A cardholder calls your contact center on a Tuesday morning. She missed November’s payment after a hospitalization, and there’s a $340 charge from before that she’s still disputing. She gets routed to the disputes queue, then to collections, then to a hardship specialist who can’t see what the dispute team already promised her. By Friday, she’s stopped answering the phone.
Most credit card collections operations are built to handle that customer the way they handled her in 2014. They’re phone-heavy and channel-siloed, measured by activity rather than outcomes. The economy that produced those operations is gone, and the customers who tolerated them are gone with it.
The macro picture is more nuanced than the trade press suggests. Total household debt reached $18.8 trillion in Q4 2025, and credit card balances hit a record $1.28 trillion. But aggregate credit card delinquency rates have actually stabilized over the past six quarters. The pain is concentrated. The St. Louis Fed found that delinquency in the lowest-income 10% of ZIP codes climbed from 14.9% in Q3 2022 to 22.8% by Q1 2025. Two cardholders sitting at 60 days delinquent are not the same customer, and an operating model that treats them identically is a problem.
A recent PAG post argued that your AI strategy is really a data strategy. The same logic applies one layer up. Your collections strategy is really an operating model strategy, and the operating model has three stages.
Level 1: Reactive and manual
The dialer is the strategy. Right Party Contact is the headline metric on the monthly report. Promise-to-Pay (PTP) numbers look respectable.
What’s missing from the report is the kept-promise rate, which is the percentage of those PTPs that convert to dollars in the bank. In most Level 1 operations, a meaningful portion of the pipeline the executive team is tracking is theater.
Level 1 operations also tend to confuse policy with discipline. There’s a Collections Policy Manual on someone’s shared drive. Nobody on the floor has opened it in 18 months. Hardship offers get made by whichever supervisor is closest to the agent’s chair when the customer starts crying.
This level passes audits in good economies. It doesn’t pass audits when the CFPB starts looking at complaint trends.
Level 2: Digitally enabled and compliant
Segmentation exists. Accounts are bucketed by delinquency age (30, 60, 90, 120+ days). The dialer is paired with SMS and email. Reg F discipline is real, with 7-in-7 call frequency limits coded into the system, validation notices going out on time, and three-year retention documented.
This is where most large-bank operations live today. It’s good. It’s not enough.
The trap at Level 2 is mistaking bucket segmentation for behavioral segmentation. Two cardholders sitting at 60 days delinquent are not the same customer. One missed a payment because she changed jobs and the autopay didn’t follow her. The other is six months into a slow financial spiral and choosing which bills to pay this month. A Level 2 operation treats them identically. A Level 3 operation doesn’t.
Level 3: Predictive and relationship-oriented
Behavioral propensity models drive the work queue, not delinquency age. The system knows which accounts are likely to self-cure, which need a single well-timed nudge, and which need a hardship conversation before the next billing cycle hits.
Channel strategy follows the same logic. McKinsey research found that 73% of customers in late delinquency made a payment when contacted through digital channels, yet most issuers still default to phone calls after 30 days. Customer preference for email and text is a fact in the data. Lenders that ignore it leave money in the field.
Offer strategy runs as a test-and-learn lab. The team knows which combinations convert better for which segments because they tested them on actual accounts last quarter and the data said so.
Hardship programs are operations, not exceptions. The cardholder from the opening scene gets routed to a single specialist who can see her dispute, resolve it, and structure her hardship plan in one interaction. The complaint that would have arrived eight days later doesn’t.
Most importantly, Level 3 operations measure what matters. Activity metrics still exist, but they sit below outcome metrics: kept-promise rate, complaint rate per 1,000 accounts, CEI by segment, and cost-to-collect per dollar recovered.
Three questions for the CEO
If you’re not sure which level your operation is at, ask your head of collections three questions this week.
- What is our kept-promise rate, and how has it moved over the past four quarters? If the answer is “we don’t track that separately,” you’re at Level 1.
- What percentage of our queue is prioritized by behavioral models versus delinquency age? If the answer involves the word “buckets” and nothing else, you’re at Level 2.
- When a customer has both a dispute and a hardship request, how many times does she get transferred? If the answer is more than once, the operating model is the problem.
Collections is one of the few functions in a card business where the operating model is the strategy. Get it right and recovery improves while complaints fall. Get it wrong and you’re funding a 2014 operation with 2026 charge-off rates.
The data work was the prequel. The operating model is the work.
Predictive Analytics Group helps clients cost-effectively bridge the gap between analysis and impactful strategic or tactical decisions. To learn more about how GOBLIN can consolidate your data environment and accelerate your AI readiness, schedule a discovery call with us.
David LaRoche is managing partner of U.S. operations for Predictive Analytics Group.
