Published in 2026. coverage 2024-2026. reviewed 2026-07-10.
Introduction
AI lending is no longer a single category with a single business model. By 2026, companies marketed under the broad banner of “AI lending” span at least four structurally different approaches: direct lenders that make and hold credit decisions themselves, network lenders that route risk to outside institutional capital, document-AI vendors that feed cleaner data into someone else’s underwriting model, and legacy digital lenders that have layered Machine Learning onto decades-old balance sheets. Treating these as interchangeable obscures more than it reveals, because the incentives, regulatory exposure, and failure modes differ meaningfully across each group. This market map organizes the landscape by business model rather than marketing language, using verified, company-disclosed information as the basis for each entry.
Direct lenders: origination and risk in one place
Upstart remains the clearest example of a direct AI lender: it originates consumer loans using a model that layers alternative variables on top of traditional credit data, and it has now been tested through a full interest-rate cycle, including the sharp increases of 2022 and 2023 that pressured loan volumes across the industry. Affirm occupies a related but distinct niche as a point-of-sale installment lender, underwriting each purchase individually at checkout rather than issuing a revolving credit line, with a revenue model that does not depend on late fees the way a traditional credit card does. Figure extends the direct-lending model into home equity, using blockchain-based infrastructure to automate loan origination and securitization end to end. What unites these three is that each holds both the underwriting decision and, at least initially, the resulting credit exposure, meaning their model performance is directly visible in their own disclosed loan-loss data rather than obscured behind a partner bank’s balance sheet.
Network lenders: routing risk to outside capital
Pagaya represents a structurally different model. Rather than originating loans directly, Pagaya integrates via API with bank and fintech lending partners and uses AI models trained across its network — not any single partner’s own portfolio — to identify creditworthy applicants a partner’s own model would have declined. Critically, Pagaya then absorbs the resulting credit risk itself through capital raised from institutional investors, letting a partner lender approve more of its applicant pool without expanding its own balance-sheet exposure. This network structure means Pagaya’s commercial success depends on convincing both lending partners and institutional capital markets that its risk-routing model is sound, a different sales motion than a direct lender pitching consumers or a bank pitching depositors. Nova Credit plays an adjacent but narrower role, translating cross-border and cash-flow data into standardized, lender-ready risk analytics that partner lenders can plug into their own underwriting rather than assuming any of the credit risk itself.
Document AI: the unglamorous layer underneath every decision
A credit decision, however sophisticated the underlying model, is only as good as the data feeding it, and a large share of that data still arrives as scanned or photographed paper documents of wildly inconsistent quality. Ocrolus has built its business specifically around this problem, converting bank statements, pay stubs, and tax documents into structured, underwriting-ready data using a Human-in-the-Loop design that pairs machine-learning extraction with human reviewers who check and correct model output before it reaches a lender’s system. Ocrolus has named lenders including Brex, PayPal, Plaid, and SoFi as customers, illustrating that document AI often sits invisibly upstream of decisions attributed to other, more consumer-visible brands. The company’s 2023 launch of Instant, a fully machine-only processing tier offered alongside its human-reviewed product, shows how document-AI vendors have approached the trade-off between speed and verified accuracy: offering full Automation only once model confidence had caught up with production lending requirements, rather than defaulting to it from the start.
Digital-origination platforms for banks and mortgage lenders
Blend and nCino sell software that automates origination and decisioning workflows for banks and mortgage lenders, rather than making credit decisions themselves under their own brand. Blend’s platform, after a 2021 diversification into title insurance that it later reversed by divesting Title365 in 2026, has refocused around its Autopilot AI Agent, which reviews borrower documents and checks compliance against lender guidelines in as little as 15 to 25 seconds — a public “reset,” in the words of its own Head of Blend, after concluding the company had spread itself across too many product lines at once. nCino’s Bank Operating System, used by financial institutions across 18 countries, introduced its AI foundations relatively early: nCino IQ launched in 2018, well before Generative AI became a mainstream enterprise priority, giving the company’s later AI features such as Banking Advisor a more mature technical base than platforms that bolted on generative AI only after 2023.
Legacy digital lenders retrofitting AI onto older balance sheets
Not every AI-branded lender started as an AI-native company. Enova International, a publicly traded online lender, has run its Colossus machine-learning underwriting platform for non-prime consumer and small-business credit for years, applying AI to a lending book that predates the current generative AI boom by more than a decade. Happen Bank, the digital-first marketplace lender formerly known as LendingClub before its 2026 rebrand, similarly represents a company that built its original marketplace-lending model well before AI was a marketing category, and has since layered additional automation onto that existing infrastructure as a chartered bank rather than a pure marketplace. These companies are useful counterexamples to the narrative that AI lending is inherently a startup phenomenon: some of the more battle-tested AI underwriting models in the market are running inside companies whose founding predates the current AI cycle by a decade or more.
Fraud and identity signals feeding into lending decisions
AI lending decisions increasingly depend on fraud and identity infrastructure that sits adjacent to, but distinct from, the underwriting model itself. Sardine combines fraud prevention, compliance, and credit underwriting signals into a single agentic risk platform for financial institutions, reflecting a broader pattern of vendors bundling categories that used to be procured separately. Socure and Trulioo both focus on identity verification at account opening, catching synthetic identities before an account is ever funded — a problem distinct from, but directly relevant to, whether the credit decision that follows is even evaluating a real person. Lenders that treat identity verification, fraud scoring, and credit underwriting as three separate procurement decisions, rather than an integrated pipeline, are increasingly the exception rather than the rule by 2026.
The regulatory backdrop shaping model design
U.S. banking regulators and the Consumer Financial Protection Bureau have continued to emphasize that existing fair-lending and adverse-action notice requirements apply regardless of whether a credit decision is made by a traditional scorecard or a machine-learning model, meaning every direct AI lender discussed here must be able to explain declines in terms regulators and consumers can understand. In Europe, the EU AI Act’s phased implementation has begun to touch creditworthiness assessment as a higher-risk use case, pushing vendors serving European banks toward heavier investment in model documentation and explainability tooling. This regulatory pressure is a meaningful design constraint across every category in this market map, not just direct lenders: document-AI vendors like Ocrolus and network lenders like Pagaya both must be able to demonstrate, to their bank and lender partners, that the data and risk-routing decisions feeding into a final credit outcome would themselves survive supervisory scrutiny.
Market shifts worth tracking
Several shifts are worth watching through the remainder of 2026. First, whether more direct lenders move toward hybrid structures resembling Pagaya’s network model, routing a portion of marginal-approval risk to outside capital rather than holding it entirely on their own balance sheet. Second, whether document-AI vendors like Ocrolus continue expanding fully automated tiers as model confidence grows, gradually reducing the share of human-reviewed volume. Third, whether legacy digital lenders such as Enova and Happen Bank narrow the perceived gap between “AI-native” and “AI-retrofitted” lenders as their own machine-learning platforms mature further. Finally, continued consolidation between fraud, identity, and credit-risk categories — visible already in Sardine’s bundled positioning — suggests that lenders are increasingly buying integrated risk platforms rather than assembling point solutions themselves, a trend that is likely to continue reshaping vendor selection criteria for both banks and fintechs.
How buyers should diligence an AI lending vendor
For a bank or credit union evaluating any AI lending vendor, the most useful diligence question is rarely “how accurate is your model” in the abstract, since accuracy claims are difficult to compare across vendors using different data and evaluation methodologies. A more productive line of diligence asks which of the business models described above a vendor actually is, what happens to a declined applicant’s data afterward, and whether the vendor can produce adverse-action explanations that would satisfy a bank examiner rather than only a data scientist. Direct lenders such as Upstart and Affirm have had to build this explainability muscle earliest, since they face direct regulatory scrutiny of their own credit decisions; document-AI and origination-software vendors such as Ocrolus, Blend, and nCino face a related but less direct obligation, since their bank customers remain the party ultimately accountable for the credit decision even though the vendor’s technology shapes the data feeding it.
The interest-rate cycle as a real-world stress test
One of the more underappreciated developments of the 2024–2026 period is that several AI lending models have now been tested through a genuine, multi-year interest-rate tightening cycle rather than only through the benign, historically low-rate environment many of these models were originally trained and marketed in. Upstart’s own experience through the 2022–2023 rate increases, and its subsequent public commentary on loan volumes and performance, gives outside observers one of the only continuously disclosed, company-specific case studies of how an AI underwriting model behaves outside of easy credit conditions. This matters more than it might initially appear, because a model’s statistical performance during a benign rate environment says relatively little about how it will behave once borrower stress increases broadly across the market, and lenders without a comparable multi-year track record are, by definition, less battle-tested than those that have already lived through one full cycle.
Conclusion
The AI lending category is best understood as several distinct business models operating under one popular label. Direct lenders such as Upstart, Affirm, and Figure hold both the decision and the risk; network lenders such as Pagaya route risk to institutional capital while retaining the decisioning intelligence; document-AI vendors such as Ocrolus supply the data quality upstream of someone else’s decision; origination-software vendors such as Blend and nCino power decisions made under a bank’s own brand; and legacy digital lenders such as Enova International and Happen Bank demonstrate that AI underwriting is not exclusively a startup phenomenon. Readers evaluating any company in this space should start by asking which of these models it actually is, since that answer determines far more about its incentives and regulatory exposure than any AI-specific marketing claim.