Payment-Method Defaults Will Shape AI Commerce

AI shopping will not remove payment choice. It will move much of that choice into software. When an agent reaches checkout, the available credentials, eligibility rules and ranking logic can determine whether the customer uses a card, bank account, wallet or installment product before a familiar payment page ever appears.

That makes the default a commercial decision. Merchants and payment providers need to define how an agent discovers methods, compares tradeoffs and obtains consent. Otherwise, the easiest credential to retrieve will quietly become the winner.

What the evidence shows

On March 3, Klarna announced that its flexible payment options would soon be supported in AI-agent shopping through Stripe’s Shared Payment Tokens. The planned availability covers eligible U.S. merchants already offering Klarna through Stripe, provided the shopping agent supports flexible payments at checkout.

Stripe described the broader architecture the same day. Shared Payment Tokens are intended to let an authorized agent initiate payment without exposing the underlying credentials. Stripe said it was expanding support beyond cards to network-led agentic payment capabilities and buy now, pay later methods including Klarna and Affirm. Its technical explanation says an agentic network token is scoped to the customer’s intent.

That infrastructure addresses credential security and method availability. It does not, by itself, settle how choices should be ranked. An agent could optimize for acceptance, speed, rewards, financing cost, merchant expense or a saved customer preference. Those objectives can point to different methods.

Klarna’s own tokenization documentation makes one control explicit: saving a method for future use requires a step-up flow and customer consent. The hard operating question begins after secure access is possible: which method should the agent propose for a particular purchase, and why?

The operating consequence

Defaults are distribution. In a conventional checkout, the shopper can scan several methods and revise a choice. In an agent flow, the software may reduce that menu to a recommendation or simply ask for approval of a completed basket. A method that is absent from machine-readable eligibility, difficult to tokenize or ranked below a stored card may lose volume without the consumer consciously rejecting it.

The merchant economics are material. Payment methods differ in acceptance cost, authorization rate, fraud exposure, settlement timing, chargeback process and return behavior. If an agent selects solely for conversion, it may route toward a method that reduces contribution margin. If it selects solely for merchant cost, it may create customer friction or suppress an appropriate financing option.

Financing raises a sharper conflict. Installments may help a shopper manage cash flow, but an agent should not infer that credit is desirable merely because it expands purchasing capacity. Eligibility, total repayment, due dates and consequences of missed payments remain part of an informed choice. A model-generated explanation cannot replace clear disclosures and affirmative authorization.

Accountability also becomes harder. When a human clicks a payment button, the interface records an obvious action. When an agent filters, ranks and presents a method, responsibility is spread across the merchant, agent, orchestration platform and payment provider. Without a decision record, a dispute can become a reconstruction exercise.

What operators should do now

Write a payment-selection policy before launching agent checkout. Specify which factors the agent may consider, which preferences are stored, when fresh consent is required and whether merchant economics influence the ranking. Make conflicts explicit. The cheapest method for the merchant may not be the cheapest or most suitable method for the customer.

Publish complete, machine-readable eligibility. Agents need current information on geography, currency, amount limits, product restrictions, settlement, refunds and recurring-payment support. Stale rules produce failed orders and support contacts. Accurate line-item and customer data matter as well; Klarna’s Stripe guidance links complete transaction data to better authorization and fewer false declines.

Log the decision path. Retain the methods considered, the policy version, material reasons for ranking, the disclosure shown and the customer’s authorization. Do not store sensitive credentials in that record. The objective is a durable explanation of the choice, not a duplicate payment vault.

Test agent selection against a visible-choice control. Compare completion, approval, contribution margin, refunds, disputes, complaints and repeat purchase. A higher first-order conversion rate is not enough if customers later feel financing was chosen for them or if expensive routing erodes the order’s economics.

Give the customer an easy override. A well-designed agent can reduce effort without erasing agency. The best default is one that can be explained, changed and revoked.

The decision

The decisive payment battle in AI commerce may occur before checkout is rendered. Operators that treat method selection as governed product logic, with explicit consent and measurable tradeoffs, can gain convenience without letting technical availability dictate a customer’s financial choice.