The top 14 international investment banks could increase their front office productivity by as much as 27–25%, according to a recent analysis by the consulting firm Deloitte titled “Unleashing a new era of productivity in investment banking through the power of generative AI.” According to the same estimate, by 2026, each front-office employee will earn an extra USD 3.5 million as a result of this productivity boost.
The inherent hazards of AI for global financial institutions need to be addressed, despite the rapid deployment of generative AI. These hazards include privacy issues, bias, performance robustness, and other possible cyberthreats, according to the International Monetary Fund.
Recently, Andrea Haskell, Deloitte Principal in Strategy and Analytics, and Val Srinivas, Head of Research at the Centre for Financial Services at Deloitte, met with Daniel Faggella, CEO and Head of Research at Emerj, to talk about new and developing generative AI tools for news analysis and summarization in financial workflows and its applications beyond trading, like M&A, advisory, and debt issuance. The trio also looks into how to rank the use cases for adopting generative AI.
The paper that follows examines their viewpoints and offers executives the crucial takeaways from their discussion:
Rethinking operating models: In order to guarantee scalability, sustainability, and responsible risk and control, operating models must be rethought when moving from proof of concept to production.
Adopting generative AI by using a prioritizing matrix: In order to prioritize implementation, use cases are mapped on a two-by-two grid according to business value and complexity. This allows users to choose between low-hanging fruit with lower complexity and revolutionary but complex prospects.
Reevaluating Operating Models
Val starts out by bringing up the same Deloitte paper about how generative AI could help investment banks increase front-office efficiencies. The topic of incorporating generative AI technology into current systems, procedures, and infrastructure is then brought up by him. According to Val, the emphasis should be on increasing productivity, reducing costs, and promoting revenue growth. Val does admit, though, that most organizations are probably putting efficiency and productivity improvements ahead of revenue growth at this point.
He illustrates how generative AI might increase productivity, particularly for young analysts in investment banks, with a use case. A significant amount of effort is spent by these analysts manually collecting and compiling data. This procedure can be made more efficient by reducing the amount of manual labor required with generative AI techniques.
As Val explains, generative AI may produce summaries or paragraphs as its initial output, but it may also expand into more diverse use cases:
How much time do we spend watching movies, listening to podcasts, or any other type of audio content? You may assume that non-text modalities will become increasingly significant as their use grows globally. Perhaps there is a video or an image of an event taking place, and it hasn’t even been converted into text yet. How does generative AI include that? And how can you use all of it to produce new outputs that combine text, music, video, and other elements?
-Deloitte’s Centre for Financial Services’ Head of Research, Val Srinivas
He goes on to discuss the importance of latency in securities trading. The true challenge, however, is whether we can rely on the results produced by generative AI in a short amount of time.
Andrea talks on how the discussion has changed and the latest developments in investment banks’ use of generative AI. Use cases and demonstrating capabilities through proof of concept were the main topics of debate at first.
Second, she talks about the ramifications for talent, such as reskilling and upskilling younger workers. She even challenges the notion that the skills students are learning in school are sufficient for the demands of the workforce.
Thirdly, Andrea suggests reconsidering operating methods and appropriately managing risks and controls. It is imperative that the crucial factors be taken into account while advancing generative AI from a proof of concept to a real production setting. In order to foster trust in the dependability of the technology’s outputs, she highlights the necessity of control points.
Andrea then gives an actual example of a “doomsday” situation that might occur if generative AI creates a public filing and a financial model takes content out of it to carry out large-scale buy or sell orders. She talks about the possible dangers, like mistakes in public filings that could have serious repercussions for companies.
Adopting Generative AI By Using A Prioritization Matrix
When the conversation turns to prioritization, Andrea refers to generative AI as “augment technology,” trying to emphasize that generative AI is meant to augment human work, not replace it entirely. She mentions that it can enhance the value of work by allowing humans to focus on higher-order tasks, as AI would handle the more repetitive or ‘grunt’ work.
Andrea addresses misconceptions about the future role of generative AI and AI in general. She refutes the idea of an autonomous future where machines fully control the world in the near term. She highlights that throughout human history, inventions have been designed to free up human time and not to replace them.
Andrea specifically emphasizes generative AI’s nature as a tool that generates content and summaries. She acknowledges the concern about AI output sounding accurate and competent, leading to potential misinformation or false confidence. She stresses the need for safeguards to protect against such pitfalls, including addressing biases in training data and incorporating human oversight.
“Our data is biased, it’s just a fact. So how do we ensure that — with any data we’re training models on — that we’re using the right techniques and guardrails to fine-tune datasets that are not biased? [All] to try to ensure we arrive at outcomes that are as least biased as possible. Then again, back to the human-in-the-loop component: That’s where you really just cannot have fully autonomous systems in this area, and in the near term, because you need that human higher-order thinking.”
-Andrea Haskell, Principal in Strategy and Analytics at Deloitte
Additionally, she dispels the misconception that AI, including generative AI, possesses proper understanding or anything that can be described as a ‘moral compass.’ Instead, she specifies that these systems excel only at predicting the next word based on training data but lack the higher-order thinking of human agents.
The current focus of the conversation around the adoption of generative AI is strategically prioritizing use cases based on a framework that considers both business value and complexity.
The matrix involves mapping use cases against two axes: business value and complexity. High-value, highly complex opportunities might be transformative, representing a significant business case, but they may take time to implement. These could be part of a roadmap for future adoption rather than the initial starting point.
On the other hand, there are “low-hanging fruit” use cases that offer business value with lower implementation complexity. In the context of banking clients, Andrea mentions examples of middle and back-office efficiency use cases that use internal datasets, minimizing reputational risk at an early stage.
By employing this framework, organizations can strategically choose where to begin their adoption of generative AI. As Andrea says, this matrix provides a way to experiment without reputational risk at this early juncture.
Later in the podcast, Andrea discusses the evolving job landscape in the context of generative AI adoption. She emphasizes that while some jobs may see a decrease in demand or even be entirely replaced, there will also be the creation of net new jobs. The key is to recognize that as AI becomes more integral, there is a growing need for fluency within organizations — specifically in data-, technology-, and AI fluency.
She notes a broader trend where every business leader is now expected to become a technology leader as well, highlighting that AI will not be limited to a specific group or department but will be integrated into various aspects of job responsibilities across the enterprise. These integrations will be seen as part of the overall enterprise strategy.
Finally, Andrea underscores the importance of reskilling the workforce to adapt to this changing landscape. The goal is to enable employees to interact with generative AI and similar technologies responsibly, thinking critically about the outputs they receive.