Beyond FICO: the alternative data powering AI-driven payday lending

The traditional payday lender of the 2000s asked for a paystub, a checking account, and a credit score, and made underwriting decisions in roughly five minutes. The modern AI-driven lender asks for read-only access to your bank account, your phone’s permissions, and sometimes your Plaid history — and makes the same decision in roughly thirty seconds, with a far more accurate prediction of whether you’ll repay. The shift from FICO to alternative data has reshaped subprime lending, and most borrowers don’t realize how much they’re handing over.

Bank transaction data does most of the work

Once a borrower connects their checking account through a service like Plaid or MX, the lender has months of categorized transaction history. The model can see paycheck timing and stability, rent or mortgage payments, recurring subscriptions, gambling activity, overdraft frequency, and balance trends. Studies from Cornerstone Advisors and the CFPB have shown that cash-flow underwriting outperforms traditional credit scores for thin-file borrowers — often the people payday lenders target. A stable income deposit pattern and consistent rent payment will get an applicant approved even with a 540 FICO, because the model has learned those signals are more predictive than the score itself.

Rent, utilities, and telecom records fill in the rest

Services like Experian Boost and dedicated rent-reporting bureaus feed lenders signals that traditional credit files miss. On-time rent and utility payments, phone bills paid without disconnection, and cable accounts in good standing all become positive inputs. For lenders, this is good news: it’s predictive data that traditional bureaus undercount. For borrowers, the trade-off is that opting in means consenting to ongoing reporting, which can also reflect missed payments going forward. The same alternative-data infrastructure that helps you get approved can hurt you later.

Smartphone metadata is the frontier and the gray zone

Some international and U.S. fintech lenders have experimented with smartphone-derived signals: contact list size and structure, app usage patterns, GPS stability suggesting consistent home and work locations, and battery and storage behavior as proxies for device age and economic situation. Academic studies from researchers including Berg, Burg, Gombović, and Puri showed digital footprint variables can equal or exceed credit-bureau scores at predicting default. Regulators in the U.S. have looked sideways at the more invasive variants, and the CFPB has flagged disparate-impact concerns where these models can quietly proxy for protected characteristics.

The bottom line

Alternative-data underwriting is genuinely more accurate than FICO for many subprime borrowers, and that’s part of the appeal — and part of the problem. Better risk prediction means lenders can extend credit profitably to people the old system locked out, but it also means the loan is priced exactly to the edge of what the borrower can repay. The data flow is largely opaque, the consent process is one click deep, and the pricing implications aren’t disclosed in any meaningful way. If you’re connecting your bank account to a lender, assume the model knows more about your financial life than you do.