AI Agents Could Reopen the Credit Card Interface

Max Levchin calls the card-payment interface “the singular best user interface ever created.” The claim is not that cards are technically sophisticated. It is that, at the point when a person has decided to buy, a card is fast, familiar, and requires almost no further thought.

That durability has constrained payments innovation for decades. Levchin points to a hard operational limit in the Visa and Mastercard system: roughly two and a half seconds for the merchant, acquirer, network, and issuing bank to complete an interaction. Miss that window and the transaction may be retried or canceled. In e-commerce, merchants can run fraud checks and other processes before submitting a card to the network. At a physical point of sale, once the card is presented, there is little time to do anything more elaborate.

Apple Pay and Google Pay made an important exception. By using secure enclaves in phones, Levchin says, they can establish and process information before the transaction reaches Visa or Mastercard. Yet he is struck less by what those services changed than by what the networks have not changed. A longer interaction could, in principle, let issuers bid for a transaction with better credit terms or make room for other forms of innovation. The foundational rules of the “Dee Hock era,” he says, remain largely intact.

Levchin thinks agents could finally reopen that interface. “The best user interface ever created is the credit card,” he says, “This may actually be finally up for renegotiation because agents are in fact smarter than pieces of plastic, and even pieces of plastic with rewritable chips.”

The stakes are large even where the individual payment is not. Levchin describes payments as the world’s largest market and says that seemingly narrow corners of it routinely become $100 billion opportunities. A chart shown in the source puts global payments revenue at $2.4 trillion in 2023, after growing 7% annually from 2018 to 2023, and forecasts $3.1 trillion by 2028.

But scale does not translate simply into pricing power. ? alex-rampell argues that the revenue-rich portions of payments tend to be low-ticket, high-frequency transactions. A hypothetical $40 trillion wire transfer represents immense volume, but no processor will capture 2% of it. The “rake” shrinks as transaction size rises.

Quick-service restaurants illustrate the inverse dynamic. Rampell says Starbucks created its stored-value payment system partly to avoid paying card-network fees repeatedly: load $50 once, and the card fees attach to that initial funding transaction rather than every subsequent coffee. For payment businesses, the aggregate opportunity often sits in small, routine purchases where convenience dominates.

That is also why Levchin treats buying coffee as a practical test for any new payment method. On a very large transfer, a buyer will weigh security, speed, cost, and counterparties. At a bagel shop, a person whose crypto-wallet passphrase is cumbersome will reach for cash, debit, or credit. A payment system that cannot win at that point of least patience has not displaced the card in the way that matters.

For all the attention paid to new payment mechanisms, consumer behavior rarely moves because a product is only somewhat more elegant. ? alex-rampell says he was surprised by the penetration of Apple Pay and Google Pay precisely because changing payment habits is normally difficult.

Their adoption, in his account, depended on several conditions arriving together. The move from magnetic-stripe cards to EMV chips shifted fraud liability toward merchants that had not upgraded their terminals. Retailers therefore had to replace their equipment. Those new terminals included contactless capability, even though few customers initially used it. Mobile phones then became ubiquitous, and Covid further changed behavior around physical checkout. The combination of a required infrastructure transition, pervasive telephony, and a changed consumer environment made tap-to-pay commonplace.

Max Levchin contrasts that transition with payment ideas that made intuitive sense but failed to cross a threshold. Before PayPal, he recalls, Mastercard and a gas-station network offered a device that customers could wave near a pump to authorize payment. The proposition was reasonable: leave the card in a pocket, use a small device attached to car keys, and fuel the car. But it was only a little better than pulling out the card already in a person’s pocket.

The lesson, in Levchin’s view, is that payments have a threshold effect. A new device or network either reaches the point where everyone needs to have it, or it disappears. There are few middling outcomes. Biometrics have met similar resistance. Levchin says the industry is “forever in search” of a compelling biometric payment method, but people still fundamentally pay with cards or phones. Amazon’s palm-scanning system at Whole Foods, which Rampell says was discontinued, was enjoyable but not meaningfully faster. Levchin’s verdict: it was “not even faster, it’s just fun.”

Crypto, for Levchin, belongs in a related but distinct category. His work before PayPal involved making cryptographic primitives work on very low-power chips. The resulting technology could encrypt and decrypt small amounts of data quickly, which led naturally toward payments. But he arrived at digital payments after the failure of DigiCash, David Chaum’s earlier attempt built around cryptographic anonymity.

Levchin recalls presenting PayPal’s approach at a cryptography conference and being booed because it was neither as secure nor anonymous as the audience wanted. PayPal’s essential insight, he says, was that people did not need anonymity to buy coffee or make online purchases. They needed to pay.

He found the original Bitcoin paper technically clever as a solution to the Byzantine generals problem, but did not expect Bitcoin to become a currency or everyday payment method. He remains unconvinced that it has become one in that practical sense, while acknowledging its success as a commodity, currency, and store of value. Stablecoins have produced clearer applications, he says, but none has displaced the canonical test: a person choosing the system to buy coffee because it is the best way to do so.

The early concept behind Affirm was not simply installment financing. It began with what ? alex-rampell calls the “pajama problem”: a person wants to buy something on a phone, but their card or wallet is elsewhere in the house. The practical question was how to complete that transaction without making the buyer get up.

Rampell’s work at TrialPay shaped the initial answer. TrialPay arranged alternative ways to pay for digital goods: a customer might receive virtual currency by signing up for an offer from an insurer, streaming service, or credit-card provider. The broader thesis was that advertising and payments could converge.

In conversations with Max Levchin around 2011 and 2012, the pair imagined an updated general-store relationship. A merchant who knew a customer could let that person take goods and settle later; online, the merchant saw only a cookie and an IP address. They wondered whether identity signals could recreate that trust. A shopper with hundreds of Facebook friends, years of uploaded photos, and signs of being a real account might be safer than an anonymous browser.

Rampell is explicit that this was wrong as an underwriting model. Still, it exposed a useful distinction. A person actively searching for credit may be a poor risk; a person trying to complete an ordinary purchase without a wallet at hand is presenting a different problem.

Levchin was initially more interested in building a better credit score than in building a lending company. After PayPal and a period running Slide, the social-media company Google acquired, he had been reluctant to return to financial services. His wife urged him to notice that he had been happiest during PayPal’s anti-fraud years, despite the strain. The possibility of using machine learning and richer data to assess credit drew him back.

The founders demonstrated the identity concept to 1-800-Flowers with a cloned checkout flow that included “pay with Facebook.” It used Facebook Connect, a threshold of more than 500 friends, and Facebook account-quality signals intended to distinguish a real person from a fraudulent account. Jim McCann, whom Rampell and Levchin met for breakfast, immediately connected it to an existing practice: service members would call to send flowers without their cards, and the company would take the order and collect later. Affirm’s proposed role was to take that repayment risk off the merchant’s hands.

The demo established that merchants could understand the proposition. It did not establish a business. The company, briefly incorporated as Expedite Software, contemplated a 7% merchant discount rate. But the offer appeared only after a shopper had already selected flowers, leaving the merchant with a simple objection: the customer was prepared to pay by card, and the new option was more expensive than card acceptance. In Rampell’s recollection, one observer looked at the setup and saw “free flowers”—a product whose economics failed when consumers did not repay.

The decisive discovery was that financing could be valuable not as a substitute at checkout, but as part of the purchase proposition from the beginning.

The 1-800-Flowers experience left Affirm, in Max Levchin’s account, in the ordinary “40 years in a desert” of startup building: enough evidence to keep working, not enough to know whether the company was headed toward anything meaningful.

At 1-800-Flowers, the financing offer came too late. The merchant saw it as cannibalizing credit-card volume, while conversion and interface quality remained unsatisfactory. Affirm was trying to persuade merchants to pay extra for a different way to process an already-intended purchase.

Beautylish, an online beauty retailer, changed the placement. It told shoppers while they were browsing shampoos and perfumes that they could pay in installments or over time. Levchin says that produced an immediate 30% increase in conversion.

The implication was much larger than the pajama problem. Financing could change what a shopper believed they could afford and therefore what they chose to buy. A consumer who came in for shoes, Levchin says, might add a bag if transparent financing made the total manageable. The merchant was not paying for an alternate rail; it was paying for incremental demand.

That made the merchant discount rate easier to justify. Small direct-to-consumer brands trying to grow their top line cared less whether the fee was 1%, 5%, or 12% if financing lifted conversion enough. Levchin remembers early merchant executives sending dashboard screenshots showing increases they attributed to Affirm, including a 35% lift at Tradesy.

Mattress brands became an especially clear fit. The direct-to-consumer mattress boom paired high-margin goods with a large upfront price and a purchase people often resist making. In Rampell’s example, a consumer lying on an uncomfortable mattress may not want to spend $1,200 at once; a recurring payment changes the decision. The product’s long replacement cycle amplified the value of conversion. If a mattress brand missed the moment when a buyer was ready, its next chance might be years away.

Not every high-fee category was attractive. ? alex-rampell notes that for-profit education providers could bear exceptionally high merchant discount rates because online-course margins were high and they expected substantial nonpayment. Affirm briefly entered the category, attracted by the prospect of subsidizing education. But Levchin says it left within roughly half a year because defaults often reflected dissatisfaction with poor-quality programs rather than ordinary credit risk. Borrowers who concluded that a degree or credential was worthless had little reason to continue paying.

The founders’ rejection of deferred-interest promotions became another defining distinction. Levchin objects to retail cards marketed as 0% APR with an asterisk: if a customer is a day late or a dollar short at the end of the promotional period, interest can accrue retroactively from the original purchase. In his formulation, someone who believed they had financed a $1,000 purchase can wake up much later owing far more.

Affirm’s “real zero,” by contrast, meant that a merchant could fund a genuine 0% loan through the merchant discount rate, with no deferred interest and no late fees. Levchin says that once a customer agrees to a three-year arrangement, “you will never get screwed, you will never be surprised to the negative.” A late payment does not rewrite the original price.

? alex-rampell argues that Affirm’s unusual strength is negative customer-acquisition cost. Most consumer companies pay heavily to acquire customers, and Rampell says the economic value of many of those businesses ultimately flows to Google or Facebook, where their demand originates. Affirm, by contrast, is paid by merchants to acquire consumers.

That arrangement works because merchants do not want to manage the financial relationship themselves. A mattress company may want the conversion lift from financing, but it does not want to send delinquency notices, manage repayment questions, or become the party telling a customer to pay for the mattress they are sleeping on. The merchant wants a third party to take on that work.

For Rampell, this differs from TrialPay’s B2B2C model. A game company might let TrialPay connect its player with an advertiser, but it did not want TrialPay to own the player relationship. With Affirm, merchants actively prefer that the lender own the financial relationship because it removes a burden from the brand while producing sales.

Max Levchin says that relationship is now central to Affirm’s strategy. He characterizes the company as having transacted with more than 50 million Americans, operating in four countries, and processing tens of billions of dollars a year. It still helps merchants satisfy existing demand—letting a consumer finance the shoes and perhaps add the bag—but it is increasingly intended to help merchants create or guarantee demand by reaching consumers who have already used the service.

That is the delayed fulfillment of the old TrialPay thesis that payments and advertising converge. Rampell saw a version of it at TrialPay; Levchin had pursued a related idea in a PayPal project intended to remarket based on what someone had just bought. The underlying behavior, they argue, became viable only at greater scale.

Longer-term loans are important to that strategy even though they are difficult. Levchin says many buy-now-pay-later products run for roughly six weeks, while Affirm will make loans extending to three and a half years. That requires genuine underwriting rather than superficial proxies such as a FICO score or social-media connections. The company must manage default and delinquency risk over a much longer period, which he says demands sophisticated machine-learning work.

The payoff is repeated contact. A 12-month loan gives the lender 12 opportunities to communicate with a customer; a three-and-a-half-year loan gives it 39. Those payment interactions can become opportunities to introduce new services. In Levchin’s framing, long-duration credit is both the most difficult part of the business and a mechanism for building a durable consumer platform.

The near-term argument over agentic commerce should not focus only on whether people will let AI choose their products, Max Levchin says. He is skeptical that consumers will broadly delegate taste-driven purchases. He enjoys comparing nearly identical bike parts and does not want a machine to remove him from that decision. The important moment may come after the decision: “that one” should be the end of buying, rather than the beginning of finding a wallet, choosing a card, or navigating a checkout path.

? alex-rampell agrees in part, but sees a more immediate use case in execution rather than product discovery. Once a customer knows the exact SKU or UPC they want, an agent could search among many sellers and buy it at the lowest cost. He compares the behavior to CamelCamelCamel, a site for tracking Amazon prices: people with more time than money already pursue those savings manually. An agent could make that work available to people who value time more highly.

The same distinction applies to payment optimization. Rampell says he has multiple cards and cannot reliably remember the best terms for each purchase: which card has the right category reward, whether the purchase falls below a particular threshold, and how the terms apply in the specific case. An agent could evaluate those choices as it compares retailers.

Levchin’s reservation is that merchant selection is not only a lowest-price calculation. A shopper is implicitly assessing whether a retailer is reliable, whether an item will arrive on time, whether it will match the listing, and whether a return will be manageable. A dated-looking site offering a cheap bike cassette may be legitimate, but the buyer is making a probabilistic judgment about delivery and quality that is difficult to formalize.

He thinks AI may already be capable of some of that reasoning. The obstacle may instead be trust: users have not yet learned to trust an agent to do as good a job as they believe they can do themselves. The adoption curve could be slow even if it eventually accelerates suddenly.

At the same time, Levchin argues that consumers are already more accustomed to delegated commerce than they acknowledge. Grocery delivery through Instacart is, in his words, “100% agentic.” A shopper asks someone else to buy milk; that person substitutes another brand when necessary; and the order arrives correctly or better than expected almost all the time. The agent is human rather than software, but the behavioral precedent is established.

The harder cases are those with subtler preferences and more costly failures: the wrong bike part, an opaque seller, a delayed delivery, or a difficult return. Working through those quirks may take longer than enthusiasts expect. But if agents earn trust after the consumer has decided what they want, they could eliminate the procedural work that has kept the card at the center of checkout.