A consumer is ready to pay. She's on a call about her overdue auto loan, she's agreed to an amount, and she wants it done. She's in the kind of hurry people are in when they've decided something and want to move on.

Before that payment can go through, she has to hear a specific set of words. The authorization language, the disclosure the law requires before money moves. It takes a few seconds, and in the moment it feels like friction, because she's already said yes and the AI agent can tell.

It read her tone at the start of the call, eased off when she was guarded, and found an amount that fit what she could pay. None of that was scripted, and if it had been it would have failed, because real people don't follow scripts. So the same agent, sensing a consumer who's ready and impatient, will reach for the helpful move and skip ahead to the part she wants. That instinct is exactly what you want in a conversation and exactly what you can't allow at a collections payment.

So Prodigal doesn't leave it to instinct. The disclosure has to happen before the confirmation can, and that ordering isn't a line in the agent's prompt asking it to remember. It's enforced underneath the conversation, where the agent doesn't get a vote. The agent can be as warm and quick as the call needs but it cannot skip these words. 

‍

‍

This is the idea behind how the whole agent is built. It's flexible in conversation on purpose, because collections calls are unpredictable and a rigid bot fails them. The decisions that have to be exact, the ones with a legal or financial right answer, are taken out of the model's hands and enforced by the system around it. A general purpose large language model gets these right most of the time, and most of the time is not a standard the Collections industry can run on. 

The required language shows the same logic. Disclosures and authorizations aren't generated by the model, because a model paraphrases and a paraphrase can drop a clause. They're produced exactly and handed to the agent to deliver word for word. The agent decides how to carry the conversation toward that moment but it doesn't get to rewrite the moment.

Before it will name a balance, a creditor, or the reason for the call, it has to know it's talking to the right person. A capable conversational agent, trying to be transparent, will lead with why it's calling. On a Collections call that instinct discloses someone's debt to whoever happened to answer the phone. Prodigal won't do it. It confirms identity first, and only then says why it's calling. 

‍

‍

The same design keeps the agent from being confidently wrong about money. When a consumer works out a settlement or a payment plan, the agent doesn't invent the terms or decide how to phrase them. The plan is checked against fixed rules, how low a settlement can go, how long a plan can run, what counts as good faith, and only a plan that clears every rule is offered, stated in the exact terms the validator returns. The agent carries the conversation toward the plan. It doesn't get to compose the plan, or round it, or soften it into something the account doesn't allow.

The engineering is there to act as a permanent guardrail, and in a regulated industry that boundary is the most critical part of the build. A collections agent is worth deploying only when it's safe and reliable first, and convincing second.