Collections is a sequence problem.

Nobody has ever cured a delinquent account with a single perfect message. Cure comes from a sequence: a first contact that lands at a moment the borrower can absorb it, on a channel they actually read; a follow-up that arrives before the situation hardens; a hardship conversation instead of a demand, if the signals warrant it; sometimes a restructure that quietly saves the account. Which step, when, on which channel, or whether to change the instrument altogether. Cure is a path through a space of choices.

This is why the AI conversation in collections keeps missing. Ask most vendors what they automate and the answer is a step: draft the message, transcribe the call, score the account. But the value was never in the step. It was in the path.

Every step in the path is bounded by law

A collections sequence is not a marketing funnel, and this is the part that makes it hard. Every arrow in the path is constrained. Contact only within permitted hours. Only through channels this borrower consented to. Not after they have asked you to stop. Not to a number that no longer belongs to them. Say the things the law requires you to say, and nothing you are forbidden to say, in a manner that cannot be construed as harassment.

And the penalties are priced per action, not per incident. Under the TCPA, an illegal call or text carries $500 in statutory damages, rising to $1,500 where the conduct is willful. The FDCPA caps statutory damages at $1,000 per action. Those are unit costs. Multiply them by the volume automation exists to deliver, and the arithmetic turns on you. In 2014 Capital One settled TCPA claims for $75.5 million over autodialed collection calls that reached more than 21 million phone numbers. The dialer did exactly what it was told, at scale, with no gate in front of it.

Why this is where safe automation pays first

Collections has a property that most of the lending lifecycle lacks, and it is enormously valuable to anyone trying to build a system that learns. The outcomes arrive quickly. A default label on a fresh origination can take two or three years to mature. But whether this contact, on this day, on this channel, moved this account toward cure is knowable in weeks.

Short feedback loops are what let a system’s claims be checked against reality before anyone bets the book on them. You can learn a sequence policy here and find out, soon, whether you were right. That is not true of a credit model whose verdict lands in 2029.

What the system has to be able to do

Put the two facts together and the requirements write themselves. A system that runs collections has to reason about a sequence, because that is where the cure lives. And it has to be stopped at every individual touch, because that is where the liability lives.

Which means the contact is checked before it goes out, every time: the hour, the channel, the consent on file, the state of the account, the wording. What fails to clear never fires. What clears leaves the reasoning that cleared it on the record, so the supervisor reviewing the queue and the examiner reviewing the year are reading the same thing. And because that record accumulates, the system can tell, later, which sequences actually reached cure.

None of this makes collections gentle by itself; that is a matter of policy, and policy is the institution’s to set. What it does is make the institution’s policy actually govern what happens, at machine volume, touch by touch. The dialer that cost $75.5 million was not malicious. It was unsupervised at the only moment supervision would have mattered.