When a consumer taps "Pay in 4" at an online checkout, the lending decision behind that button happens in under 200 milliseconds. That's less time than it takes the browser to render the next page. In that window, the system pulls bank transaction data, computes dozens of features about the borrower's financial behavior, runs a fraud screen, scores creditworthiness, and returns an approve/decline decision with a specific credit limit. The consumer sees either a payment plan or a polite rejection. They don't see the infrastructure that made the decision, and until recently, most of the industry didn't either.
Embedded lending has become one of the fastest-growing segments in consumer and small business finance, and the AI systems running underneath it are more sophisticated than the simple "buy now, pay later" branding suggests. I've spent the past year tracking how the underwriting stack has evolved since the BNPL correction of 2022-2023, and what's emerged is a credit infrastructure that looks very different from traditional installment lending.
The Decision Stack in 200 Milliseconds
The AI behind an embedded lending decision isn't a single model. It's a pipeline of models running in sequence and in parallel, each handling a different part of the risk assessment.
The first stage is data ingestion and feature computation. When a borrower initiates a lending request, the system pulls real-time signals from multiple sources: bank account data (via open banking APIs or account linking), device fingerprints, behavioral signals from the browsing session, and any prior transaction history with the lender. This data has to be assembled and transformed into model-ready features in 50 to 100 milliseconds. The features themselves are things like average daily balance over the past 90 days, income regularity, recurring expense patterns, NSF (non-sufficient funds) frequency, and velocity of recent credit applications.
Running in parallel with the credit scoring is a fraud detection layer. This model evaluates whether the person making the request is who they claim to be and whether the transaction shows signs of fraud. Device intelligence, IP geolocation, behavioral biometrics, and velocity checks all feed into this score. If the fraud score exceeds a threshold, the transaction is declined regardless of creditworthiness.
The credit model itself is typically an ensemble; I've seen implementations using gradient-boosted decision trees trained on the lender's proprietary repayment data. The model outputs a probability of default, which maps to an approval decision and a credit limit. Some lenders run multiple model variants for different product types (pay-in-4 vs. longer-term installment loans) and select the appropriate offer based on the borrower's profile and the merchant's product configuration.
The entire pipeline returns a decision to the checkout interface, which presents the offer or the decline. From the borrower's perspective, it's instant. From an engineering perspective, it's a carefully orchestrated sequence of API calls, model inferences, and business logic running under hard latency constraints.
Consumer BNPL After the Correction
The buy now, pay later sector went through a painful recalibration in 2022 and 2023. After years of aggressive growth, several major BNPL providers reported rising charge-off rates as consumer spending shifted and interest rates climbed. Klarna, the largest pure-play BNPL company, saw its valuation drop from $45.6 billion in mid-2021 to $6.7 billion in a 2022 funding round before recovering to roughly $14.6 billion by mid-2024. The company processes $45.6 billion in gross merchandise value annually and serves over 150 million users globally, but the path to its 2024 IPO filing was shaped by the credit losses it absorbed during the correction.
The correction forced a shift in how BNPL lenders think about underwriting. Early BNPL models were relatively simple; some providers approved borrowers based on little more than an email address, a phone number, and a soft credit pull. The losses that followed were predictable in retrospect. What changed was the depth of data these lenders now use to make decisions. Klarna's current underwriting stack incorporates open banking data, merchant-level transaction history, and real-time affordability assessments that go well beyond the traditional credit bureau score.
This matters because BNPL borrowers are disproportionately thin-file consumers; people with limited or no traditional credit history. A FICO score, if one exists at all, doesn't capture their financial behavior with much fidelity. Cash-flow underwriting, which evaluates a borrower's income and spending patterns from bank transaction data, provides a richer picture. The tradeoff is computational complexity: parsing months of bank transactions, categorizing each one, and computing meaningful features is significantly harder than pulling a three-digit score from a bureau.
Upstart has been one of the most visible proponents of this approach. The company's models incorporate over 1,600 variables, including education and employment data, to predict default risk. Upstart argues that its models approve 27% more borrowers than traditional models at the same loss rate. Whether those numbers hold across economic cycles is still being tested; the company's charge-off rates rose during 2022-2023 alongside the rest of the industry, though they've since stabilized.
B2B Embedded Lending and Platform Credit
The consumer BNPL story gets most of the attention, but I think the more structurally interesting development is happening in B2B embedded lending, where platforms extend credit to their own merchants and sellers using proprietary transaction data.
Shopify Capital has disbursed over $5 billion in merchant cash advances and loans since launch, underwriting each offer based on the merchant's sales history on Shopify's platform. The data advantage is substantial: Shopify sees every transaction, every refund, every seasonal pattern in a merchant's business. A traditional lender evaluating the same merchant would need to collect tax returns, bank statements, and financial projections. Shopify already has the signal it needs.
Amazon Lending operates on the same principle, extending credit to third-party sellers on its marketplace. Amazon's underwriting relies on seller performance metrics, sales velocity, inventory data, and customer reviews. The loans are repaid automatically through deductions from the seller's Amazon payouts, which reduces collection risk and simplifies servicing.
Stripe has built what I'd describe as lending-as-a-service infrastructure. Through Stripe Capital and its broader financial services platform, Stripe enables its business customers to offer credit products to their own users. The underlying data is transaction flow through Stripe's payment processing; authorization rates, average transaction values, chargeback rates, and revenue trends. Stripe's position in the payment stack gives it a data asset that's difficult for a standalone lender to replicate.
What these platforms share is a structural information advantage. They don't need to ask borrowers to submit documentation because they already observe the borrower's economic activity in real time. The AI models they build on top of this data are trained on outcomes they can directly measure: did this merchant's sales grow after receiving capital, did repayment proceed on schedule, did the business churn from the platform? This closed feedback loop produces underwriting models that improve quickly with scale.
The Infrastructure Layer
Behind many embedded lending products sits a layer of infrastructure companies that provide the plumbing: credit decisioning APIs, bank data aggregation, identity verification, and compliance automation.
Stripe's role here is dual. It's both a platform lender (through Stripe Capital) and an infrastructure provider that other companies build on. Its Treasury and Issuing products allow platforms to embed financial services, including lending, without building the regulatory and operational infrastructure from scratch.
The data aggregation layer is equally important. Companies like Plaid and MX provide the bank account connectivity that makes cash-flow underwriting possible. Without reliable access to transaction-level bank data, the entire real-time underwriting model falls apart. The quality and coverage of these data connections directly affects the accuracy of the credit models that depend on them.
For a broader view of the companies building in this space, the lending segment of the directory tracks over 40 companies across infrastructure, data, and application layers.
Regulatory Friction Points
The CFPB issued an interpretive rule in 2024 that treats BNPL providers as credit card issuers under Regulation Z. This means BNPL lenders must provide billing statements, allow consumers to dispute charges, and issue refunds for returned products in the same way credit card companies do. The rule adds compliance cost, but it also creates a clearer regulatory framework that larger players like Klarna and Affirm are better positioned to absorb than smaller competitors.
State licensing is another friction point. In the US, lending is regulated at the state level, and embedded lenders need licenses in each state where they operate. Some platforms use bank partnerships to lend under a bank charter, which preempts state licensing requirements but introduces its own regulatory complexity. The OCC's "true lender" rule and related court challenges have created uncertainty about when a platform's bank partner is the "true lender" versus when the platform itself needs a license.
For B2B embedded lending, the regulatory picture is somewhat simpler. Merchant cash advances, which Shopify Capital primarily offers, aren't technically loans under most state laws; they're purchases of future receivables. This distinction matters because it means MCAs aren't subject to state usury caps or many consumer lending regulations. Whether this regulatory gap persists is an open question. Several states have introduced commercial lending disclosure requirements, and the trend is toward more oversight, not less.
The intersection of AI and lending regulation raises its own issues. The Equal Credit Opportunity Act (ECOA) and fair lending laws require lenders to explain adverse actions, meaning they must tell declined applicants why they were declined. When the decision comes from an ML model with hundreds of input features, generating a clear and accurate adverse action reason is a non-trivial engineering problem. Some lenders use post-hoc explanation methods like SHAP values to identify the features most influential in each individual decision; others maintain simpler, more interpretable models specifically for the adverse action notice.
Where Embedded Lending Goes From Here
I expect embedded lending to continue growing, but the growth rate will depend on credit performance through the current rate cycle more than on any technical advancement. The AI infrastructure is already capable of making sub-second credit decisions at scale. What remains uncertain is whether the models trained during a period of low interest rates and high consumer spending will perform well in a higher-rate, slower-growth environment.
The platforms with proprietary transaction data will have an advantage. Their models are trained on first-party behavioral signals that don't degrade the way third-party credit data can. But proprietary data also creates concentration risk: if a merchant's entire lending relationship depends on a single platform, the platform's underwriting decisions carry outsized influence over which businesses can access working capital.
I'm watching three specific things over the next 12 to 18 months. First, charge-off rates at major BNPL providers through the rest of 2025; if they stabilize below 2022 peaks, it validates the underwriting improvements. Second, how the CFPB's BNPL rule affects smaller providers' unit economics once full compliance is required. Third, whether any of the B2B platform lenders disclose enough performance data to independently evaluate their underwriting quality; so far, most keep that data close.
Embedded lending isn't going to replace traditional bank lending. But for the segments of the market where speed matters, where borrowers are thin-file, and where the lender already has transaction data, the AI-driven embedded model is becoming the default. The question isn't whether it works. It's whether it works well enough, for long enough, across a full credit cycle.