US banks and payment firms are expanding the use of agentic AI in payments, account opening and other financial services. But a central question remains unresolved: who is responsible when an AI agent makes the wrong transaction, is manipulated into fraud or acts beyond a user's authorization?
The answer affects whether consumers can recover lost funds and how banks, payment networks, merchants, technology providers and fintech platforms divide operational risk. Bain estimates that annual US spending on agentic commerce could reach $500 billion by 2030. Clear liability rules may be essential as transaction volumes grow and businesses consider wider deployment.
Waller flags liability for an agent's wrong purchase
Federal Reserve Governor Christopher Waller addressed the issue at the Sibos conference on September 29. He said broader adoption of agentic commerce—particularly when AI makes decisions on a consumer's behalf—depends on trust between buyers and sellers.
Waller outlined a straightforward example: a consumer asks an agent to buy a sweater, but the agent orders the wrong size and the bank approves the payment. Payment authorization does not mean the agent bought what the consumer intended. Existing e-commerce frameworks, including payment-network rules, consumer protections and relevant regulations, could be adapted for agent-led transactions, he said, while regulators could also test new approaches.
The risks are not limited to hackers or fraudsters manipulating an AI system. An agent may stray from its prompts or settings, and ordinary mistakes by consumers, merchants or others in the transaction chain can also cause problems. If an agent transfers money outside an account, responsibility could fall on a user who configured it incorrectly, a model provider that allowed it to exceed its authority, or a bank that approved a transaction despite warning signs.
Lead Bank proposes a liability waterfall
Kansas City-based Lead Bank, a state-chartered bank that provides banking-as-a-service infrastructure to fintechs, digital-asset companies and consumer brands, is developing a framework for agent-led banking and payments.
CEO Jackie Reses said traditional identity and verification systems were built around people and often designed to exclude bots. That approach is changing as machines outnumber people and operate at far greater scale. Transaction errors are inevitable, she said; the priority is to establish controls, programming requirements and a framework for assigning responsibility when an agent goes off course.
Lead Bank's proposed liability “waterfall” would allocate responsibility according to the type of failure and which party had control. If an agent exceeds the authority a consumer explicitly granted, the consumer should not be liable. If a bank controls the allocation of permissions but fails to enforce the limits, the bank should bear responsibility. AI providers would be responsible for problems such as model defects, unsafe tool calls or inadequate safeguards. A merchant could be liable if it knowingly encouraged an agent with limited permissions to exceed them.
The approach aims to build responsibility into systems before a dispute occurs rather than settle every case afterward. In practice, the outcome would still depend on how permissions were configured, whether the bank received signals of unusual activity and whether each party can show it controlled the relevant part of the process.
Primitive develops a protocol for agent-led finance
Primitive, a technology startup founded by former US Bancorp chief digital officer Derek White, is also working on operational and liability standards for agentic AI. White said consumers are already using models such as ChatGPT to find financial-service providers and may eventually ask agents to open accounts at designated banks.
One proposed standard is the Agentic Financial Services Protocol, intended to help companies build AI agents to shared safeguard requirements. White said it calls for a record of each agent session, including whether the agent followed governance rules, which large language model it used and how many tokens the session consumed.
The records could be compiled for regulators or auditors, making it possible to trace how an agent reached a decision. For banks, a detailed audit trail could also help establish whether a problem stemmed from a user's instructions, model behavior, a tool call or the transaction-approval process.
Controls exist, but regulatory alignment remains open
Reses said current technology can constrain agent behavior and support accountability. Lead Bank's programmable interfaces can monitor deviations, record system attestations, periodically update “know your agent” information, issue transaction credentials and revoke them when needed.
The larger challenge, she said, is aligning banking and payments rules with agentic AI activity. Without clearer legal and regulatory expectations, banks may struggle to determine which controls satisfy their obligations, while consumers may not know in advance who will reimburse them if a transaction goes wrong.
Card networks, e-commerce platforms, payment providers, technology companies, the FIDO Alliance and banks are all involved in developing standards. Consumer concerns about using AI agents to pay remain, even as market estimates point to potential growth. As more agents gain the ability to search, open accounts and make payments, financial institutions will need to define how permissions are recorded, transaction credentials are managed, models are audited and responsibility is assigned when an agent exceeds its authority.