Agentic Banking Creates Two-Sided Trust Problem
In an evolving landscape of banking, artificial intelligence (AI) agents are positioned to play pivotal roles in transactions. Financial institutions are increasingly employing these agents for tasks ranging from customer service to fraud detection, while customers themselves are utilizing AI to conduct transactions, negotiate deals, and manage payments. This dual application of AI has given rise to a critical challenge for banks: determining the appropriate level of authority to grant AI agents and establishing a robust framework for trust as machines begin to interact with one another more frequently.
During a recent address at the Sibos conference in Miami, Federal Reserve Board member Christopher Waller raised a pertinent question regarding the implications of these AI interactions: "Who is on the hook if an agent makes the wrong purchase?" Waller underscored trust as a significant hurdle in the realm of agentic commerce, identifying three core issues that need to be addressed—authentication, liability, and fraud.
As the tumultuous landscape continues to evolve, banks are actively adapting their strategies to address these concerns. AI agents have already begun to permeate the internal operations of financial institutions, and customer-facing agents are starting to gain traction. Testimonies from banks, payment networks, and regulatory bodies reveal the speed at which this dual-agent model is taking shape.
Inside the Bank: Agents on the Payroll
A notable development occurred in late September when Kris Brewster, interim CEO of LHV Bank, a U.K. challenger bank, announced via LinkedIn that the organization had launched its agentic AI project primarily aimed at managing customer emails, an initiative powered by Gradient Labs. Brewster emphasized the importance of developing comprehensive guardrails over rapid deployment, stating, "In a bank, the guardrails matter more than getting live quickly." Recognizing the potential challenges, he characterized the project as an exploratory phase without revealing immediate results.
From its inception, LHV Bank approached the implementation of agentic AI with a cautionary mindset, committing to a controlled environment focused on key attributes such as explainability, auditability, and accountability. By consciously prioritizing human oversight, the bank aims to safeguard customer interests and promote a responsible deployment timeline—an estimated six months for a regulated institution.
During JPMorgan Chase’s second-quarter earnings call on July 14, CEO Jamie Dimon highlighted the bank’s extensive integration of AI, reporting nearly 1,000 use cases, although he specified that around 50 are particularly impactful, encompassing areas such as risk assessment, marketing, fraud detection, and documentation. Dimon also pointed out the ongoing retraining of staff, as certain positions experienced a reduction of 30% to 40%. However, it was announced that most affected employees were offered alternative roles within the organization. His counsel to CIOs was notable: banks do not exclusively benefit from AI, as competitors also leverage the same technology, thereby emphasizing that the ultimate beneficiaries should be the customers.
Additionally, JPMorgan’s CFO, Jeremy Barnum, provided insight into the bank’s fiscal approach, noting that while expenditures on AI tokens were minimal in the first half of the year, a significant uptick is forecasted for the latter half.
In a distinct paradigm, BNY Mellon has adopted a model wherein AI agents are thought of as colleagues rather than tools. In an OpenAI case study, the bank illustrated its use of "digital employees" equipped with identities and specific workflows. These agents can perform tasks that range from payment validation to enhancing code security. Chief Data and AI Officer Sarthak Pattanaik highlighted how this evolution allows human staff to focus on training and supporting these digital entities.
At the Counter: Agents With a Card
The complexity of identity and accountability deepens when considering customer-facing agents. On September 21, Danske Bank and Mastercard jointly announced what they referred to as Denmark’s inaugural AI-enabled payment transaction. A customer tasked an AI agent with reserving a coffee tasting session through Mastercard’s Priceless.com platform. The agent not only facilitated the booking but also managed the payment via a Danske Bank card. While this initial instance was singular rather than a widespread deployment, it demonstrates a fundamental principle: the significance of customer trust and control over transactions.
Waller’s insights bring to light the necessity of maintaining trust in AI interactions. He classified the operational dynamics into two frameworks: the agent-assisted model, in which a buyer uses an agent to explore options before making final decisions themselves, versus the agent-delegated model, where a buyer permits an agent to handle purchasing and payments on their behalf. Given the inherent risks of unintended purchases within the delegated framework, Waller emphasized the urgent need for stringent trust mechanisms and safeguards.
He also touched upon significant shifts required in authentication processes—transforming the focus from merely verifying buyer authorization to confirming whether an AI agent holds the requisite permissions to execute payments. Traditional fraud detection systems, tailored for human behavior, may not function optimally in this new context, necessitating adjustments to compliance rules and models.
Waller also indicated that the corporate payments sector should be a targeted area for deploying such technology due to its nature of recurring transactions within pre-established parameters. Nevertheless, he cautioned that larger transactions elevate risks, hence stricter controls and monitoring are essential.
The Caller Nobody Planned For
These challenges extend beyond transactions. At the Sibos conference, Johan Gerber, Mastercard’s executive vice president for network products and real-time payments, shared an intriguing incident involving a bank’s call center that received a request from an AI agent looking to renegotiate the terms of a customer’s credit card. The absence of a standardized procedure for managing such requests posed a new challenge—one the bank did not anticipate.
Gerber pointed out that banks and card networks might not have the luxury of choosing whether to engage with these interactions, especially when customers can share their credentials with AI agents. This evolving trend signifies a transformative shift in how banks may need to operate.
Three Questions for Leadership Teams
The rising prominence of AI agents in the banking sector necessitates leadership teams to engage with three critical questions:
-
Can you distinguish an agent from a person? Waller insists the conversation should embrace identifying the authority assigned to agents rather than focusing solely on the identity of the payer, emphasizing that organizations need systems capable of verifying an agent’s representation.
-
Which actions can you reverse? There are risks associated with allowing automated systems to make irreversible payments without adequate supervision. Institutions must decide which functions require human intervention and which can progress autonomously.
- What will it cost, and who is accountable? As highlighted by Barnum and Waller, understanding budgetary implications and determining accountability is essential prior to scaling pilot projects.
Across the board, banks are forging ahead with AI agents. From LHV’s implementation for customer communications to JPMorgan’s extensive AI use cases, BNY’s collaborative approach, and Danske Bank’s trailblazing AI payment transactions, the landscape is rapidly transforming. Given these dynamics, Waller’s query about accountability in the event of errors remains highly pertinent. The banks that can preemptively address these responsibilities will likely harness the benefits of AI most effectively.