AI voice or a human caller: choosing per bucket, not per belief

Most arguments about AI calling in collections are conducted at the extremes. One side demonstrates a flawless scripted reminder and declares the floor obsolete. The other plays a recording of a confused bot and declares the technology unready. Both are answering a question no operator actually has, because the real decision is not whether to use AI voice but where in the book to use it.

What separates an easy call from a hard one

Early-bucket contact is mostly information transfer. The borrower knows they are late, the amount is not in dispute, and the useful outcome is a date and a payment link. That conversation has a shape, and a competent agent that follows the script, speaks the borrower language and captures the promise correctly does it as well as a person, at any hour, without a bad day.

Hard-bucket contact is negotiation. The borrower has a reason, the reason is often true, and the outcome depends on judgement about what this person can actually pay and what the lender will actually accept. That is human work, and spending human time on reminders is what leaves too little of it for the accounts where judgement changes the outcome.

A defensible split

• Pre-due and 1 to 30: automated voice and messaging, with a human path available on request.
• 31 to 60: automated first contact, human follow-up where the first contact produced a dispute, a hardship signal or a broken promise.
• 61 to 90: human-led, with automation handling reminders around agreed dates and confirmations after payment.
• 90 plus and legal-stage: human, with the platform doing evidence and scheduling rather than talking.

The split is a starting point, not a law. It should move as you learn, and the direction it moves is the interesting number: if automated contact keeps producing kept promises deeper into the book, push the line down. If it produces complaints, pull it up.

The three failure modes worth designing for

• The agent cannot understand the borrower. Two failed recognitions in a row should route to a person, not a third attempt.
• The borrower is distressed or angry. This needs a human immediately, and the handoff should carry the transcript so nothing is repeated.
• The conversation leaves the script. An agent that improvises in a regulated conversation is a liability; an agent that says it will have a colleague call back is not.

How to measure the decision rather than argue about it

Run both on the same cohort for a month. Compare kept-promise rate rather than contact rate, because contact is cheap and promises are the thing that turns into money. Compare complaint rate per thousand contacts separately for each. Compare the cost of the human hours you freed against what those hours produced when they were pointed at the harder bucket. The answer that comes out is specific to your book, which is why nobody else can give it to you.