Generative artificial intelligence (Gen AI) has become the topic du jour in recent years, permeating corporate boardrooms and reshaping industries. OpenAI's ChatGPT, among others, has further shifted the horizon. Many industries, including financial services - the traditional bastion of data-driven decision-making, are making rapid strides in harnessing Gen AI’s possibilities. 

Exciting as this sounds, firms are also recognising the need to invest in overcoming its pitfalls. In this blog series, we unpack both sides of this revolutionary technology, within the realm of financial institutions.

The genesis

The emergence of Gen AI owes much to the remarkable surge in computational horsepower through cloud computing and the evolution of graphic processors capable of tackling resource-intensive calculations. Gen AI has transitioned from the experimental playground of tech giants to the drawing boards of organisational strategy. It is increasingly finding its way into the arsenal of financial institutions that are constantly in quest of sharpening their competitive edge.

Understanding the perils

Financial Institutions have three broad reasons to be circumspect. 

Adverse outcomes: The foremost concern revolves around Gen AI’s occasional tendency to unbridled responses leading to undesirable outcomes. The notorious ‘hallucinations’ and biases observed in AI-generated responses are especially relevant to address in the financial sector, which demands precision and accountability with a lot at stake.

Legal and ethical quagmires: The legal and ethical terrain of Gen AI remains murky. Organisations must navigate the treacherous waters of copyright infringement and unlicensed content in training data while meticulously assessing the ethical implications of using AI-generated content for client interactions.

High costs: Building and maintaining AI infrastructure, especially when developing proprietary large language models (LLMs), can be financially daunting. Costs encompass CPUs, resource-intensive graphic processing chips (which can exceed $40,000 each), and expenses related to training and maintaining datasets.

Addressing the concerns

These concerns resonate across industries. But it is interesting to see how financial institutions are tackling them head-on.

Technical safeguards: These are strategies to enhance the LLM architecture to improve contextual recognition while minimising the risk of hallucinations. Banks are refining data ingestion and indexing processes, optimising retrieval/filtering algorithms, and implementing user experience guardrails such as citation links and disclaimers.

‘Prompt’ engineering: Crafting precise prompts guides the model in producing domain-specific responses. Techniques like ‘one-shot’ or ‘multi-shot’ prompting help the model understand user requirements. ‘Chain-of-thought’ prompting aids in addressing complex queries requiring reasoning.

Data usage and protection¹: Financial institutions keen on deploying LLMs for financial intelligence are also emphasising the use of proprietary data for model training, minimising reliance on publicly available data. Some banks exclusively utilise their internal data ecosystem to mitigate risks associated with internet-sourced data. They are putting in place robust measures to prevent leakage of personal identifiable information that ensure models discard personal data after customer interactions.

Periodic evaluation²: In addition to the above, banks must establish governance mechanisms to regularly assess model performance and prevent drift or bias from taking root.

Evaluating use cases

While the financial industry unanimously acknowledges the revolutionary potential of Gen AI, practical, scalable use cases are still in the works. These use cases could broadly be categorised as:

Efficiency boosters: Applications targeting enhanced employee productivity and cost savings

Value proposition augmenters: Innovations that enrich client value propositions

Synthetic data generators: These leverage Gen AI to create pseudo data for training existing machine learning models

What lies ahead

As financial institutions are increasingly drawn into the realm of Gen AI, the path forward would demand meticulous consideration of the risk-reward dynamics inherent to each use case.

Close scrutiny of challenges across the dimensions of technology, data, ethics, finance, and legality before committing substantial investments, would save on a lot of anguish later.

We have worked out the evaluation of one use case for our readers below (see, An illustration), to drive home the point.

In our forthcoming blog series, we will explore the value chain of the banking, buy-side and markets industries, assessing Gen AI's potential to enhance value propositions, elevate productivity and redefine the client experience.

Stay tuned. And let us know if you would like us to dissect any particular process!

 An illustration: Working capital lending analysis

Gen AI shows bright potential in the domain of working capital lending analysis. This analysis involves the identification of new products and credit offerings, and the early detection of credit risk red flags.

Relationship managers (RMs) typically pore over cash flows, profit and loss statements, and balance sheets while doing this analysis - seeking signals in cash seasonality, creditor aging, and asset-liability dynamics (see Figure 1).

Figure 1: Typical analysis process

In this use case, let us look at the transformation of this process with the aid of LLMs. LLMs possess the strength to capture context and discern patterns, rendering them adept at generating insights. LLMs can be trained to detect this shift in asset and liabilities, identify cash flow patterns, and pinpoint concentrations within supplier and customer relationships.

Figure 2: AI enabled process

Example prompt 

“Context: <12 to 18 months of Financial Statements of Acme Inc>

Question: Please give me your best analysis and factors for month-over-month seasonality in cash flows for Acme Inc for July2020-June2021, showing minimum cash balance, maximum cash balance, and average cash balance. Please give me numeric values for these flows and draw a graph with over time with minimum, maximum and average cash balance

Do not use any information that is not available in the above context”

Answer

“Analysis of month-over-month seasonality in cash flows for Acme Inc for July 2020 to June 2021

Factors:

• Seasonality in sales: Acme Inc's sales are typically higher in the summer months, as consumers drink more beverages during this time. This leads to higher cash flows from operating activities in the summer months.
  
• Seasonality in expenses: Acme Inc's expenses are also typically higher in the summer months, as the company spends more on advertising and marketing during this time. However, the increase in expenses is not as large as the increase in sales, resulting in higher net cash flows from operating activities in the summer months.
  
• Dividend payments: Acme Inc pays quarterly dividends to its shareholders. These dividend payments are made in February, May, August, and November. Dividend payments reduce Acme Inc's cash balance.
  

Minimum, maximum, and average cash balance:

The following table shows the minimum, maximum, and average cash balance for Acme Inc for each month from July 2020 to June 2021:

Graph of minimum, maximum, and average cash balance:

The following graph shows the minimum, maximum, and average cash balance for Acme Inc for each month from July 2020 to June 2021:

Analysis of the graph:

The graph shows that Acme Inc's cash balance typically increases from July to October, as sales are higher in the summer months. The cash balance then declines from November to February, as the company pays dividends to its shareholders and spends more on advertising and marketing during the holiday season. The cash balance then increases again from March to June, as sales are higher in the spring and summer months.

Acme Inc's cash flows are typically seasonal, with higher cash flows in the summer months and lower cash flows in the winter months. This seasonality is due to the seasonality in sales and expenses, as well as the company's dividend payment schedule. This makes Acme Inc a candidate for higher ”

Analysis for the use case

• Introducing LLMs into the equation, as depicted in Figure 2, thus enhances efficiency and scalability, enabling RMs to cover multiple clients.
  
• The fiduciary and reliability risks associated with this LLM-driven transformation are managed through a rigorous review process by the RMs. 
  
• Legal and ethical concerns are controlled through the implementation of robust data controls during training and prompting. 

Figure 3: Risk-reward plot

The rewards, on the other hand, are abundant across the spectrum - efficiency, accuracy, scalability, and elevated customer experience. This rendering transforms working capital lending analysis into an ideal candidate for the Gen AI metamorphosis.

References:

1: https://bit.ly/3Q3nFzJ
2: https://a-teaminsight.com/blog/building-trust-in-ai-an-imperative-for-widespread-adoption/