Industry
Banking
Banking Services Transformation: Unlocking Exponential Value
Highlights
- The financial services industry has witnessed widespread adoption of artificial intelligence technologies including GenAI.
- For financial institutions, adoption of AI agents is the natural next step, given their ability to act independently without the need for explicit human guidance.
- Financial institutions must explore and experiment with AI agents and consider adoption at scale once they start seeing business value.
On this page
Introduction
The rise of generative artificial intelligence (GenAI) in the financial services industry set the foundation for the rapid creation of humanlike content and quick, intelligent decision-making.
AI agents are a natural progression from applications backed by GenAI or large language models (LLMs), which are primarily informational in nature. AI agents integrate with external tools to execute multi-step tasks and achieve goals. Using the reason-act-observe-learn paradigm, as well as building blocks such as models, tools, instructions and memory, AI agents can personalize interactions, automate complex repetitive tasks, and continuously improve over time to deliver more relevant and effective outcomes. AI agents thus have the potential to change the game given their ability to independently perform tasks that require human judgement, without the need for explicit prompts or human guidance.
While the terms agentic AI and AI agents are often used interchangeably, agentic AI is a broader capability framework that enables AI agents to act with greater autonomy and adaptability, often by orchestrating multiple agents to achieve complex goals. In our view, it will take a little longer for agentic AI to become mainstream in the financial services industry, given the technology is still evolving. Consequently, we will limit this point of view to the role of AI agents in transforming the financial services industry.
In our experience, the financial services industry abounds with transformation opportunities using AI agents. Such agents may use LLMs such as GPT-4 (where general reasoning and understanding of language is essential) or small language models (SLMs like Phi 3) (for low-compute or edge devices or where cost, speed, and privacy are important). Additionally, industry or domain specific models (ILMs) such as FinGPT can be leveraged for highly regulated and accuracy-critical use cases. We use the terms LLMs, SLMs, and ILMs interchangeably since any of these can be used as the ‘brain’ of AI agents.
Despite the tremendous transformative potential of AI agents, widespread adoption in the financial services industry is a way off. This can be attributed to the absence of regulatory guidance on the use of AI agents and concerns around their ability to consistently deliver trusted outcomes. Additionally, financial institutions need guidance on areas where AI agents can add value and be deployed safely. However, to avoid being disrupted by fintech or big tech, financial institutions must experiment with AI agents, remain abreast of developments, and nimbly pivot as they start seeing value.
Applicability
Identifying the right areas in financial services to deploy AI agents is critical.
To accomplish this, it may be prudent to objectively review the entire business process for transformation opportunities using AI as well as non-AI levers. Financial institutions may find that adopting process simplification or deterministic workflows, based on traditional workflow engines or business process management (BPM) suites or technologies such as robotic process automation (RPA) work equally well.
While agents are ubiquitous in the software landscape, AI agents use AI models as their ‘brain’ to think, reason, plan, and act, and can play a significant role for processes with:
- Complex multi-step task execution: loan origination and underwriting, customer onboarding including ‘know your customer’ (KYC) and anti-money laundering (AML).
- Dynamic goal-oriented decision-making: personalized financial advice or product recommendations, dynamic credit limit adjustments.
- Evolving rules: real-time fraud detection.
- Voluminous structured and unstructured data: credit assessment for small and medium businesses
- Voluminous data processing at high speeds where human interventions are desirable but not possible: high frequency trading.
We believe that financial services organizations will create an ecosystem of wide-ranging purposive agents for different functions. While these agents will vary across different dimensions including technologies used (from ‘low code, no code’ frameworks to manually written code), types of agents (goal-based agents, utility agents and so on), scalability, and risks, they will coexist and seamlessly work together. Many areas will benefit tremendously from the adoption of AI agents (see Table 1).
Area |
Scope |
Agent complexity |
Use case complexity |
Business impact |
Business decision-making |
Real-time data analysis, monitoring, summarization Business analytics, presentations, and dashboards |
Medium |
Medium |
Productivity gains and enhanced employee experience |
Contact center operations |
Upsell/cross sell recommendations Personalized communications Actionable insights |
Medium |
High |
Productivity gains and elevated customer experience |
Investment banking |
Analysis of market volatility Suggestions on trade booking and hedge positions |
High |
High |
On-demand provision of quantitative computation and actionable insights for business decision-making |
Customer due diligence (CDD) |
Real-time customer profiling Risk assessment and scoring Verification and watchlist screening using real-time data |
High |
Medium |
Productivity gains; human intervention limited to exception management, freeing up human resources for higher value adding tasks |
Risk management |
Real-time tracking of competitors, news, price changes, and market buzz to generate early warning signals and briefs with upgrade or downgrade recommendations Real-time assessment of creditworthiness based on alternative data analysis Disaster management or stress testing simulation Real-time fraud detection and remediation |
High |
High |
Productivity gains, provision of actionable insights in real-time, autonomous decisions with human intervention to validate agent decisions |
Compliance |
Analysis of structured and unstructured documents, contracts Risk identification in legal documents and risk summaries |
Medium |
Medium |
Efficiency gains and human supervision for validating results |
Investment management |
Financial research and strategy intelligence Buy and/or sell recommendations with commentary Customized investment plans and advisory for retail and institutional investors |
Medium |
High |
Real-time provision of actionable insights for business decision-making |
Sales and marketing |
Customer lifecycle optimization Spend analytics and next-best action recommendations |
Medium |
Low |
Better customer experience and retention, higher precision campaigns |
Table 1: Areas in the financial services sector where AI agents can add value
From theory to action
For implementing AI agents, we recommend that financial institutions adopt a phased approach.
Phase 2 represents the ideal target state (see Figure 1).
Figure 1: Phased strategy for adoption of AI agents
The evolution of business use cases and technology along with necessary guardrails, observability, and evaluation occur incrementally and in parallel through the phases. Since AI advances are occurring rapidly, it is extremely important to ensure that AI agents are implemented using an evolutionary architecture. An enterprise level platform can be built to implement enterprise approved architecture patterns and archetypes, promote reuse, implement use cases in a federated way, and assimilate evolutionary patterns as they emerge.
- Start small in a sandbox environment, identifying a single business process with low customer and regulatory impact.
- Ensure buy-in from both business and technology stakeholders.
- Run pilots across different types of AI agent frameworks, use cases, and tools to identify the right fit based on organization specific requirements.
- Demonstrate the value of implementing AI agents.
- Implement simple business and technology metrics.
- Create enterprise-specific decision trees to identify the right use cases for implementation.
- Implement guardrails and observability sufficient for the use cases.
- Build foundational AI agent platform by leveraging insights and components from phase 0, while continuing to scale use case delivery.
- Move to use cases with medium impact.
- Generate regulatory and compliance reports.
- Transition to multi-agent collaboration, establishing standards for orchestration patterns.
- Ensure a scalable AI agent development lifecycle incorporating all necessary building blocks like registries for agents and automation.
- Improve upon the observability and guardrails in the platform.
- Identify an AI agent steward to promote trust and acceptance.
- Transition to complex use cases leveraging an ecosystem of AI agents.
- Optimize existing AI agents to increase the value of outcomes based on feedback and self-learning mechanisms.
- Embrace the fundamentals of FinOps to reduce costs—for example, deploying a lower cost model that delivers acceptable outcomes.
- Integrate with AI agents of trusted partners as well as agent and tool marketplaces for ecosystem play.
- Use the observability framework to analyze and optimize multi-agent behavior in a complex ecosystem of AI agents to enhance the maturity of the framework over a period of time.
- Employ LLMs or other agents to evaluate performance.
- Optimize the platform infrastructure (through caching, load balancing and so on) for efficient multi-agent delivery at scale, paving the way for agentic AI where a group of AI agents accomplish a business goal with minimal human intervention.
As financial institutions explore and experiment with AI agents, they must keep in mind that adopting AI agents comes with its own risks. Designing AI agents for the financial services industry requires balancing autonomy with compliance, precision, and control. Let us examine the key risks.
Budget: High costs and misaligned business objectives are big issues which can result in heavy capital expenditure as well as operational expenses.
Operational: A malfunctioning AI agent system that frequently leaks information or fabricates incorrect content or makes unreliable decisions, when scaled up, can become a massive operational risk. This can cause systemic failure and reputational loss.
Security and compliance: AI agents that are not adequately robust to withstand adversarial attacks or are not interpretable can lead to regulatory violations. Regulators demand clear rationale and logic for AI-driven decisions in the financial services industry—an AI system that lacks transparency and auditability, or is not sufficiently aware of regulations can result in serious compliance challenges. Critical regulatory mandates include:
- Notifying customers in advance that they are interacting with an AI system.
- Ensuring explainability and traceability of how AI-based systems reach decisions, especially in areas like dynamic pricing and credit assessments that have significant consequences for customers.
- Adopting an outcome-based approach rather than a prescriptive one by managing risk at various phases of the AI lifecycle:
- Design
- Development
- Deployment
- Operations and monitoring
- Test, evaluation, verification, and validation (TEVV).
This will require banks to put in place checks and balances to ensure safe, secure, explainable, privacy-enhanced, fair, accountable, and reliable AI systems.
Tackling root causes
Financial institutions will need to take the following steps to address the risks associated with incorporating AI agents into their operations.
Cost control
Implement a hybrid model comprising AI agents, robotic process automation (RPA), and LLMs to contain operational expenses (OpEx) related with training and maintenance. In addition, BFSI firms must exercise care while identifying use cases—only complex, multi-step use cases requiring proactive decision-making should be chosen for deploying AI agents.
Outcome assessment
Measure the quality of tasks completed by AI agents. This will entail testing fault tolerance and communication protocols, checking authentication and authorization across agent interactions, maintaining SOC 2 Type I and Type II compliance, and performing independent audits.
Impact assessment
Assess the impact of incorporating AI agents on product safety, liability, and security among others. Evaluate and adopt measures needed to ensure AI systems are accountable and capable of combating harmful bias. The assessment must be carried out by techno-functional people with the capability to evaluate from multiple dimensions—technical, human, socio-cultural, and legal.
Control assurance
Use control evaluations that test the safety of AI agents. AI systems can potentially cause systemic failure resulting in loss of data, malfunction of business systems, and stalled processes. Robust testing is therefore imperative. This can be achieved by engaging two teams—one stages control failure while the other takes action to prevent it.
Data quality and privacy
Ensure AI agents have access to accurate enterprise data. This is key as agents are only as good as the data they act upon. If banks use data from third parties, it is even more important to ensure the reliability of the data source. Additionally, banks must deploy data masking and privacy controls to avoid breach of sensitive proprietary information and to comply with privacy regulations. Tools used by AI agents should be finely permissioned based on role, task, and sensitivity.
Orchestration controls
Define and set guardrails on design aspects such as hierarchy, communication protocols, specificity of tasks assigned to agents, and mode of collaboration when multiple agents collaborate to achieve business outcomes.
Looking ahead
AI agents are set for astounding growth in the coming years.
AI agents can set the stage for expanding financial institutions’ portfolio of products and services in innovative ways, in turn unlocking immense value. However, given the potential flipside of large-scale adoption of AI agents in financial services, we recommend a cautious approach spanning a three-year horizon.
The way forward lies in defining a phased adoption strategy traversing the critical steps of controlled pilots, larger scale trials, and extended trials with an ecosystem of AI agents performing a variety of activities. Only after navigating these phases must financial institutions consider adoption-at-scale, which will depend on the behavior of autonomous AI agents and regulatory guidance. That said, human supervision is a must. AI agents cannot be allowed to act autonomously, an AI agent steward should be ultimately responsible and accountable.
