Banks Grapple with Trust and Liability as AI Agents Transact with Each Other
Financial institutions are deploying AI agents internally while customers use their own for transactions, creating a two-sided trust problem where current regulations lag behind.
The increasing deployment of artificial intelligence agents on both sides of financial transactions presents a complex new challenge for banks: establishing trust and accountability when machines increasingly interact with other machines. As financial institutions leverage AI for internal operations like customer service and fraud detection, customers are simultaneously gaining access to agents capable of making purchases and negotiating on their behalf. This convergence necessitates a re-evaluation of established security protocols and regulatory frameworks.
Federal Reserve Board member Christopher Waller recently highlighted this emerging dilemma, posing the critical question: "Who is on the hook if an agent makes the wrong purchase?" He identified authentication, liability, and fraud as the primary obstacles to widespread adoption of agentic commerce. While the regulatory landscape remains uncertain, banks are actively exploring and implementing AI agents, signaling a rapid shift in the operational paradigm.
Internally, banks are cautiously integrating AI. LHV Bank, for instance, has deployed agentic AI for customer emails, emphasizing guardrails for explainability, auditability, and clear accountability. JPMorgan Chase is exploring nearly 1,000 AI use cases, with a significant focus on risk, fraud, and operational efficiency. CEO Jamie Dimon noted that while AI adoption offers little unique competitive advantage for banks, it necessitates retraining staff and may lead to job reductions, with the primary benefit accruing to the customer.
BNY Mellon offers a different approach by treating AI agents as "digital employees" with defined identities, access controls, and specific workflows. This model emphasizes the evolving role of human employees in training and supporting these AI counterparts. The common thread among these initiatives is a gradual, documented approach to AI deployment, with clear definitions of agent capabilities and responsibilities.
However, the complexity escalates when AI agents operate on the customer's side of the transaction. A pilot program by Danske Bank and Mastercard demonstrated an AI agent handling a payment reservation and transaction on behalf of a customer, highlighting the need for secure, verifiable transactions that maintain consumer control. This scenario amplifies the trust and control issues raised by Waller, particularly in agent-delegated models where agents are authorized to shop and pay autonomously.
The core challenge lies in adapting existing frameworks for authentication, fraud prevention, and liability to a scenario where the parties involved in a transaction are not always human. Banks must define clear protocols for agent verification, establish mechanisms for recourse in case of errors, and determine how to attribute responsibility when an AI agent causes financial loss. The current regulatory environment, designed for human-to-human or human-to-system interactions, is ill-equipped to handle machine-to-machine financial dealings.
As agentic AI becomes more prevalent, the financial sector faces a critical juncture. The development of robust security measures, clear legal frameworks, and industry-wide standards will be essential to foster trust and enable the safe and effective integration of AI agents into the banking ecosystem. Without these, the potential benefits of AI in finance may be overshadowed by the risks of unprecedented fraud and liability disputes.