The way lenders decide who gets credit and on what terms is changing.
Institutions that once made a single credit decision per borrower and applied it to every purchase that followed are rebuilding the model around real-time behavior. What a consumer earns and owes today matters more than what their score said six months ago.
The shift is already playing out across consumer lending, B2B payments and the infrastructure beneath both.
This month, Plaid launched a sequential foundation model that reads the order and timing of a consumer’s full transaction history rather than reducing it to a single score, according to a company blog post. In early tests, the model cut default risk by 13.6% at a 70% approval rate and reduced losses from returned payments by 26.5%.
Overzealous fraud controls are costing merchants an estimated $50 billion annually, according to the PYMNTS Intelligence tracker “Orchestrating Trust: The Future of Fraud Prevention in Payments.” Nearly half of merchants estimated that up to 5% of legitimate orders are incorrectly blocked as fraudulent, turning away the customers that lenders’ risk systems were designed to protect.
Static Credit Scores Miss What Transaction Sequences Reveal
The difference between static and real-time credit model approaches shows up most clearly when two borrowers look identical on paper. Plaid described in the post two consumers with nearly identical 90-day financial summaries. They have the same income, average balance and number of overdrafts. For one, a single overdraft appears after an unusual expense. The account recovers on the next paycheck. For the other, loan repayments drain the account within 24 hours of each paycheck. Small transfers fill the gap. The cycle repeats two weeks later. A static model sees the same numbers. A model that reads sequence sees two different borrowers.
That distinction costs merchants as much as it costs lenders. The PYMNTS Intelligence tracker found that 85% of merchants said their top fraud challenge is stopping bad transactions without driving away good ones.
In March, Billtrust launched its Agentic Credit Lines product, embedding credit limit recommendations directly into the software finance teams use to track what customers owe. The product draws on data from 13 million business buyers and 25 years of B2B payment history to surface risk before it becomes a missed invoice.
“It’s time to stop reacting to risk and start getting ahead of it with technology that elevates human judgment instead of replacing it,” Billtrust Chief Product Officer Lee An Schommer said in the March 3 launch announcement.
At Deutsche Bank, Chief Risk Officer Marcus Chromik told McKinsey that credit reviews that once took days now run in near real time.
“Clients no longer wait weeks for a decision,” Chromik said, per the report.
The PYMNTS Intelligence playbook “The ABCs of AI Credit: A Playbook for Issuers” identified four metrics that now define issuer performance, including approval rate, false decline rate, operating costs and fraud losses. All four improve when the system reads behavior rather than applies fixed rules. Approval rates rise because the model distinguishes an unusual purchase from a genuine risk. False declines fall, meaning fewer good customers are turned away. Operating costs drop as fewer transactions need human review. Fraud losses fall because criminals who bypass fixed rules still leave traces in the transaction sequence.
Agentic Credit Underwrites Each Transaction Individually
The shift changes what a credit decision is, Affirm President Libor Michalek told PYMNTS in April. Rather than approving a borrower for a credit line and applying the same rules to every purchase that follows, agentic credit evaluates the transaction itself, including what it costs, what the consumer already owes and what their cash flow looks like this month.
“We’re taking into account … what that translates to on a per-month obligation, and how does that relate to their cash flow, their existing debt…,” Michalek said.
Borrowers that traditional scoring couldn’t identify become visible. Those already stretched get declined before the damage compounds.
The systems beneath the decisions, or the pipes that move data between a bank, a payment network and a merchant, need to be fast and connected enough to act before a transaction fires. The PYMNTS Intelligence playbook found that without that foundation, real-time credit intelligence can’t operate at scale. Established card issuers updating older systems face a more difficult transition than FinTechs and digital lenders that started on newer architecture from day one.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
