Industry
Banking
Banking Services Transformation: Unlocking Exponential Value
Highlights
- The use of AI has exploded in financial services with many banks and financial institutions running pilots to assess the viability of adopting AI across various functions.
- Banks are seeing value at the proof of concept or pilot stage; however, this is not translating into scalable benefits, compromising returns on AI investments and resulting in waste of time, effort, and resources.
- To enhance AI RoI, banks and financial institutions must adopt a set of best practices, grounded in data science, at all six stages of the lifecycle of an AI project, which will greatly improve the success of AI initiatives.
On this page
Why AI RoI is elusive
The last two years have witnessed a massive growth in initiatives to explore how AI can be applied across sectors, and the banking and financial services industry has been no exception to this trend.
The intent typically is to automate or semi-automate the work of humans. While these initiatives start out as experiments or pilots, in our experience, many of them are poorly conceived, planned, and executed. Consequently, though banks achieve success at the proof of concept (PoC) stage, AI projects frequently do not translate into tangible and scalable benefits despite investing heavily, resulting in misdirected efforts and wasted opportunity.
With AI evolving rapidly, new opportunities are emerging, and financial services firms expect to embed AI into their processes and capture benefits. However, since AI is new to many financial services leaders, there are potential pitfalls. To overcome these pitfalls and enhance RoI from AI initiatives, we recommend a set of best practices to be followed across the lifecycle of AI projects. These best practices are grounded in data science, and are organized by lifecycle phases to help banking leaders bring them in at the right stage of the project, in turn improving the success of AI initiatives.
In an earlier thought paper, we explained why having the right mix of the four key pillars—people, technology, data, and domain—is critical to convert PoC value into long term returns on investment (RoI). In this paper, we highlight the considerations which become critical at each life stage of an AI project and the associated best practices.
How AI complements data science
The financial services industry has been an early adopter of traditional data science techniques.
Risk modeling, quantitative modeling in asset-liability and liquidity management, and credit scoring for consumer lending are well established.
These techniques utilize structured and/or quantitative data from consumers and markets. However, banks also leverage significant amounts of relevant, unstructured data. For instance, banks frequently process invoices receipts, purchase orders and other documents that are typically received as image files—extraction of information from such documents has become easier with AI technologies. Similarly, other lines of business in the financial services sector also need to process unstructured data. However, traditional data science approaches have not been able to utilize such unstructured data in a big way—with its ability to process and unlock value from unstructured data, AI can fill this gap and unlock exponential business value.
Generative AI (GenAI) techniques come with significantly enhanced capabilities to process unstructured data and vendors are releasing products, some of which have broader capabilities necessitating more effort to implement. At the same time, there are other products suitable for narrower use cases, but require comparatively less effort to execute, making it faster to build production grade solutions. AI features can be added into existing IT products already in use in the financial services industry, easing the process of incorporating AI into workflows. These advancements are reducing the investment needed to incorporate AI capabilities into banking workflows, in turn opening new transformation opportunities.
The AI lifecycle
A typical AI project traverses six stages across its lifecycle.
Each stage involves several steps (see Figure 1) aimed at capturing and addressing risks and directing effort where the RoI is the highest.
Figure 1: Lifecycle stages in an AI project
Financial services leaders must look at business or IT processes that use unstructured data—that’s where potential AI use cases arise. From a retail banking perspective, acquiring and cross-selling is an important area where AI technologies can help better exploit available data on customer or prospect behavior. A leading global European bank incorporated GenAI into campaign analytics, cross-sell, and next best offer processes leading to increased business. Leveraging GenAI in the risk management function can help banks speed up the analysis of financial reports, analyst reports, and documents that are key to lending to the small and medium enterprise (SME) segment. In the consumer lending space, a global US bank invested heavily in social media profiling to help predict customer defaults. The financial services industry uses unstructured data in several areas, translating into a wide-ranging list of potential use cases for AI deployment. With traditional data science approaches lacking the capability to exploit unstructured data, harnessing it demands the deployment of large language models (LLMs).
From the long list of potential use cases identified in stage 1, a few should be selected for running PoCs. The objective is to assess feasibility and get an approximate evaluation of the benefits vs costs (initial investment and ongoing costs) as well as uncertainties or risks. This will help make the right decision in rolling the PoC into production.
To prioritize the right use cases, financial services firms will need to navigate the following steps:
- Assess feasibility to rule out obviously infeasible ones. For example, selling personal loans to a prospect. An experienced customer may feel comfortable purchasing such a service without speaking to a human, but the average customer may not be willing to sign up without speaking to a human. AI cannot replace the human connection that many seek.
- Ascertain costs and benefits, considering maintainability, reliability, and uncertainties. Scale and stability play a big role here. Costs are relatively fixed, while benefits are proportional to the scale. For instance, automating 80% of loan workflows may be beneficial for a large or mid-size bank, but it may not be viable for a small credit union. Additionally, making changes to the underlying processing logic in response to frequent changes in forms and documentation can be expensive.
- Check if digitization is needed as a precursor to the PoC or production stage. For example, consider the mortgage process, where typically, a loan advisor talks to customers to understand their situation in detail before recommending a suitable type of mortgage loan. A significant portion of this information may reside in the advisor’s head or may be captured as notes that only the advisor understands. Standardizing and digitizing this process is a prerequisite for AI adoption.
- Ensure that the team prioritizing the use cases has a broad understanding of the domain and technology across software systems, machine learning (ML), and AI. This will help determine if the business goals for a use case are better served by traditional ML or even a business rule-based system or a combination of the two.
Clearly, inferring where AI can help and where it cannot is crucial. Often, the most workable solution is one where AI supports humans to improve their productivity and quality of work rather than fully replacing them.
The intent of the PoC is to determine whether to proceed with AI. This means that the scope of the PoC must extend to evaluating or validating aspects whose outcome will influence AI investment decisions. Furthermore, the evaluation must be end-to-end, focusing on aspects such as integration of the AI components into the workflow, user and customer experience, computing costs, security considerations, regulatory compliance, and ease of making changes.
The scope of the PoC must also help identify the workflow areas where using AI can add value and determine how it will work with humans. Traditional software systems give consistent, reliable, and deterministic answers which is not a realistic expectation with AI. To efficiently harness the potential of AI, financial services firms must adopt a rule-based approach to demarcate activities that can be handled by AI into two categories: those that can be handled reliably by AI alone and those that will need human intervention or oversight. For example, AI can be leveraged to automate data extraction from invoices and automated checks can be put in place to confirm the total and so on, skipping human checks. On the other hand, for processes such as high value loan approvals, where the stakes are high, a human must review every case. Ensuring cost-effective PoCs may necessitate manual data transfer or copy-pasting data between systems without modifying production systems.
The next step is execution, where obtaining real user experience on the actual live workload of users is crucial. This necessitates caution, specifically, if the workflow does not involve a human in the loop. It would be prudent to have a human reviewer monitoring the AI, especially when the stakes are high, until the model starts performing reliably. Regulatory factors also come into play, for example, recommending loan products is subject to the appropriate sales mandate where banks are required to act in customers’ best interest—in such a scenario, human monitoring becomes necessary to ensure compliance.
The final step involves evaluating the PoC across parameters like user adoption and user experience. There are two aspects to this:
- Users accustomed to doing things in a certain way may resist the change. Business leaders must dig deep to identify valid, material concerns around quality of AI responses and customer experience.
- Identifying the types of users that will benefit substantially from AI adoption and those that will benefit to a lesser degree—this could depend on users' skills, experience, or baseline performance. Studies show that AI boosts the performance of users whose prior performance was below average.
At the end of this stage, financial services firms will have a clear picture of:
- The types of activities AI can help with.
- Whether the activities can be fully automated or human supervision is necessary.
- The types of users that will benefit most and those that will benefit to a limited extent.
- A high level design of the end-to-end system and workflow.
If the PoC indicates that incorporating AI will add value, the next stage is implementation. This stage involves finalizing the end-to-end solution architecture and testing, defining strategies for user adoption, ensuring ongoing quality assurance, adapting to change, and measuring customer experience.
Solution architecture and testing: In this step, the integration of AI components into production workflows is completed and comprehensive testing is performed. Additionally, edge cases in scope that were not covered in the PoC are tested.
Customer or user experience: Users are ultimately responsible for their actions, whether or not they take AI assistance. Consequently, building trust in the AI system is important to drive users to use and rely on it. Ensuring the explainability of AI responses is critical to building trust. For example, if a sales agent relies on AI systems to recommend suitable loan products, backing those recommendations with detailed explanations will greatly assist the sales process.
Quality assurance: AI systems are not as consistent as traditional business logic-based software systems. So, continuous monitoring and review of AI outputs are critical to ensuring quality. AI models will need to be periodically fine-tuned, and guardrails may need to be adjusted as well. Consider, for example, an AI chatbot that guides customers in choosing a credit card. A sample of the AI responses/interactions must be reviewed by humans on an ongoing basis to help identify any erratic behavior or systematic errors which can then be fixed.
Resilience to change: Financial services firms are not static—they must adapt nimbly to an evolving market to retain a competitive edge. This may necessitate changes to their product and service portfolio as well as internal IT applications or workflows. Modifications may also be needed to align with changing regulations. At the same time, LLM models too are evolving and improving. Financial services firms must design AI systems incorporating the necessary degree of flexibility to accommodate such external demands. A common practice is to architect the system adding a business logic or rule-based layer on top of the core LLM model. This layer can be quickly reconfigured, allowing the AI model to be rapidly adapted in response to business changes. For example, when recommending banking products, an AI model can be used to identify the product type that the customer is most likely to buy (unsecured credit card, secured credit card, or pre-paid card and so on), with the final recommendation based on business logic.
This stage involves transitioning the actual tasks from the existing process to the new AI process. The first step is to identify an overall owner, whose role will entail:
- Enabling a smooth transition and ensuring that the new process consistently functions as expected.
- Aligning stakeholder expectations, coaching users, and managing users’ concerns around the new process.
- Understanding the key performance indicators (KPIs) of the users. These KPIs may need to be refined to suit the new AI based process. For example, if a part of the work is automated via AI, KPIs should be redefined to measure the work that remains with users.
- Working with the developers to ensure that edge cases that were not covered in the PoC or in stage 4 or discovered subsequently are tested at this time.
- Establishing new guardrails and improving existing ones.
- Managing change arising from changes in the financial product and service portfolio, upstream or downstream IT systems, regulations, as well as upgrades to newer AI models.
The initial one or two PoCs or AI deployments are typically driven and backed by an executive sponsor. These initial deployments serve as a learning ground to understand what it takes to make AI work. Banks can expect meaningful benefits only after AI is more widely deployed.
So, how can financial services executives create an ecosystem for scaling AI within their firm? In our view, firms must establish a centralized AI Center of Excellence (AI-CoE) staffed either with in-house talent or a hybrid team that includes external consultants. The AI-CoE must oversee AI projects, foster experience sharing, and centrally manage shared AI computing resources, software subscriptions, and data governance policies.
Typically, the AI-CoE members actively engage with stakeholders of various business functions to evaluate and shortlist areas or use cases where AI can potentially add value and then go through the five AI lifecycle stages to operationalize suitable ones. The level of involvement of the CoE members versus that of business function members would depend on the skills available in the business unit. Ideally, CoE members must handle aspects requiring deep AI skills and expertise leaving domain aspects to business functions.
AI benefits will vary widely from one process to another, but it is undeniable that the aggregate net benefit will be significant. Over time, banks will benefit from a critical group of people, with knowledge of AI applications, across various teams, helping to effectively harness the benefits of AI.
In summary
Though the potential of AI to deliver untold benefits to the financial services industry is indisputable, returns have been elusive, despite investing millions of dollars.
Furthermore, the technology is evolving rapidly—the advent of GenAI and agentic AI are a case in point. Financial services firms do not have complete clarity on what works and what does not, especially as there are limited number of practitioners with real experience in deploying AI into production workflows.
However, we believe that AI in production use cases is set to explode in the coming years though it is difficult to estimate the pace at which these changes will occur. Most use cases will involve a human in the loop to varying degrees, evolving along two parallel streams. First, technology products in the market will increasingly incorporate AI into their workflows—most commercial or off-the-shelf products being used by banks will increasingly use AI-assisted processing and reduce human effort. The second stream is in-house development activity that leverages AI to improve productivity. They could take the form of coding assistants, document analysis and summarization, which is frequently an intermediate step in several processes such as loan approvals, automated extraction of information and so on.
To support these activities, banks and financial institutions must ensure that on the ground AI teams are led by experienced leaders—individuals with experience in deploying data science solutions even if they lack AI experience given the significant overlap in learnings. Financial services firms must build an AI-CoE to scale AI across the organization. However setting up the AI-CoE will come with its own pitfalls, and banks and financial institutions must consider partnering with a service provider with the requisite technology and domain expertise as well as AI-skilled resources.