Building Trustworthy Agentic Ai Systems FOR Personalized Banking Experiences

Abstract

Artificial Intelligence (AI) systems have become ubiquitous in contemporary society and have the potential for transformative impact on user behavior. These systems are capable of learning autonomously to personalize their behavior to deliver improved user experiences. However, there exists the potential for unintended consequences, as the same agentic features associated with positive outcomes may also increase the capacity for negative outcomes. Financial services is an example of a domain where deploying AI systems with agentic features would be high risk. The automated decision-making capabilities of these systems could influence billions of dollars. Nonetheless, they would be entrusted with taking actions that affect users without human oversight, such as reallocating entire portfolios of assets in ways that users do not wish. Therefore, a foundational requirement for adopting these systems must be the capacity to build shared norms of beneficial behavior prior to their deployment.

Several commercially available AI systems with agentic features are already deployed in the domain of personal banking. Automated personal financial management combines categorization of transactions followed by predictions of future expenditure and savings to improve budgeting decisions, amongst other impacts. Digital banking assistants embed intelligent conversational agents used primarily for accessing banking information and services. These systems typically operate in conjunction with non-intelligent user interfaces, and thus the extent of agentic features in user-bank interactions is limited. However, envisionable advancements include wider adoption of natural language processing capabilities, comparative financial analysis, and customized query suggestions.

Agentic AI systems must operate under a formal specification of trustworthiness constraints. Therefore, AI agents must embody the technical requirements for trustworthy AI systems. Management of risks associated with agentic AI is a dangerous task given the scale of money flows in financial markets and the unprecedented scale, scope, and speed of analysis, prediction, and execution in such markets. At the same time, AI systems that endow agents with entity-level legal ownership and agency create a scarcity that could be captured in trust funds in the form of wealth to protect a material asset class from better prediction by other agents (all other predictions being sub-optimal). Glücksspiel unter Vertrauensbildung para-poker could be a conceptually rigorous game of chance. Thereby, the proposed system could help to promote beneficial forms of AI agency while governing risks effectively.