Banks Want Agent Commerce Rules

Who writes the rules for agent commerce when no regulator has?

On September 22, six banks from five countries published a paper called "Building Trust in Agentic Commerce." Bank of America, Capital One, ING, NatWest, Commonwealth Bank of Australia, and ASB Bank jointly proposed five principles they believe should govern how AI agents spend money on behalf of humans. Transparency, safety, privacy, choice, interoperability.

The paper is voluntary and non-binding, with no compliance requirements, no enforcement mechanisms, and no timetable.

Eight days later, Mastercard published its own trust framework with five pillars and an enforceable certification mark. The difference between those two documents is the difference between asking the industry to behave and making the payment rail refuse to process transactions that don't.

What the banks actually wrote

The paper's definition is narrow and deliberate. Agentic commerce is "the use of AI agents to help make or facilitate payments between a consumer and a merchant for a product or service." It excludes peer-to-peer transfers, B2B accounts payable, and machine-to-machine settlement. Consumer checkout only.

Within that scope, the banks identify a spectrum of autonomy. On one end, an agent searches for products and a human completes the purchase. On the other, an agent identifies a need, selects a product, enters payment credentials, and executes the transaction without the consumer reviewing the final choice.

The banks are watching both ends accelerate. Constructor launched Stripe-powered Agentic Checkout in October. Meta and Sierra announced the Personal Agent Protocol the same month. The infrastructure is live and scaling. But the paper notes that industry standards, policies, and consumer protections are "yet to be fully defined or established."

So six issuers wrote down what they think the standards should say.

Five principles, zero enforcement

The five principles read as reasonable and toothless in equal measure.

Transparency asks that all parties know when an agent is involved and on whose behalf. It flags a specific conflict: agents may prioritize products that deliver higher commissions or lower costs for their providers, even when another option offers better value to the consumer. The paper asks for disclosure but stops short of requiring it.

Safety wants consumers to have control over the authority they grant to agents. It calls for secure, auditable methods for entering payment credentials, and says dispute resolution should involve all relevant parties with liability reflecting where the error entered the transaction.

But the paper doesn't define what that liability allocation looks like, and no party is assigned default responsibility.

Privacy and data acknowledges that agentic commerce generates new categories of information: conversational prompts, decision logs, intent mandates, purchase details. The paper says providers should maintain auditable records of consumer instructions, authentication, intent, and transaction outcomes. Enough to reconstruct the path from instruction to purchase when something goes wrong.

Choice addresses platform lock-in. Consumers and merchants should be able to pick their agents, wallets, and payment services without unreasonable restrictions. At the same time, service providers can choose which agentic commerce services they support based on safety, commercial, or legal grounds.

Interoperability calls for common connections on core protections while leaving room for competing features.

Each principle is a sentence that could anchor a regulation, but none of them do. The paper explicitly states: "The principles described are voluntary and non-binding. Nothing in this paper requires the authors or other industry participants to take any particular action."

Mastercard's answer: architecture, not principles

On September 30, Mastercard published "Trust for agentic commerce," introducing the Agent Pay Trust Framework built around five pillars: identity, intent, controls, trusted execution, and intelligence.

The pillars map onto the bank paper's principles, but Mastercard operationalized each one.

Where the banks ask for transparency about agent identity, Mastercard verifies it cryptographically through Agent Pay Ready certification. Where the banks ask for auditable records of consumer intent, Mastercard built Verifiable Intent, a signed record connecting the consumer, their spending instructions, and the merchant interaction.

On the question of controlling agent authority, Mastercard encodes spending constraints directly into Agentic Tokens, including maximum amount, currency, merchant category, and expiration, checked during network authorization. A transaction that violates them never reaches the issuer.

The bank paper proposes that agents should be safe, while Mastercard's framework defines safety at the wire level and rejects transactions that fall outside it.

This isn't a coincidence of timing. Mastercard had been building Agent Pay infrastructure since acquiring BVNK and launching Agentic Tokens alongside the Visa/Mastercard authorization split that emerged in October. The bank paper gave Mastercard a public framework to respond to, and Mastercard responded with the argument it was already prepared to make: trust at the principle level is necessary, trust at the protocol level is what scales.

The liability question the paper sidesteps

The bank paper's safety principle says that "liability outcomes [should reflect] where risks and errors are introduced." It's the closest the paper comes to addressing the central question of who pays when an agent buys wrong.

But it doesn't answer it.

PYMNTS and Visa's Global Digital Shopping Index found that 93% of merchants believe the AI provider should absorb the financial loss for incorrect agent purchases. Merchants see a transaction they didn't control generating a chargeback they can't defend with existing evidence frameworks.

The chargeback gap is already documented. As of November 2026, no card network has published an agent-specific reason code, no dispute process involves the agent platform as a party, and no liability framework assigns default responsibility to the entity that deployed the agent.

The bank paper acknowledges this gap. It says dispute resolution should "involve all relevant parties" and that "any party holding a potential liability [should be] able to require authentication." But it names no party and assigns no default. A subsequent paper, with no announced date, will explore how to translate principles into "protocols, industry standards and policies."

Meanwhile, specific products are already filling pieces of the gap. Chargebacks911 launched AI agent purchase indemnification. Amex announced Agent Purchase Protection in April. These are vendor-specific programs, not industry standards.

What the paper tells you about the banks' position

The six banks that wrote this paper process millions of card transactions daily across their issuing portfolios. They sit between the consumer and the card network. When an agent-initiated transaction goes wrong and no one else absorbs the loss, the issuer does.

The paper is the issuers telling the rest of the value chain: we see what's coming, and we need the protocols to carry more of the safety burden. Agents should identify themselves, intent should be auditable, and disputes should involve the agent platform alongside the merchant and issuer.

Bank of America's Mark Monaco said the industry needs "thoughtful approaches to identity, authorization, fraud prevention, liability management and customer protection." Capital One's Todd Kennedy said his bank is "committed to helping shape it the right way." Both executives work at institutions that will absorb the losses if the ecosystem fails to self-organize.

Principles versus protocols

New payment channels follow a sequence: principles first, then protocols that encode what the infrastructure will actually do, then regulation that enforces what the market must do.

The bank paper occupies stage one, Mastercard's Agent Pay Trust Framework sits at stage two, and regulation, the third stage, doesn't exist yet. No jurisdiction has enacted rules specifically addressing autonomous AI purchasing.

Visa took a different path. Its Trusted Agent Protocol verifies agent identity per transaction but doesn't encode spending constraints into the credential itself. Visa's Intelligent Commerce platform supports multiple protocols, including TAP, Stripe's ACP, and Google's UCP. Mastercard locks policy into the token, while Visa trusts the edges to enforce it.

The banks' five principles align more closely with Mastercard's approach than Visa's. The paper's emphasis on auditable intent records and authentication requirements maps to Verifiable Intent and Agent Pay Ready certification. Visa's per-transaction verification model puts more trust in merchants and issuers, exactly the parties the bank paper says need more data to manage safety risks.

What happens next

The banks promised a follow-up paper on implementation. No date. The paper invites "feedback and engagement with consumers, merchants, AI agent providers, payment schemes, issuers, acquirers, industry bodies, regulators and policymakers."

In practice, the implementation is being written by Mastercard, Visa, Stripe, Google, and the FIDO Alliance through competing and overlapping protocols. ACP handles checkout, UCP covers discovery and post-purchase, and AP2 manages authorization through cryptographically signed mandates. Mastercard's Agent Pay encodes constraints into the credential while Visa's TAP identifies the agent to the merchant.

The bank paper sits on top of all of them. It's the institutional vocabulary for discussing what these protocols should accomplish, without specifying which one should win.

The 93% of merchants who think AI providers should bear the loss aren't waiting for a principles paper. They're waiting for a protocol that enforces liability assignment at the transaction level. The banks know this. Their paper says so, carefully, in language that avoids committing to any specific protocol, network, or timeline.

Six banks named the problem, and the card networks are already building the answer. Whether the banks' principles shape the protocols or the protocols make the principles moot depends on how fast that follow-up paper arrives.