Bank customers are already using AI programs to search the internet for the best deals on their behalf. And if those programs make bad purchases, sometimes banks get the blame. iStock illustration
The rapid evolution of agentic AI will create a new fraud-prevention quandary for local banks and credit unions: How do you tell a legitimate bot from a criminal one?
Existing fraud-fighting tools are built to look for robot-like activity. If a transaction looks like it is being made by a bot rather than human, it can be flagged as fraud.
But that doesn’t work with artificial intelligence agents a depositor might be using to do their shopping for them.
“It’s not your normal authentication anymore,” Hometown Financial Group Chief Information Officer Sheryl Shinn said. “It’s looking more at behavior, your typing cadence, your mouse movements, cross-channel activity. If it thinks it’s a bot, it’s going to block the transaction.”
Robot Shoppers Becoming Big Business
Agentic shoppers could represent $190 billion to $385 billion in U.S. e-commerce spending by 2030, according to Morgan Stanley.
As the use of agentic AI is rising, banks need to be prepared for a rapid evolution, said Kate Walton, J.P. Morgan Payments’ head of solutions for merchant services.
“It took more than two decades to build the digital commerce ecosystem we see today,” Walton said in an email. “We believe agentic commerce will evolve more rapidly, but it won’t be at scale tomorrow. Like earlier phases of digital commerce, it will require trust by consumers, clear value for merchants, and transparency for ecosystem participants, including payment providers.”
Some institutions already have their own AI agents who allow for transactions to be made. Metro Credit Union’s Chelsea bot allows its customers to move money and make loan payments.
But people are already using AI programs such as ChatGPT, Claude and Gemini as agents to search the internet for the best deals on their behalf, experts say. Additionally, these programs can be used to analyze financial trends and patterns to help users save money and improve their financial health.
But they aren’t foolproof.
“Any LLM that generalizes beyond its training data will either hallucinate invalid outputs or suffer mode collapse, failing to produce the full range of valid responses,” according to a recent research paper by Georgia Tech and OpenAI.
And as with Zelle fraud, when AI agents make mistakes, experts say, customers might blame their bank for not stopping or reversing the transaction – even though they theoretically authorized it by handing their wallet over to that agent.
“A lot of these banks are just concerned that even if they’re not liable, the customers tend to trust their community bank the most,” said Daniel Hartman, a banking law expert at Boston law firm Nutter. “We’re in the very early stages of this new technology, and they are just trying to stay abreast of everything that’s going on and what their customers are looking to do.”
What It Will Take to Manage AI Agents
As agentic AI usage rises, banks must evaluate whether their technology can safely verify these transactions, experts said.
Regulatory frameworks and internal guardrails will also be necessary to protect institutions from errors, Hartman said.
Financial institutions must extend existing fraud prevention tactics to AI-driven transactions. While older systems detect human patterns, they must now identify abnormal AI agent behavior and verify the agents themselves, J.P. Morgan’s Walton said.
“The technology itself is advancing quickly, but broad adoption will depend on participants across the ecosystem aligning on how agents are identified, authorized to act, and held accountable when executing a transaction on behalf of consumers,” she said in an email.
But doing that will be complicated.
“We believe this will require close collaboration among financial institutions, merchants, technology providers, payment networks and industry groups to establish common frameworks that give consumers and businesses confidence in agent-driven transactions,” she said.
Sam Lattof
What’s Being Done So Far
A complicated effort or not, those industry-wide solutions are still emerging.
A national Agentic Payments Alliance has been created to standardize protocols like authorization, Hartman said. Formed by enterprise infrastructure provider Rain, the coalition includes Visa, Mastercard, Fiserv and Remitly.
Additionally, some banks are testing out new tools, such as a trusted contact person who can verify suspicious transactions, said MassBankers President and CEO Kathleen Murphy.
The Mass Bankers trade group has also recently supported the formation of a National Blockchain Network and is a part of the newly formed BankChain Alliance. This national blockchain could further aid efforts to verify fraud, including from agentic AI, Murphy said.
“If there’s a financial transaction involved, there would be rules that would have to be set,” she said of BankChain’s work. “It would be a person that would be setting those rules to determine what conditions have to be met before a transaction can take place. What we’re very aware of is just how quickly the world is changing, and that we want to be there for our members to help them and to create an infrastructure that will be there when they are ready to utilize this technology.”
With how fast AI developments are occurring, staying ahead of the curve can appear difficult. With how much trust consumers put in banks – especially community banks – these institutions will be relied on to put protections in place.
“I think that one of the top risks that regulators – and, I think, lawyers – all agree on is one of the major risks is to do nothing,” Nutter’s Hartman said. “This is just one of those things where customers are going to be dealing with this, and so if you don’t take the bull by the horns, then AI is going to be forced upon you. When you’re in that type of situation, you’re never dealing from a place of strength, and that’s not what banks are known for.”
