Industry
Banking
Banking Services Transformation: Unlocking Exponential Value
Highlights
- Banks and insurers are turning to GenAI to usher in disruptive transformation across the banking, financial services, and insurance (BFSI) value chain.
- Defining a clear strategy and choosing the right approach to contextualize large language models (LLMs) for BFSI applications are key to success.
- While the final decision will depend on organizational requirements and business strategy, a hybrid approach that combines RAG and fine-tuning is preferable for BFSI firms.
On this page
On this pageinpage
Introduction
In the banking, financial services, and insurance (BFSI) sector, artificial intelligence (AI) has unlocked new opportunities and innovations.
BFSI firms have been quick to take advantage, leveraging AI to tap into rich data and extract actionable insights. In addition, the industry has traditionally deployed AI backed virtual assistants or chatbots in various areas such as customer support, risk assessment, financial advisory, and fraud detection and more, for over a decade.
Going forward, the adoption of AI is set to increase, and initiatives will predominantly focus on innovation and revenue. According to the BFSI industry report of the TCS AI for Business Study, 88% of Pacesetters (those organizations in the survey with stronger overall financial performance) are primarily focused on using AI for innovation and revenue growth. The study also revealed that 61% have AI implementations in-process or completed while 35% are planning AI projects. Unsurprisingly, firms are well aware of the need to ensure trustworthy AI outcomes which will necessitate strong governance processes—79% of BFSI firms are already reworking or planning to rework how they operate across the enterprise.
BFSI firms are now turning to generative artificial intelligence (GenAI), the next evolution in AI, which has tremendous potential for disruptive transformation across various functions like risk management, fraud detection, financial advisory, underwriting, claims processing, financial contracts analysis, personal financial management, portfolio assistance, among others. GenAI backed by large language models (LLMs) can analyze data and provide crucial business and operational intelligence to improve decision-making and help BFSI firms effectively handle their complex operations. The demand for efficient GenAI tools and platforms is therefore increasing rapidly—our study revealed that 55% of BFSI firms plan to create enterprise-specific LLMs for GenAI implementations.
While BFSI firms are optimistic about GenAI, challenges abound around data inaccuracies, fair and ethical use, transparency, trustworthiness, and hallucinations in the use of LLMs. Moreover, there are concerns around privacy and security and inadequate control over LLM responses. In addition, GenAI models will also increase the demand for compute resources. However, use of GenAI models is increasing in the BFSI industry and it is imperative to establish a safe and secure framework for implementing it.
Breaking through the bottlenecks
Several challenges will need to be overcome before BFSI firms can embark on building GenAI models.
- Training: Extensive data is required to train the model and produce accurate results. Obtaining vast amounts of quality data is difficult and time-consuming. Techniques such as synthetic data generation, data augmentation, and feature addition can be explored.
- Data security: GenAI models, which often handle sensitive data, are vulnerable to cyberattacks. Storing sensitive data on a virtual private cloud (VPC) within the hyperscaler environment can help mitigate the risk.
- Bias: Depending on the quality of the training data and the model architecture used, biases may creep in giving rise to questions on the fairness and transparency of GenAI models. Balancing the training datasets through data augmentation to increase the size of the minority datasets can help.
- Privacy: A data or privacy breach can occur if customer data becomes identifiable in model outputs, exposing BFSI firms to regulatory scrutiny and penalties. BFSI firms will need to establish a framework to ensure responsible use of AI.
- Hallucinations: GenAI models often produce factually incorrect responses. Using contextualization techniques can help as they deliver outputs that include citations to the original sources of the information which ensures traceability. BFSI firms must adopt the human-in-the-loop model which means resources will need to be deployed to refer to the original sources and verify the authenticity of the content generated, in turn minimizing hallucinations.
- Ownership issues: Disputes over the ownership of training or generated data can result in intellectual property and copyright infringement lawsuits. Policies and regulations in this area are still evolving and need regular review. BFSI firms will need to establish processes to ensure continuous monitoring of evolving rules and policies by setting up dedicated AI teams to fulfil legal obligations.
A step-by-step strategy for GenAI adoption
According to our study, a substantial 55% of BFSI firms plan to build enterprise-specific LLMs for GenAI implementations.
We believe that GenAI can be leveraged to design several innovative offerings across the BFSI value chain. This will require BFSI organizations to define a systematic strategy for GenAI adoption and contextualize LLMs to enhance efficiency and drive significant business outcomes. As part of GenAI strategy definition, BFSI firms must traverse the following steps.
- Ideate: Identify specific problem areas or use cases where GenAI solutions can be used. A dedicated office must be set up to accelerate the idea incubation process.
- Define prioritization framework: Prioritize key projects and initiatives that can significantly impact business and are relatively easy to implement.
- Experiment: Define a pipeline of projects to experiment with GenAI—this could include running pilots and proofs of concepts (PoCs)—to manage the complexity of developing, testing, and deploying GenAI models. This will ensure robust performance and seamless integration into operational workflows.
- Build reusable components: Create reusable components such as vector databases, model catalogs, and frameworks to help accelerate the development cycle and save time and cost.
- Upgrade infrastructure: Enhance the existing compute infrastructure based on the underlying workload, latency requirements, peak throughput, and evolving business needs.
- Establish governance: Monitor GenAI models to ensure accurate performance. Laying down stringent governance standards and frameworks is critical to ensure safe, responsible, ethical, and transparent use of GenAI systems.
Approaches to contextualize LLMs for BFSI applications
In our view, BFSI firms can adopt one of three approaches to customize or contextualize existing, pre-trained LLMs to develop applications for specific use cases.
Figure 1 depicts the three approaches as well as the use cases for which the approaches can be leveraged.
Figure 1: Approaches for developing LLM applications
Prompt engineering
This approach envisages leveraging pre-trained LLMs through an application programming interface (API) framework. As a result, it reduces capital expenditure by eliminating costly and time-consuming data collection, in turn speeding up project timelines and maximizing return on investment (RoI) while minimizing delays and budget overruns. Specific instructions are given to the existing LLM model to customize responses, reduce bias, and enhance performance, especially for tasks spanning areas such as summarization of financial reports, translation of legal documents, risk assessment, customer service, due diligence processes, and internal support. However, there is scope for hallucinated responses that may have adverse implications for financial decisions.
Retrieval-augmented generation (RAG)
In this approach, in addition to training data, a well-established, reliable data source, which may include domain-specific knowledge bases or the organizational database, is used to deliver LLM output that is accurate, relevant, and aligns with user queries. Vector databases are used to ensure efficient data handling. It significantly reduces infrastructure costs and deployment times, enables scalability in searching and retrieving information from heterogeneous elements such as financial transactions, stock trading, legal contracts, macro-economic developments, among others to benefit functions like investment advisory, claims management, underwriting, and portfolio advisory. However, multiple references in the sourced content can result in inaccurate or incomplete responses from the LLM model.
Fine-tuning
This approach leverages balanced, domain-specific data to specialize pre-trained, generic LLMs, adapting them to new tasks in specialized domains such as legal, finance, personal financial management, and so on. This approach is especially suitable for the BFSI sector given the twin needs of higher accuracy and to adapt the model to specific segments in alignment with the requirements of individual firms.
Moreover, firms can create multiple contextual models to address a variety of business requirements—infrastructure can be reused, optimizing costs for hosting each model. However, fine-tuning requires huge amounts of diverse, high-quality data and significant computational power.
The model can be further improved by constantly recalibrating it in response to human feedback and adapting it to changing business requirements. While this will boost reliability and user satisfaction, it can be cost-intensive and complex.
How TCS can help
In our view, BFSI firms must adopt a combination of RAG and fine-tuning (bolstered by human feedback) approaches to contextualize LLMs.
Though the final choice will largely depend on organization specific requirements and business strategy, a model built using a hybrid approach of RAG and fine-tuning is preferable for BFSI firms as it enhances the ability to understand technology and nuances related with specific domains.
Successful execution of GenAI projects will demand in depth knowledge of the BFSI domain, technology expertise, robust partnerships with cloud and technology providers, and experience in managing large transformation programs.
TCS has entered into alliances with several AI leaders to offer best-in-class services to support GenAI adoption at scale, create enterprise-centric models, establish responsible AI frameworks, and manage complex data sets (see Figure 2).
Figure 2: Capabilities that enable TCS customers to develop GenAI applications
Looking ahead
Adopting GenAI is crucial for BFSI firms to gain the operational and business intelligence necessary for strategic decisions, innovation, and creation of new value.
Unarguably, GenAI will help banks and insurers position themselves at the forefront of innovation and maximize growth.
GenAI is the future of the BFSI industry—the opportunities for transformative impact are tremendous. We believe that BFSI firms must look at GenAI adoption from a transformational lens and move beyond using it to develop point solutions for specific use cases. And needless to say, firms that leverage GenAI to innovate and rethink the way they operate will realize big wins.
