An AI assistant that finds a product and an AI agent that pays for it create different problems for the businesses handling the transaction. Federal Reserve Governor Christopher J. Waller put that distinction at the center of a September 29 speech at Sibos in Miami, as the payments industry considers how automated shopping could develop.
In his published remarks, Waller distinguished purchases where a person retains control from those where an agent is authorized to act. He described agentic commerce as early-stage and identified authentication, liability and fraud as central obstacles to trust. The speech raises questions for the industry; it does not announce a new rule assigning responsibility for an AI agent’s mistakes.
For merchants and payment providers, the commercial issue is more specific than whether a model can recommend a good product. They need to know whether the action presented at checkout matches the authority the customer actually granted.
A budget is only one part of permission
Imagine a hypothetical office authorizing an agent to reorder a particular printer cartridge for no more than $80. A purchase of a different cartridge for $60 meets the price limit but fails the task. A correct cartridge arriving after an important deadline may also fail the intended purpose.
A useful design review would therefore ask how the system represents product requirements, approved sellers, delivery needs and permission to substitute. It would also ask what happens when those requirements conflict. Returning to a person for a decision may be a successful outcome, even if it reduces the number of transactions completed without intervention.
That changes how a pilot should be measured. Counting completed purchases alone can reward a system that quietly makes too many assumptions. A better business test would include whether the purchase met the original request and whether exceptions were handled in a way the buyer could understand.
Keep evidence of the decision
Waller highlighted the difficulty of determining who is responsible when an agent makes an incorrect purchase and the possible role of standards recording buyer intent. For a company evaluating a system, that suggests asking for an understandable transaction record before expanding access.
In the hypothetical cartridge order, the record should let a reviewer connect the instruction, selected item, final price and authorization. A record that merely states that an agent approved the transaction would leave the most important question unanswered: approved according to which instruction?
This is an operational question as well as a dispute question. Customer-service staff need enough information to explain an order without reconstructing a long technical conversation. Buyers need a clear route to pause future activity while a problem is investigated.
Adoption depends on both sides of checkout
A retailer can make its catalog easier for automated tools to read while still requiring explicit approval at payment. Those are separate design choices. The value of improved discovery should not automatically be treated as evidence that delegated purchasing is ready for every customer or transaction.
Waller’s speech points to an industry still working through trust and standards. For businesses, the useful test is whether an automated purchase remains understandable and controllable when something goes wrong. Faster checkout has limited value if resolving one mistaken order consumes the time saved by many correct ones.
Our coverage of Microsoft’s agent-testing skill explores another part of evaluating automated systems: evidence that a safeguard works.
Illustrative conventional payment equipment; the parking meters pictured do not demonstrate AI-agent purchasing. Photo: Raphael Loquellano / Pexels.