Roughly 45 million Americans have no credit score at all. Another 28 million have files so thin that traditional scoring models can't generate a reliable number. That's 73 million adults, nearly a third of the credit-eligible population, locked out of mainstream lending by a system designed in the 1980s. The cost of this exclusion isn't abstract; it means higher-interest payday loans, denied apartment applications, and a persistent wealth gap that compounds across generations.
The FICO score, first introduced in 1989, was built for a different credit economy. It evaluates five factors: payment history, amounts owed, length of credit history, new credit inquiries, and credit mix. All of these are derived from bureau tradeline data, meaning you need to already have credit to get credit. The model is backward-looking by design; it tells lenders how someone managed debt in the past, not whether they can afford a new obligation today. For the millions of recent immigrants, young adults, and cash-preferring consumers who don't carry revolving debt, the FICO model has nothing to say.
I don't think FICO is broken, exactly. For the population it was designed to score, it works reasonably well. But the assumption that bureau data is the only data worth scoring against has aged poorly.
What Alternative Data Actually Means
The term "alternative data" gets thrown around loosely, so it's worth being precise. In the context of credit underwriting, it refers to any signal not found in a traditional credit bureau file. The categories that matter most today are:
Bank transaction data. Checking and savings account history, including income deposits, recurring expenses, overdraft frequency, and average daily balances. This is arguably the single most predictive alternative data source because it reflects real-time financial behavior rather than a lagging summary of credit usage.
Cash flow analysis. A step beyond raw transaction data; this involves modeling income stability, expense volatility, and discretionary spending capacity. A borrower who earns $4,200 per month with steady direct deposits and $800 in discretionary surplus looks very different from someone with the same income arriving in irregular freelance payments.
Rent and utility payments. On-time rent payments are among the strongest predictors of mortgage performance, yet they've historically been invisible to credit bureaus. Programs like Fannie Mae's positive rent payment reporting and Experian Boost have started to chip away at this gap, but ML-based underwriters were doing it years earlier.
Employment and income verification. Real-time payroll connections through providers like Argyle or Pinwheel, replacing the old model of submitting pay stubs that may be weeks old by the time a loan closes.
Behavioral and device signals. This is where it gets controversial. Some lenders look at how applicants interact with a loan application: time spent on each page, device type, typing patterns. The predictive value here is real but ethically fraught, and I expect regulators to draw harder lines around this category in the next two years.
The Companies Building This
Upstart is the largest pure-play AI lending platform, publicly traded since late 2020. Their models ingest over 1,600 variables per application, including education, employment history, cost of living by geography, and hundreds of behavioral features in addition to traditional credit data. As of their most recent earnings, they've facilitated over $40 billion in loans across roughly 100 bank and credit union partners. Their approval rates run about 27% higher than traditional models at the same loss rate, according to their own published data. The asterisk here is that Upstart took heavy losses when rates spiked in 2022-2023, which tested whether their models were truly cycle-resilient or had been trained predominantly in a benign credit environment. They've since tightened their credit box and rebuilt. Whether the models hold through a full recession remains an open question.
Zest AI has carved out a different niche. Rather than originating loans directly, they sell model-building software to credit unions and community banks. Their core pitch is explainability: every credit decision their models produce comes with reason codes that satisfy regulatory requirements under the Equal Credit Opportunity Act. For a 500-branch credit union that wants to modernize underwriting without rebuilding its compliance infrastructure, this is a practical offering. They claim their models reduce charge-offs by 20-30% while increasing approval rates, though independent verification of those numbers is limited.
Ocrolus sits upstream of the decision itself. They specialize in document-level data extraction, parsing bank statements, pay stubs, tax returns, and mortgage documents using a combination of OCR and ML classification. Lenders who want to underwrite based on bank statements need to first turn those PDFs into structured data, and doing that accurately at scale is a harder engineering problem than it sounds. Ocrolus processes millions of documents per month and feeds into the underwriting workflows of companies like Cross River Bank, Brex, and several large mortgage originators.
Petal targets the thin-file consumer directly. They issue Visa credit cards underwritten primarily on cash flow data rather than credit scores. Applicants connect a bank account; Petal's models analyze income, spending patterns, and savings behavior to determine creditworthiness. It's a clean test case for whether cash-flow underwriting works at the consumer level, and their loss rates have reportedly tracked in line with traditional card issuers despite serving a population that would largely be declined by conventional scoring.
Heron Finance focuses on the small business lending side, using open banking data to assess SME creditworthiness in markets where bureau data on businesses is sparse or unreliable. Their approach relies heavily on categorizing and analyzing business bank transactions to model revenue stability, supplier payment patterns, and cash reserves.
The Regulatory Overlay
None of this happens in a vacuum. The CFPB has been paying close attention to algorithmic lending since at least 2022, and the regulatory framework creates real constraints on what alternative data models can do.
Under ECOA and Regulation B, lenders must provide specific reasons when they deny credit or offer adverse terms. A traditional scorecard can point to "insufficient credit history" or "high utilization ratio." An ML model that weights 1,600 features needs to produce similarly specific, actionable explanations. This is a genuine technical challenge; post-hoc explainability methods like SHAP values get you partway there, but translating a model's internal feature importance into consumer-facing adverse action notices that actually make sense is still more art than science.
Disparate impact testing adds another layer. Even if a model doesn't use race, gender, or other protected characteristics as inputs, it can still produce discriminatory outcomes if its features correlate with protected classes. Zip code is the classic example, but subtler proxies exist in transaction data, device data, and educational history. The better AI lending companies run extensive disparate impact analyses and can demonstrate that their models reduce demographic disparities compared to FICO-only underwriting. But "better than FICO" is a low bar when FICO itself has well-documented scoring gaps across racial lines.
I think the regulatory posture will tighten further. The CFPB's proposed rulemaking on automated valuation models and their broader interest in algorithmic accountability signal a direction of travel. Companies building alternative credit models today need to treat explainability and fairness testing not as compliance checkboxes but as core product requirements.
Where This Is Going
The trend line is clear; the question is speed of adoption.
Real-time income verification is becoming table stakes. The days of uploading a PDF pay stub are numbered; lenders will connect directly to payroll systems and bank accounts and verify income in seconds rather than days. This collapses the underwriting timeline and opens the door to true instant lending at the point of sale.
Embedded lending is the distribution model that makes AI underwriting economically interesting at scale. When a SaaS platform, e-commerce checkout, or payroll provider can offer credit decisions in milliseconds using an API call to an AI underwriter, the total addressable market for alternative scoring expands dramatically. Shopify Capital, Square Loans, and Amazon Lending are early versions of this model, and all of them rely on proprietary transaction data rather than bureau scores.
The incumbents are moving faster than I expected. JPMorgan Chase, Wells Fargo, and Capital One have all made serious investments in ML-based underwriting over the past three years. When I first started covering this space, the conventional wisdom was that large banks would resist alternative data because of regulatory conservatism. That hasn't held. What's actually happening is that the big banks are adopting these tools internally while the fintechs compete on the edges of the credit box that traditional lenders still won't touch.
The question isn't whether ML-based credit scoring replaces FICO. FICO itself is adapting; FICO 10T incorporates trended data, and the company has partnered with alternative data providers. The real question is how quickly the lending industry shifts from "credit score plus manual underwriting" to "real-time, multi-signal, continuously updated risk assessment." Based on what I'm seeing in the market today, that shift is about 60% complete for personal loans, maybe 30% for mortgages, and still early innings for small business lending. The next three years will close most of those gaps.