Six banks are calling for clearer dispute processes and liability allocation in agentic commerce as AI agents are given greater authority to complete purchases. In their “Building Trust in Agentic Commerce” principles paper, the banks say liability can become unclear when an AI agent exceeds the authority a customer gave it.
ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest Group published five voluntary principles covering transparency, safety, privacy and data, choice and interoperability. The banks said agentic commerce has already moved from closed trials into live pilots and early scaling while standards, policies and consumer protections remain under development.
The paper describes different degrees of autonomy, from agents that help consumers search while a human completes the purchase to agents that can identify a need and buy on their own. The banks said more autonomy can create mismatched expectations between consumers, agents and merchants, including uncertainty over liability when an agent exceeds its authority.
Payment rail changes the recovery path
Bridget Hall, director of account-to-account payments for the Americas at ACI Worldwide, told TechInformed that the first question is whether an agent stayed within the authority the customer gave it.
“Dispute process and agentic guardrails will need to be better defined as agentic commerce becomes more common,” she said.
How the agent pays can then affect recovery, she said. According to Hall, credit card systems already have processes for handling unauthorized charges, although the industry has yet to determine how those processes would apply when an authorized agent goes beyond a customer’s instructions. The Federal Reserve’s FedNow service and The Clearing House’s RTP network, the two U.S. instant-payment networks, work differently because settled payments are irrevocable.
“The customer’s bank can ask for the money back, but the receiving bank is not required to return the money,” Hall said. She cautioned that the industry still has more to determine when an agent payment goes wrong, regardless of the payment rail.
Banks may not see what the agent was told
Resolving the dispute also depends on evidence that may sit outside the payment system. The banks’ paper says agentic commerce can generate conversational prompts, decision logs, intent mandates and purchase details, and calls for auditable records showing consumer instructions, authentication, intent, transaction decisions and outcomes.
Hall said banks may need the customer’s original instructions, price limits, merchant requirements, product details and a record of what the agent did. ISO 20022 is making payment messages richer, she said, but that level of context “is beyond what exists within a payment today” and may remain inside the system where the agent is operated.
The five principles are voluntary and nonbinding, and the banks said a later paper will address implementation through protocols, standards and policies.