Why AI Agents Could Finally Reinvent the Credit Card

Summary

• The credit card remains the best payment UI, but agents may reopen the interface. In the closing exchange, Alex Rampell argues that agents are smarter than “rewritable, chipped plastic,” so negotiations could eventually make agentic commerce and payments possible. Max Levchin is skeptical of agentic shopping but bullish on agentic payments. Rampell’s caveat is that people may still want to choose themselves, as with his bike-parts example.
• Payments is the world’s largest market, yet its most profitable opportunities are small-dollar niches. Rampell’s examples contrast a potentially enormous but difficult $40 trillion wire transfer with everyday payments where convenience dominates. His abandoned PayMeSooner idea exposed the B2B gap: GE can pay in 90 days, while a small merchant may factor the receivable at 15%, even though the borrowing is effectively against GE’s credit. They decided it was not a big business, while noting that accounts-payable and accounts-receivable financing can work.
• Visa and Mastercard’s 2.5-second transaction window is a roughly 60-year-old fossil. Apple Pay and Google Pay use secure elements to do work before the networks process the card, but Levchin’s surprise is that the networks never introduced a newer standard—such as allowing 15 seconds for additional innovation or asking issuers to bid for better credit quality.
• Levchin’s crypto verdict is earned, not reflexive. Before PayPal, he was sent away from a cryptography conference for presenting a non-anonymous digital-payments idea, and he attended DigiCash’s bankruptcy ceremony at Stanford. He admired Bitcoin’s Byzantine Generals solution but never believed it would work as a payment method; he says it has succeeded as a currency, asset, and store of value. Stablecoins have clear uses, but coffee remains the practical test: small payments are ruled by UI, while huge transfers justify optimizing safety, speed, and cost.
• Affirm’s origin fused the “pajama problem” with 1800s general-store underwriting. Recognition-based credit—such as knowing a customer through social signals—could substitute for a wallet. Levchin wanted to build a strong credit score and let others lend; Rampell focused more on completing purchases from the couch. Levchin built an all-night PHP 1-800-Flowers demo using Facebook Connect, and Jim McKelvey responded positively. Product-market fit came through Beautylish, where installments lifted conversion 30%, revealing that the product solved a budget problem and could serve as a sales tool.
• The mattress-in-a-box wave created room for genuine 0% loans, while Levchin attacked deferred-interest cards. An HBR article around the time of Casper said people replace mattresses every 7 years; several companies emerged, with compressed memory-foam mattresses offering high margins and MDR flexibility. Affirm also tried for-profit education, where MDRs could reach 50%, but exited after about half a year because customers often refused to pay for worthless education. Levchin’s 0% has no asterisk: no late fees, no deferred interest, and no retroactive interest.
• Affirm’s underappreciated assets are negative CAC, the customer relationship, and long-term underwriting. Merchants pay Affirm to acquire customers, unlike TrialPay, where Rampell merely connected merchants and users. Affirm has transacted with more than 50 million people in America and operates in four countries, while shifting from fulfilling demand to helping merchants generate it. Some Affirm products run as long as 3½ years, versus roughly 6 weeks for BNPL, requiring machine-learning underwriting rather than a FICO or Facebook shortcut.

Deep dive

1. Tap-to-pay happened by accident; the 2.5-second rule never changed

• Rampell’s explanation for the change in contactless behavior is that consumer habits are extremely difficult to change, until merchants are forced to change their terminals. Magstripes were easy to copy, so merchants had to adopt cards with safer chips and new machines. Those machines also supported contactless payments, though nobody initially tapped. COVID forced merchants to change independently and made tapping common.
• A point he makes about the acronym: the E in EMV is for Europay, not Eurocard.
• Levchin’s technical layer is that the networks impose a hard 2.5-second limit on the transaction among the issuing bank, merchant, and acquiring bank. Offline, that leaves little time for intelligent antifraud work; online, a merchant can run checks before submitting the card.
• Apple and Google Pay use secure elements in their chips, allowing them to establish knowledge of the card and perform tasks before Visa or Mastercard even speaks to it. Levchin’s surprise is that the 2.5-second constraint did not need to remain: the networks could ask issuers to bid for better credit quality or allow companies 15 seconds for other innovations. Instead, rules from roughly 60 years ago remain almost unchanged.

2. The world’s biggest market, and its smallest profitable niches

• Rampell’s paradox is that payments is the world’s largest market, yet almost everything is small. A $1 trillion wire transfer would not be easy or obviously profitable: he imagines America announcing $40 trillion in debt, or Elon Musk—or Rihanna—being worth $40 trillion and imposing a $40 trillion tax. The largest transfers justify spending more time finding the safest, fastest, cheapest method, while low-dollar payments are dominated by convenience.
• Their abandoned B2B idea began with Rampell’s April 2011 email asking Levchin about Bill Me Later. Rampell had bought the PayMeSooner domain. The logic was that GE could pay a small merchant in 90 days, forcing the merchant to finance payroll or factor the receivable at roughly 15%, while GE could issue bonds at SOFR plus 10 basis points. The merchant was effectively borrowing against GE’s credit, yet factoring was treated as selling future money, so usury laws did not apply.
• They concluded that PayMeSooner was not a big business, although Rampell notes that there are good businesses in both accounts-payable and accounts-receivable financing.

3. The ideas that never happened—and why crypto isn’t payments

• Biometric payment never became the movie-like replacement for cards. Levchin recalls a MasterCard and gas-station network device that let drivers wave a wand to pay; he thought it would replace credit cards, but it did not. The lesson is that payments require critical mass: “It’s okay” is not enough. Everyone must use the widget or network, or it disappears.
• Amazon’s palm payment at Whole Foods appealed to Levchin, though the group joked that it might not be faster. The imagined interaction was: “Tell me my future and give me these grapes and fruits,” followed by identifying the palm.
• Levchin’s crypto scar tissue is personal. He presented a digital-payments idea at a cryptography conference and was humiliatingly sent away because it was not anonymous. He also attended DigiCash’s bankruptcy ceremony in Stanford’s courtyard, where cypherpunks lamented that digital payments had not yet arrived. David Chaum’s blind-signature idea was brilliant, he says, but the market did not receive it; PayPal entered after many enthusiasts had already left.
• He admired Bitcoin’s solution to the Byzantine Generals Problem and said the mathematics and cryptography were known but the method surprised him and everyone else. Still, he never believed Bitcoin would change payments, and still does not regard it as a payment method. He says it has proven successful as a currency, asset, and store of value, while stablecoins have clear uses.
• His test is the practical coffee purchase. For a huge transfer, people will spend time optimizing safety, speed, and cost. For coffee, a long wallet password sends the buyer back to cash or a debit or credit card. As payment size falls, user-interface quality and comfort matter more.

4. Affirm’s origin: the pajama problem meets the 1800s general store

• The founders met at the Allen & Company conference in March 2009. Levchin’s wife was due March 16 and their son was born March 28, so he initially resisted attending; Rampell told him that if labor began, they would put him on a plane home. Rampell was running TrialPay, which exchanged actions such as signing up for GEICO for virtual goods such as FarmVille coins. Levchin had sold Slide to Google and had become dissatisfied with his post-PayPal work.
• Levchin’s wife reminded him that the PayPal antifraud years had been exhausting but also made him happy, and urged him to give payments another chance.
• The underwriting insight was to recreate the 1800s general store, where a shopkeeper recognized customers and let them pay later. Rampell connects this to Israeli grocery tabs and Japanese stores that recorded purchases against a business card. Modern social signals could provide recognition: 500 Facebook friends and thousands of photos might indicate that someone is a low credit risk. By contrast, a customer with an 800 FICO score receives credit offers in advance, while someone who goes to Google and says, “I need credit,” might be judged a bad risk.
• Levchin originally wanted to build “a great credit score” and let someone else handle lending. He considered buying data and asked Mark Zuckerberg whether Facebook would provide its data. Rampell’s motivation was more transactional: a television ad could create demand while he was in bed, but he did not want to get up to find a card. That was the “pajama problem.”

5. A PHP demo, Jim McKelvey, and forty years in the desert

• At the 2012 Allen & Company conference, they arranged breakfast with Jim McKelvey after seeing his name on a list. Levchin stayed up all night and built a PHP clone of the 1-800-Flowers site, with checkout through Facebook Connect and social signals that could help establish whether the buyer was a real person. He joked that AI could build it in five seconds today.
• McKelvey said the idea was good. People associated with the service already called to ask for flowers for their wives and then said they did not have a credit card. The founders offered to take responsibility if the customer did not pay. Amit Shah later took on that responsibility and became an early, enthusiastic supporter.
• Levchin improvised a pricing sheet in Word or Excel with a 7% merchant discount rate. The charge was an MDR—giving the merchant cash before the consumer paid—not an actual consumer APR. Finance hire Rob Fife looked at it and said, “Oh, free flowers.”
• The first merchants did not immediately adopt it. McKelvey objected that the charge was worse than his card transactions, and other merchants stalled. Levchin describes the startup experience as traveling “40 years in the desert.”
• Their friend Nils Johnson’s company Beautylish, which sold cosmetics and beauty products online, finally implemented the product. By then, the name was Affirm, after an earlier Expedite name. Offering customers three installments or payment 30 days later lifted conversion 30% immediately. The founders realized this was not just a payment problem; it was a budget problem. They repositioned it as a sales campaign, and direct-to-consumer merchants accepted high MDRs because installments increased sales.
• Tracy from Tradesy sent dashboard screenshots saying the “Affirm effect” had increased sales 35%. Mattress companies soon followed.

6. Mattress margins funded true 0%—and the war on deferred interest

• An HBR article about Casper around that time said people replace mattresses every 7 years. The idea was not yet established, but four or five companies emerged around it. Mattress makers clustered in Utah, compressed memory foam into boxes, and operated with high margins, creating flexibility around MDR.
• That flexibility let Affirm offer a genuinely 0% loan. A $1,200 mattress is difficult to buy all at once, but dividing it into $30 payments can materially increase conversion.
• The highest MDR charges, Levchin says, come from some for-profit education organizations, where charges can reach 50%. He cites University of Phoenix as an example and tentatively identifies it with Apollo. Affirm tried education and coding courses but exited after roughly half a year because customers often refused to pay after deciding that the education or degree was worthless.
• Levchin objects to store cards advertising “0% APR” with an asterisk. The terms can require at least one penny of principal to be paid on time; if the customer is even one day late, interest is calculated from the beginning. A $1,000 purchase can become $3,000 after two years. Affirm’s zero has no asterisk: it does not charge late fees or defer interest, and it does not use those terms to squeeze customers.

7. What people miss about Affirm: negative CAC and a duration moat

• Levchin says advertising and payments are getting closer. Rampell connects this to TrialPay and to PayPal Shops, a failed project that he tried on Back Market and that estimated what a buyer might purchase later. Affirm has shifted from merely satisfying existing demand to helping merchants create demand.
• Rampell’s VC joke is that 90% of the consumer companies he sees make him want to buy Google or Facebook stock, because those companies supply the customers. Affirm is unusual because it has negative customer-acquisition cost: merchants pay it to acquire customers.
• Unlike TrialPay, which connected Zynga or Netflix to users without owning the relationship, merchants want Affirm to own the customer relationship and handle payment communications, including notices when a borrower is late. That is especially valuable for products with 12-month, 39-month, or even 3½-year loans.
• Affirm has transacted with more than 50 million people in America and is available in four countries. Its longer-term products also create a difficult underwriting problem. Levchin contrasts roughly 3½-year products with BNPL’s approximately six-week duration; he does not present 3½ years as the average BNPL loan. FICO and Facebook shortcuts are insufficient, so advanced machine learning is required to control defaults and delinquencies. Managing those long-term products and owning the customer relationship supports future services.

8. Why there’s no new PayPal mafia—and the agentic-commerce close

• Levchin’s “old answer” comes from Jimmy Soni’s The Founders, which he praises for its extensive interviews. At PayPal, a common answer to “What will you do after PayPal?” was “start my own company.” The company also strategically attracted founders, helping produce YouTube, Yelp, Founders Fund, and LinkedIn.
• His “new answer” is that the group learned one another’s character under pressure. They worked in sweaty rooms, argued around whiteboards, and saw each other tired and frustrated. Levchin recalls seeing Elon Musk sweaty, tired, and disgusted in the company kitchen, and Peter Thiel calling while raising a fund to wonder whether they had money. Seeing that extraordinary people were also ordinary and human became part of the inspiration to attempt large ideas.
• Levchin is not optimistic about agentic shopping but is very optimistic about agentic payments. Rampell argues that the mistake is treating the robot as the thing to buy: people may enjoy choosing, as he does when comparing parts for his two bikes. The credit card remains the best user interface created so far, but agents are smarter than rewritable pieces of chipped plastic, so negotiations could eventually make agentic commerce possible. Today, saying “I want it” is the moment payment begins.
• When Erik asks about purchases requiring research, Rampell says AI is another tool to consult, like a friend. Once the buyer knows the particular SKU, comparing 19 options across sellers becomes a task for someone who values time over money. He cites CamelCamelCamel as his favorite example.