Six banks propose disclosure and audit-trail rules for AI shopping agents, ahead of any regulator

Six banks -- ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest -- published a joint paper on Sept. 22 called "Building Trust in Agentic Commerce," proposing five voluntary principles for the AI agents now shopping and paying on customers' behalf. The paper defines agentic commerce as "the use of AI agents to help make or facilitate payments between a consumer and a merchant," and is explicit that it "does not prescribe conduct, any commercial position or any particular method or timetable for implementation" -- this is a position paper, not a standard, and the banks say a second, more technical paper is coming to translate it into actual protocols.

The paper's own account of why it exists is unusually candid for a financial-industry document. "Consumers are unclear if AI agents will act in their interests," it says. "They are concerned that AI agents may buy the wrong thing or spend too much -- or even worse, lose their money to scams and fraud." It names specific unsafe practices already happening: agents that request a customer's card details and enter them directly into websites, and agents that steer purchases toward payment methods carrying weaker fraud protections. Merchants get a matching worry -- rising disputes and chargebacks "for reasons outside their control," with no clear line on who is liable when an agent, not a person, made the purchase decision.

That distrust already shows up in usage data from both sides of the transaction. British retailer John Lewis reported that AI-originated shopping searches grew from 0.3% to 2.5% of its traffic over the past year -- an eightfold rise, and one the retailer says is accelerating. On the willingness side, a separate PYMNTS Intelligence survey found just 24% of Americans say they'd let an AI agent both shop *and* pay, against 22% who'd let one merely start the research and roughly half who report using AI assistance in a purchase in some form. That's a genuinely different data point from the banks' own paper -- PYMNTS's survey measures consumer willingness, the banks' paper describes provider risk -- and the gap between them is itself the story: usage is climbing while trust to hand over the actual payment step is not.

That ladder, drawn directly from the paper, is the clearest evidence that this document is preemptive rather than reactive: the banks place most of today's actual shopping agents at or before the "human reviews the final purchase" stage, with the fully autonomous end of the scale still mostly theoretical. The five principles -- transparency, safety, privacy & data, choice, and interoperability -- are each written as a hedge against the risks that specifically appear as autonomy increases, not the risks of AI shopping tools as they exist today.

“Customers need to trust that they remain in control of how payments are made and that their money is safe.” — Mark Brant, Chief Payments Officer, NatWest

Brant's framing is also a fair description of who actually carries the risk while these principles remain voluntary and the follow-up implementation paper is still unwritten.

The paper arrives without any binding force behind it, which is exactly what distinguishes it from the one piece of US federal legislation aimed at the same problem. Sen. Mark Warner's discussion draft of the AI AGENT Act -- released June 29, not yet formally introduced -- would go considerably further: it would require platforms to let authorized AI agents connect on functionally equal terms to a human user, impose non-waivable fiduciary-style duties on agent providers (safeguard data, avoid conflicts of interest, follow instructions), and make the FTC the principal enforcer, including a registration framework for agent providers and civil penalties for violations. Where the banks' paper is voluntary and industry-authored, Warner's draft would be federal, mandatory, and enforced -- and it explicitly names agentic commerce and account management as covered activity. Neither document references the other, but they're answering the same gap from opposite directions: one from the institutions that move the money, one from the government that could compel them.

The immediate test of whether "Building Trust in Agentic Commerce" amounts to more than a position paper is the promised follow-up: a second document meant to turn these five principles into actual protocols and industry standards, with no date yet attached. Until that lands, the paper's own conflict-of-interest warning is worth sitting with on its own terms -- it flags, as a live risk, that an AI agent "may prioritize products, payment methods or other services that deliver the best financial outcomes for their providers" over what's actually best for the customer. That is a risk the paper names in its own agents, built by its own industry, without yet proposing a mechanism to catch it. This newsroom's earlier coverage of Meta's Muse Connector Platform -- and Amazon's decision to block that agent from its own storefront -- is the live version of exactly the access dispute the banks' "Choice" principle is trying to get ahead of.