When an AI agent completes a fraudulent transaction, consumers say the AI company should be held accountable. In a recent survey of 2,026 U.S. adults, 65% of respondents placed responsibility on the company behind the agent. That’s a reasonable instinct, since the company built the agent, but banks and merchants still handle the disputes, and they’ll need a record of what the agent was authorized to do.
Axios recently reported that Morgan Stanley started blocking third-party agents while JPMorgan, Bank of America and Citi are placing restrictions on their use. Banks struggle to tell whether a customer authorized a given agent, and a shared login makes that harder because the credentials look exactly like the customer’s. Blocking is defensible, but each legitimate customer turned away is a false positive that can cost the business a sale.
Adoption Stops at the Wallet
The same U.S. consumer survey found that 68% of respondents use an AI agent at least once a week for a personal task. Even with emerging habits, routine use has yet to translate into a willingness to hand over money. Half of respondents (50%) are uncomfortable with an agent using their credit card details to complete a purchase, even if they requested it, and a further 32% won’t share any information with an agent at all. That discomfort ranges from 65% among Boomers to 34% among Millennials.
In addition, willingness to share data decreases as it becomes more sensitive. Location (38%), purchase history (34%) and calendar access (34%) lead the list of what respondents would share with an agent, while only 15% would share financial information. That spread argues for controls that scale with the sensitivity of the action, which a blanket block can’t do but a scoped permission can.
Of the 885 respondents who gave an opinion, 14% pointed to themselves, 11% to the merchant and 10% to their bank or card provider. A customer who blames the AI company may still contact the bank or merchant that processed the payment, and that team has to determine who initiated the transaction. A shared password gives it little to go on.
The Customers Can’t Be the Check
Often, a confirmation prompt asks the customer to judge an agent’s request, and 43% of respondents say they lack confidence in their ability to tell whether a request for information or account access is legitimate. Their worries point at the agent.
A scammer gaining access to the agent or their accounts, and the agent sharing personal or financial information without permission, each topped the list at 20%. Controls that rely on the customer catching a bad request will miss a large share of such requests.
Scoped Authority Beats Blocking
Similar to how a valet key allows a driver to access a car without opening the trunk, RFC 9396 extends OAuth so that a client can request fine-grained permissions rather than a broad scope. For example, a payment of a stated amount can be made to a named merchant from a specific account, with certain permitted actions, such as initiating or canceling said payment.
Permissions like this give the banks something to compare against. Any agentic session that fits the granted (and limited) scope can proceed without friction. Anything that steps outside or arrives with identity, device and behavioral signals that don’t match the customer’s history receives a closer look before money moves. Banks can also record which agent acted and under what authority, so that an investigator can later reconstruct a disputed transaction.
Recourse Backs Up Controls
Asked what would increase their trust in agents, half (50%) pointed to human support when something goes wrong, and 42% preferred the ability to reverse mistakes. Strong identity verification followed at 39%, and 23% said that nothing would increase their trust.
Support and reversal are operational capabilities, so fraud teams will have to build them with customer service and payments operations rather than alone. A bank that can scope an agent’s authority, record what it did and reverse a mistake quickly answers the question customers are already asking. Consumers have already named who they’d blame, but the records of what an agent was allowed to do will shape how each dispute ends.
Join our LinkedIn group Information Security Community!
——————————————————-
Click Here For The Original Source.
