Collections has always relied on a blunt instrument: call the borrower, ask for payment, and try again tomorrow if it doesn’t land. Behavioral AI is starting to change that math. Instead of treating every past-due account the same way, it reads how a borrower actually responds—timing, tone, channel—and adjusts outreach accordingly. The pitch is that softer, better-timed messages recover more, sooner, and with less overhead than a phone-first approach ever could. The harder question is whether that pitch holds up against a real portfolio.
One lender decided to find out directly. Its collections team had tried texting past-due borrowers before, but the process broke down fast: every reply had to be read, answered, and logged by hand, which doubled the workload for every message sent. “It was double work, just too much involved there,” says the institution’s VP of Risk Management. The manual effort was shelved, but the instinct that texting could work stuck around.
That instinct led the team to KredosAi, an AI-driven texting platform, and to a more disciplined way of testing it: rather than switch over wholesale, the lender split its own delinquent accounts in half. One side kept working the phones. The other moved entirely to KredosAi’s AI-managed texting, which adjusts tone, timing, and channel in real time based on borrowers’ responses. Two years later, the split is still running.
“The KredosAi test pool is anywhere from half a point to three-quarters of a point less than the other test pool of normal calls,” the VP says, describing a gap that holds month after month in early-stage delinquency, the window where a late payment either recovers or turns into something worse. On a $100 million loan book, half a point is $500,000 that stays current instead of sliding further behind, every month.
Delinquency wasn’t the number the VP reached for first, though. Asked what he’d bring to his own CFO, he pointed to capacity: running that portfolio via text messaging meant the lender never had to scale up an offshore collections operation, roughly the cost and ramp-up of 10 additional hires, just to keep pace with a growing delinquent book. Response quality held up too. Softer, earlier-stage messages get answered more often, and not one borrower has complained.
None of it came from taking a vendor’s pitch at face value. It came from insisting AI prove itself against the institution’s own status quo before earning any trust. Getting there required building an AI vetting process from scratch, including confirming exactly where borrower data was and wasn’t going, but that was the price of admission, not the headline.
What the results didn’t change is the philosophy underneath them. “We don’t have the philosophy that AI is going to replace people,” the VP says. “We’re a people bank, and always will be.” The AI isn’t standing in for a person; it’s making sure fewer borrowers ever reach the point of needing a difficult phone call at all.
