Banks, lenders and card issuers
The first thirty daysare a factory.Run it like one.
Early delinquency is high-volume, low-discretion and tightly regulated, and on most floors it is still a person with a headset, a list printed overnight and a daily target. It is the part of the book where speed and coverage decide the outcome, and both are things software is better at than staffing.
The operating model
Automate the volume.Escalate the distress.
A book does not decay uniformly. A Day-1 miss is usually forgetfulness; a Day-90 account is distress and legal process. Putting the same treatment on both is how floors simultaneously over-contact the first and under-serve the second.
| Stage | What the customer is | The decision that has to be instant |
|---|---|---|
| Pre-due | Not delinquent. Needs a reminder and a way to pay. | Send the link now, capture intent, involve nobody. The cheapest cure in the whole book. |
| 1–30 days | Forgot, or a short cash-flow wobble. Still reachable and still willing. | Contact the entire bucket inside 48 hours, capture a structured promise, confirm it in writing. |
| 31–60 days | Hesitation. Credit reporting starts to matter to them. | Automated first touch, a person on the broken promise and on any complaint. |
| 61–90 days | Specialist territory. Field and legal are queued behind this. | Triage only (willing, distressed or disputing), then route to the right person. |
| 90+ days | Distress, restructuring and statutory process. | A person owns the recovery. Automation confirms logistics and keeps the clock. |
The engine reflects this shape rather than being configured into it: as exposure and days past due rise, the expected value of a human conversation overtakes the expected value of another dial.
Where the money is
Three changesmove the P&L.
- 01
Work the failure the hour it happens
A failed mandate opens a case the minute the bounce arrives, and the first compliant touch, a written notice with a pay link, goes out in the next permitted window. Re-presentation is proposed for when the account is likely to be funded, around the salary credit, with at least 48 hours between attempts, and a person confirms it before it goes to your loan system.
- 02
Cover the whole early book, not a sampled list
Coverage is the lift in early delinquency; conversation quality matters more further down. A floor that can only dial a fraction of the bucket is choosing which accounts to let roll, usually by ticket size rather than by where the intervention would have mattered.
- 03
Make promises into artefacts
Promises made is a vanity number. Amount and date captured on the call, confirmed in writing with a pay link before the call ends, a reminder on the day, and kept or broken worked out from the ledger rather than from anyone's honour.
Context
What the publishedresearch supports.
Third-party figures, given with their sources so you can weigh them yourself. These are published ranges for digital-first and generative-AI collections programmes generally. They are not measurements of a Collections Hub deployment, and we do not present them as ours.
Cost to collect
McKinsey reports up to a 40% reduction in operating cost for generative-AI-enabled credit customer assistance, and a 15% cost-to-collect reduction in a retail banking case study.
Recovery
The same body of work reports roughly a 10% recovery uplift, with more customers cured through self-service channels and digital payment volumes doubling in the case study cited.
Experience
Customer-satisfaction improvements of up to 30% are reported alongside those cost and recovery figures, an outcome that is unusual in collections, where cost reduction normally comes at the customer's expense.
Sources: McKinsey & Company, The promise of generative AI for credit customer assistance (2024) and Holistic customer assistance. We will walk through the methodology behind our own figures, and their limits, in the walkthrough.
What stays human
Judgement is nota volume problem.
The point of automating the early book is that hardship, disputes, settlements and complaints stop competing for the same hour.
Hardship
Job loss, illness, bereavement. A hold, then a person, quickly.
Disputes
Already-paid claims and fee arguments, as a queue with evidence attached.
Settlements
Always a person. Voice agents have no settlement or waiver tool, by design.
Complaints
Routed with the recording and the decision record already retrieved.
Fewer contacts, the same recovery, and an answer for every call we made.
Questions
What lendersask first.
Where should voice agents not be used?
Late buckets. Pre-due and early delinquency are coverage problems and suit automation; the middle is triage; distressed and legal accounts belong to a specialist. The engine reflects that: field visits only open up from 31 days past due and legal notices from 61, and as exposure rises it routes to people rather than dialling harder.
Do you work with our existing agencies?
Yes. An agency's own dialler records can be loaded and checked against the same caps and calling window, and agency floors that run on Collections Hub use the same gates, rubric and evidence trail as your own. That is what turns "our vendor handled it" into something you can actually show.
Start with onebucket, one month.
A pilot on a defined slice of the early book, with a comparison group, and an audit pack when it ends.