Summary
Stripe is reportedly exploring a new payment network for AI agents that could reduce reliance on Visa and other traditional card schemes. Yet the larger obstacle is not routing a payment: it is proving user intent, allocating liability, winning merchant acceptance, and creating enough trust for all participants.
The problem Stripe is trying to address
Stripe is reportedly exploring a new payment network for AI agents, with the potential goal of reducing dependence on Visa and other traditional card networks. The idea arrives as consumer-facing AI agents move beyond search and recommendation into execution. Agents can browse websites, fill in forms, compare options, and increasingly attempt to complete purchases on a consumer’s behalf.
That shift changes the nature of a card transaction. In a conventional checkout, the consumer usually selects the item, reviews the terms, and confirms payment directly. In an agent-mediated purchase, the consumer gives a request in natural language. The agent then interprets that request, searches across websites, makes decisions along the way, and may submit the final order without the consumer reviewing every field.
The central issue is that a payment can be authorized by the infrastructure and still be wrong for the person who initiated it. Consider a request for a two-night hotel stay costing no more than $600, within walking distance of a venue, and cancellable without charge until the day before arrival. An agent might select a hotel for $580 in the right location, but book a non-refundable room. The card is valid, the merchant is legitimate, and the amount is within budget. Nevertheless, the purchase may violate the consumer’s actual instructions.
That distinction—between payment authorization and fulfillment of intent—is the core challenge facing agent payments. A new network may change how a transaction is routed, but it does not by itself establish what the user authorized or who should bear the loss when the agent gets it wrong.
Intent is missing from much of the payment chain
The first risk is that an agent may not fully understand the request. Natural-language instructions do not have a standard format. A user may assume that refundability, delivery timing, exchange rights, or exact product specifications are obvious, while the agent may fail to represent those conditions explicitly.
Existing card systems are generally designed to evaluate a payment using information such as the account, merchant, amount, and risk signals. They do not necessarily carry the complete set of conditions behind a purchase. An issuer may be able to identify suspicious activity, but not know that the consumer required free cancellation. A merchant may know what was ordered, but not the constraints that led the agent to select it. A payment service provider or acquirer may have to assess the transaction without access to either side’s full context.
The second risk appears even when the consumer’s instructions are clear. The agent can still deviate during execution. Possible causes include hallucinated information, outdated listings, lost context in a multi-step workflow, misread checkout options, software bugs, or malicious instructions embedded in a webpage. A human approval step can reduce some of these risks, but it also limits the convenience that makes autonomous agents attractive. Allowing an agent to act independently improves automation while increasing the consequences of an execution error.
Every participant needs a different form of assurance
Consumers need confidence that an agent will purchase the intended product or service, protect payment credentials, and provide a clear remedy when something goes wrong. The source analysis cites an NMI survey of 1,000 U.S. adults in which 69% said they did not trust AI to handle payments safely, 70% wanted to review or veto an AI decision before an order was placed, and only 10% were willing to give AI complete control over purchases. These figures do not establish a universal market view, but they illustrate the trust gap that agent commerce still has to close.
Merchants, PSPs, and acquirers face a different problem. They must distinguish legitimate shopping agents from malicious bots and defend themselves when a transaction is later disputed as unauthorized or inconsistent with the customer’s instructions. Risk providers can evaluate signals supplied by merchants, but those signals are richer when the consumer is shopping directly on the merchant’s website or application. When an agent moves across multiple sites, the relevant context may be fragmented or unavailable to the merchant.
Issuers face a similar information deficit. They must decide whether to approve a transaction without seeing the full user request, and later determine whether a dispute resulted from account fraud, an agent’s mistake, or a purchase the consumer genuinely authorized but later regretted. Without reliable evidence of delegated authority, issuers have limited ability to separate these cases consistently.
Card networks, meanwhile, would need rules for delegated commerce. Those rules could cover how authorization evidence is transmitted, what counts as a failure to follow instructions, and how responsibility is allocated among the consumer, agent provider, merchant, acquirer, and issuer. If every party sees only one part of the transaction, disputes become a contest over incomplete records.
The risk of a negative network effect
These gaps can reinforce one another. If consumers do not trust agents, merchants have less incentive to support them. If merchants provide limited context, PSPs and issuers may apply stricter controls. Lower approval rates and more interrupted checkouts then make agent shopping feel unreliable, further reducing consumer willingness to delegate purchases.
This is a negative network effect: authorization uncertainty discourages acceptance, limited acceptance weakens the user experience, and a poor experience prevents the market from reaching the scale needed to justify broader infrastructure investment.
Agent companies also need a sustainable way to monetize transactions. Facilitating a purchase does not by itself establish who should pay the agent or what obligations the agent has assumed. Affiliate arrangements and merchant commissions are frequently proposed, but they require agents to establish relationships one merchant at a time and manage different terms, tracking systems, and settlement processes. The source analysis cites OpenAI’s Instant Checkout as an example of the difficulty: launched in September 2025, it had by March 2026 narrowed back toward product discovery and merchant-owned checkout. The example is presented as evidence that a new interface does not automatically solve the commercial and operational problems of agent checkout.
Merchants may also question the economics. If they must pay an affiliate fee while taking on disproportionate fraud, dispute, and fulfillment risk, their incentive to accept agent-originated transactions may remain weak. A new network therefore has to offer more than another route for submitting payment data. It must create a credible framework for permissions, evidence, risk sharing, and merchant value.
What a new network would have to prove
If Stripe advances this exploration, its meaningful test will not be whether it can remove certain intermediaries from the payment path. The harder test is whether it can make user intent legible and enforceable across the participants in a transaction.
A workable system would likely need structured representations of consumer constraints, clear confirmation rules, auditable records of an agent’s decisions, and mechanisms for pausing or challenging a transaction. It would also need to give merchants, PSPs, acquirers, issuers, and dispute teams enough context to evaluate what happened without exposing unnecessary personal information.
This is closely related to the concerns of institutional wallets and custody operations: who may initiate a transaction, what limits apply, which conditions must be satisfied, and how authorization evidence is preserved across a multi-step workflow. Automation does not remove the need for approvals, segregation of permissions, monitoring, and audit trails. In many cases, it makes those controls more important.
Stripe’s reported exploration therefore points to a broader infrastructure question. Agent payments are not simply traditional card payments with an AI interface placed on top. They require a shared language for intent, delegated authority, execution evidence, and liability. Until those elements are accepted by users and supported by merchants and financial institutions, a new payment rail may improve transaction mechanics without resolving the deeper coordination problem.
The market’s deadlock is structural: every participant can reject the risk, but no participant yet has a complete view of the consumer’s intent. Breaking that cycle will require cooperation across the payment stack, not only a replacement for the existing network.
Source: link