Open Agentic Commerce
One of the big fights of the last decade in financial services policy is that of “Open Banking.” The basic premise: you as a consumer should have a right to your own financial data, typically to share with fintech services, rather than having it gatekept from you by the banks.
Something similar is happening today in agentic payments: first big tech companies, and now large banks, are blocking the actions that personal agents can execute as consumer proxies. Amazon is blocking personal shopping agents, and JPMorgan and Citi only enable read-only access of your personal financial data via Plaid (a byproduct of the Open Banking fight!) — while preventing any AI agents from making transactions on your behalf.
As we think through what the future of commerce looks like, the number of AI agents performing tasks on behalf of humans is going to expand exponentially. But that potential will be capped if consumers aren’t actually permitted to use them. This is, at its essence, about consumers reclaiming control over their financial and commercial lives from institutions that have historically gatekept such access.
Open Banking
The Open Banking fight started during the drafting of the Dodd-Frank Act, stemming specifically from a part known as Section 1033: “Consumer rights to access information.”1 This section defined how consumers could have access to their own financial data, establishing statutory consumer-access rights (in other words, banks can’t block you from using your own data). But it went further than just suggesting consumers had to show up themselves to access their data. Per the text, and very relevant to today, the statutory definition of “consumer” reads: “an individual or an agent, trustee, or representative acting on behalf of an individual.”2
Of course, today, those agents need not be human. Congress obviously wasn’t contemplating AI agents when it wrote Dodd-Frank — but technology changing doesn’t obviate the underlying principle: consumers should be able to choose who, or what, transacts on their behalf.
Here’s how it worked back then: early data aggregators like Yodlee, and later Plaid, connected fintechs to consumers’ bank accounts, typically by having consumers share their credentials and then “scraping” account information. (This practice was known as “screen scraping.”) The banks then often tried to block this access, which resulted in those aggregators negotiating bank-by-bank API-access data-sharing agreements, typically at steep cost, so they could get real-time windows into your account balance and ability to pay, and then pass that information onwards to fintechs like Venmo or Mint.
The 1033 rulemaking, long-delayed, was meant to address this. It would formalize the consumer’s right to their data, which in practice would cut down on the fees charged by banks for fintech data access, and expand the set of fintech companies able to offer these services to customers.
Likely because of its pro-consumer bent, 1033 actually garnered bipartisan support, including from then-House Financial Services Committee Chairman Patrick McHenry. McHenry made it a signature piece of his pro-innovation Committee agenda to push the Biden Administration, specifically then-CFPB Director Rohit Chopra, to finalize the rule.3 The central tenet of the 1033 fight was that consumers, not banks, should determine which third parties could access their financial data on their behalf.
This bipartisan support underscored: in order for innovative fintechs to serve consumers better product experiences, banks needed to be pushed to discontinue gatekeeping data that wasn’t actually theirs.
The Rise of Agentic Commerce
Enter personal AI agents. There have been myriad advances in AI over the past two years, and those advances are only increasing in pace. But one of the biggest shifts legible to the average person is the advent of agents, persistent AI-powered programmatic actors that can execute tasks on your behalf.
For example: Hermes Agent, by Nous Research – an agentic harness atop open-source models, with which you can do something as simple as organize your email, or as complex as run a small business. Consumer agents like Instinct, and Meta’s Muse, meanwhile, are set to execute less complex and more “normie” tasks like shopping for groceries, or watching three different sites where a winter coat is on sale, and purchasing it from whichever offers the best price.
The threat agents potentially pose to incumbent businesses is not just that they might lose their relationship with consumers: agents, in practice, can dramatically reduce the friction that otherwise protects incumbents from competition. Most consumers compare credit card or loan rates, or insurance policies, on places like NerdWallet, but don’t spend hours deep in diligence – and certainly not routinely, in real time, as rates adjust. An agent, meanwhile, can monitor all of these in persistent fashion for a consumer, and take action on their behalf to optimize their savings or cashflow, at negligible cost and trivial effort. Consumer bargaining power potentially expands enormously in an agentic world, as a result.
It is here that agents are starting to hit a wall: while there are many companies like Stripe and Shopify that have made it easy for agents to transact, there are others, like Amazon, that are blocking them outright. If you set your Muse to purchase a carton of eggs on Amazon, it will be stopped at the login screen, and any attempt for it to enter your login credentials will be met with an error message, and even an account freeze.
Similarly, your Muse (via Plaid linkage) can tell you when your Chase credit card has a balance that needs to be paid. But should you direct your Muse to then pay the card, it is blocked from doing so by JPMorgan.
Amazon and JPMorgan want to maintain their direct relationship with a consumer, mine data about their behavior in the case of Amazon, and keep their funds within the four walls of the bank in the case of JPMorgan.
A consumer meanwhile may wish to retain a merchant relationship, while merely choosing a different intermediary (their agent) to manage it. This is where things really resemble the motivators behind open banking, as adopting a better tool or service interface should not necessitate abandoning the underlying provider, and a large incumbent should have to justify preventing a customer from using an agent to perform an otherwise-permitted task.
The Current Landscape
For the full potential of agentic commerce to be realized, we need standards set that enable agents to execute tasks on behalf of consumers, while preventing large incumbents from gatekeeping access from those very consumers they’d otherwise serve.
Tempo, Visa, and Stripe have built the Machine Payments Protocol (or MPP), enabling agents to transact in stablecoins or credit cards. Concurrently, Stripe, Meta, and Sierra have built the Personal Agent Protocol (or PAP), which provides a way for agents to interact with websites and execute actions. Together, these protocols enable agents to access a website via a PAP API, decide what they want to buy, and check out, with the transaction executing via MPP or another payment gateway like Stripe Link.
But the Personal Agent Protocol is opt-in – and right now, walled gardens like JPMorgan or Amazon have little incentive to opt in and participate, barring negotiating some outrageous take rate, and are content to continue using their market power to gate access with impunity. As with open banking, nominally permitting access doesn’t mean much if incumbents can impose fees or restrictions that are effectively prohibitive for competitors. Further, should a consumer wish to use an agentic tool that has not entered into one of these protocols or a direct relationship with a vendor, they will be similarly constrained. The current setup both limits consumer choice of vendor, and potentially consumer choice of agent – which leads, ultimately, to higher prices for consumers.
Banks and merchants have real and legitimate concerns about agents becoming possible vectors for fraud or unauthorized transactions. A faulty or compromised agent could execute transactions a consumer never authorized or take actions that might otherwise compromise a platform’s security. Merchants could be left to handle costly returns or bear liability for actions they had little ability to prevent.
But those are arguments for building portals for secure, permissioned access, not for prohibiting agents altogether. Open Banking again offers a useful lesson here, in that secure, standardized interfaces are preferable to forcing third parties into credential sharing or screen scraping. Mercury, for example, takes the approach of allowing agents to prepare payments, while requiring the customer to approve and execute them.
Further, JPMorgan can say that they’re developing their own agents.4 But those agents will be internal to JPMorgan, and will have little incentive to recommend a better rate at a rival bank. So too with Amazon: an Amazon agent might be able to find the best price for a good on Amazon, but that might not be the best price available to the customer across the entirety of the internet.
What Should Be Done
While the key for technological proliferation is rarely to go running to the government (though exceptions exist! See: the GENIUS Act), this is a place where some basic guidance from the FTC (perhaps in conjunction with CAISSI at the Commerce Department, the CFPB, and the banking regulators) could serve a positive role in helping the market to coalesce around open norms before incumbents obtain sufficient lock-in, freezing the current landscape to their advantage. Such FTC guidance could establish principles distinguishing between legitimate agent-access restrictions and potentially exclusionary practices for commercial actors. The Office of the Comptroller of the Currency (OCC) and National Credit Union Administration (NCUA) could then follow with guidance tailored to banks and credit unions. Dodd-Frank Section 1033 itself offers a model here, directing regulators to coordinate while avoiding regulations that unduly favor any particular technology.
We should carry over the consumer-choice principle underpinning Section 1033’s definition of an agent or representative.5 And we propose these potential practical criteria, drawn from the 1033 fight:
At initiation, a consumer establishes an agent’s authority, and imbues it with appropriate credentials, as well as task parameters.
Objective safeguards then determine whether that agent can exercise its authority.
Platforms and agent providers should operate under a clear liability framework, distinguishing between unauthorized actions, failures of platform security, and transactions authorized by properly authorized agents. Platforms should not automatically bear liability for mistakes made by independent agents acting within a consumer’s authorization.
A dominant commerce platform or bank should explain restrictions on covered tasks, and offer a way for agents or users to challenge them. The agent performing the task, similarly, must have a way to meet that standard.
A dominant commerce platform making an exclusive partnership the only route an agent can take to perform a task should be scrutinized, and pricing comparison performed between baseline and that channel.
Any interactivity between agent and platform is bounded to that interaction; it is not a blanket permission for the platform to collect data wholesale.
With standards like these established, agents can make decisions that better serve the consumer, and will index on things like quality or price, rather than being funneled into a constrained set of options.
Open Agentic Commerce
The push for Open Banking resulted in better experiences and lower prices for consumers. As agents become principal commercial actors, it is imperative that we lay the groundwork now so that an agent acting on your behalf is equally capable – if not more so – of delivering value and cost savings to you as you would be acting on your own.
The underlying principle in Open Agentic Commerce is the same one that animated the Open Banking fight: consumers, not the institutions they do business with, should decide which tools act on their behalf.
The alternative is agents constrained to a functional sandbox of commercial lock-in, where incumbents use their market power to maintain their market position.
America has an opportunity to rewrite and update the rules of commerce. Let’s not squander it.