct customer use has heightened reputational risk that will require very careful design and edge case testing before they are ready to be launched into live production. Extensive fine turning and integration of Retrieval Augmented Generation (RAG) engineering will also need to be explored. The chatbots must be highly accurate and use a tone and style that is professional and polite, and be consistently robust against inappropriate or discriminatory responses.
We expect to see major development efforts throughout 2024, and possibly some limited incorporation of LLMs into existing chatbots, but a completely LLM-powered chatbot live in production at a major FSI firm likely won’t be ready this year.
Investment Banking to Experiment With Generative AI
For many years, investment banks have struggled retaining junior staff, and have been eager to use automation to make a smaller staff more efficient while improving their quality of life. Banks dream of one day having a Generative AI assistant pull information from financial documents to build usable financial models and presentation slides.
This vision of a “robo-banker” isn’t practical in 2024. Financial documents and earnings reports contain graphs and tables with subtle footnotes that are very challenging for LLMs to interpret correctly. LLMs also have difficulty with the multi-step calculations needed to arrive at financial metrics that aren’t explicitly in the documents (e.g. ROIC or EBITDA). We see an encouraging trajectory for Generative AI agents to eventually improve with these tasks, but current error rates are far too high to be relied upon in an industry where tolerance for any numerical mistake is notoriously close to zero.
However, more modest goals are currently achievable which can improve banking process flows today. For instance, Generative AI agents can make financial documents queryable, and can be designed to provide approximate answers along with clickable footnote links that can take junior bankers directly to the document or part of the document relevant to the question. In other words, a major time saving improvement versus the basic “Control + F” technique.
Similarly, there is near term potential for Generative AI agents to assist in knowledge retrieval from the huge body of the firm's own deal assets and pitch decks. When a new deal arises and a junior banker needs to comb through past similar deals, Generative AI agents can assist with scanning internal documents and providing clickable links to the relevant asset. Retrieval Augmented Generation (RAG) engineering is likely needed here, but the low risk of this use case makes it a good candidate for near-term productionalization.
Generative AI’s potential is not limited to junior bankers only. Senior bankers will find that there are significant opportunities to enhance the CRM systems that they rely on. Many banks already use predictive modeling in these systems to flag clients likely to undergo key corporate events so that they can be targeted for outreach. Generative AI can help enrich these existing models by more effectively bringing in NLP features derived from news and management commentary. This enhancement of existing predictive modeling is a likely candidate for 2024 production, since firm model risk management teams can continue to evaluate model effectiveness with traditional machine learning performance metrics like precision and recall rates.
Finally, we also anticipate that banks will continue to experiment with other ways to leverage Generative AI to enhance their CRM systems. LLMs can provide a chat interface that bankers could talk to, to quickly retrieve useful CRM information and go back and forth asking clarifying or varying questions. Bankers sitting down for a live client meeting could ask the CRM bot questions about the company, its competitors, or about the corporate officers, and history of interaction touchpoints recorded in the CRM system related to these people and firms, and can get real time natural language answers with links to relevant information.
Unlike the high-risk customer facing chatbots, such internal facing chatbots have much lower risk of reputational damage and a more relaxed tolerance for errors, misswordings, or impoliteness. Therefore, we anticipate that improvements to CRM systems will represent some of the first Generative AI use cases to reach production in investment banking.
Looking Ahead to (Generative) AI Excitement
As evidenced by the growth in technology spending across the FSI industry, firms see investment in data and AI as being more important than ever. Generative AI represents the most exciting new opportunity in this race for competitive differentiation. Firms are eager to be the first to market with these powerful new tools, and are pushing ahead with use case development across many lines of business.
Nonetheless, we expect them to prioritize initial productionalization of use cases that keep humans in the loop, have minimal direct exposure to clients, and require reasonable accuracy in line with the current capability of the technology. Higher risk use cases will also see significant experimentation, but we expect firms to prudently keep them out of production in 2024.
*Gartner Press Release, "Gartner Forecasts Worldwide Banking and Investment Services IT Spending to Reach $652 Billion in 2023," June 21, 2023. https://gartner.com/en/newsroom/press-releases/2023-06-21-gartner-forecasts-worldwide-banking-and-investment-services-it-spending-to-reach-652-billion-in-2023. GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved.