Somewhere between the time you tapped "Place Order" and the confirmation screen appeared, a machine made a $200 bet on your character. It took about 800 milliseconds. You didn't notice... and you weren't supposed to.
In 2026, Buy Now, Pay Later (BNPL) is not just a button anymore. It's not even just a product. It's a layer of invisible financial plumbing running beneath modern commerce — a real-time probability engine deciding, before the confirmation page even finishes loading, whether you're a fraud bot, a high-risk borrower, or somebody worth extending credit to.
BNPL is taking the world of e-commerce by storm. While short term lending is hardly a novel concept, it was popularized in the digital economy by the Swedish company Klarna. The space is now dominated by many players, including PayPal, Affirm, and Afterpay (Block).
The global BNPL market hit roughly $560 billion in gross merchandise volume in 2025, growing at nearly 14% year-over-year. By 2026, transaction value is forecast to push past $565 billion. We're talking about hundreds of billions of dollars moving through systems that have to make lending decisions faster than you can blink. What does credit even mean at that speed?
FICO scores are starting to feel like fax machines
Here's the thing about a traditional credit score: it tells you what someone's debt looked like last month. Maybe two months ago. It's a rearview mirror in a car that needs to make split-second turns.
BNPL underwriters have been building something different. Using a vast amount of alternate data available at their disposal, they plug into a set of signals to answer a narrower, more immediate question: should we lend to this person, for this purchase, right now? Not "what did their debt look like six months ago" but "what does the risk look like in this exact moment?"
One piece of this is behavioral biometrics. The system watches how you interact with your device — your typing cadence, your scroll speed, whether you're copy-pasting credit card numbers or typing them from memory, and so on. These aren't creepy surveillance metrics for the sake of it. They're bot detectors. When a fraud ring is stacking loans across dozens of accounts, the humans behind those accounts don't type like real shoppers do. They move too fast, too precisely, or in patterns that scream automation. Companies like BioCatch, Sardine, and others have built entire platforms around reading these micro-behaviors to flag synthetic identities before they ever reach a checkout.
The other piece is open banking. Providers like Affirm have started using real-time bank account data — actual balance checks and cash flow patterns pulled through APIs via partners like Plaid — to underwrite loans at the point of sale. Instead of asking "what does your credit report say?" the system asks "do you have the money right now?" Affirm rolled this approach out more broadly in early 2026, pulling real-time balance and income signals from linked bank accounts into every lending decision.
The pitch from the industry is that these behavioral and cash-flow signals are significantly more predictive of repayment than traditional bureau data alone. Affirm claimed their systems to be 3-4x more accurate than traditional credit cards. That sounds plausible, but whether it's actually true at scale is a different question — one that got a very public stress test in late 2025.
How's it working out for Klarna?
While all of this fancy underwriting tech sounds great on paper, the real test is what happens when a company bets billions on it... and then has to show the receipts.
Klarna went public on the NY Stock Exchange in September 2025. They sold shares at $40 each. It was a big deal. The Swedish BNPL giant had been valued as high as $45 billion during the pandemic, cratered to under $7 billion in 2022, and was now coming back to market at around $14 billion. It was a real comeback story.
Then, two months later, they released their first earnings report as a public company.
Revenue looked great — $903 million, up 26% from the year before. But the number that spooked everyone was the money Klarna set aside to cover loans that might go bad — aka their "loss provisions" — jumped 102% in a single year. The company was now reserving nearly twice as much for potential defaults as it had the year before. The stock dropped 9% in a day, and went below their IPO price.
And then the lawyers showed up...
In December 2025, a class action lawsuit — Nayak v. Klarna — was filed on behalf of investors who bought in at the IPO. The core allegation is straightforward: you can't tell investors your risk models are best-in-class, go public on that story, and then double your rainy-day fund two months later. The argument is that Klarna knew, or should have known, that a lot of its borrowers were financially stressed people taking out interest-bearing loans for everyday purchases. We're talking fast food deliveries. Groceries. Burritos. Stuff that gets eaten before the first payment is even due.
Klarna sees it differently. Their argument is essentially an accounting timing issue. When you make a longer-term loan, the rules say you have to set aside money for potential losses right away — on day one. But the interest income from that loan comes in slowly, over months. So the books look worse early on, even if the loans are fine. It's front-loaded pessimism, not actual defaults.
But regardless of who wins the lawsuit, the episode raises a question the whole industry has to sit with: if your system is fast enough to approve a $15 Chipotle order in under a second, is it also smart enough to know that the person ordering is quietly stacking five of those loans across three different providers?
The self-driving wallet: when AI decides, not you
Now forget the button entirely. The next phase of BNPL isn't about whether you choose to split a payment. It's about whether your AI agent does. A "Buy for Me" Agent doesn't ask "Should I split this into four payments?" Instead, it evaluates your liquidity and market conditions in real time to make wiser financial decisions on your behalf within acceptable thresholds set by you.
Imagine that your grocery bill is $280. Your AI detects that the idle cash in your checking account could earn up to 4% in a short-term investment strategy. The agent splits your grocery bill into four payments not because you lack the cash but because deploying your funds elsewhere yields better incremental returns. BNPL has become a temporary liquidity bridge where people see groceries and commodities but agents see capital allocation. Traditional BNPL providers can sell liquidity-as-a-service to become the preferred liquidity API for autonomous agents.
In a world where BNPL becomes a utility layer, the underwriting war isn't about who approves more users, but who build models capable enough for autonomous wallets to plug into.
The ghost becomes the driver
For years, the ghost in the checkout was invisible underwriting. Consumers thought they were choosing how to pay. In reality, they were being scored in real time by systems optimized for yield.
Now, the ghost is evolving.
As AI agents manage end-to-end shopping experiences, underwriting becomes a competitive battleground between machines. BNPL providers will compete over integration trying to convince autonomous wallets that their models are trusted enough to plug into automatically. The decision shifts from should this person get credit, to should this agent receive liquidity access?
The ghost in the checkout is no longer haunting the system, but driving it. The institutions that end up thriving in this transition will be those whose models can withstand the scrutiny of markets, regulators and federal agencies. Somewhere between tapping “Place Order” and the confirmation screen, the ghost in the checkout isn’t just betting on you anymore — it’s preparing to manage you.