Voice agents built on large language models are increasingly used for customer calls. But in regulated settings, finishing the task is not enough: the agent must also follow a conversational flow prescribed by law. A new arXiv preprint focuses on this problem in the context of debt collection, where legal rules shape how agents may speak and act.
The authors describe an approach for training voice agents to transfer to regulated procedures they have not seen before. Rather than only optimizing for task completion, the method aims to align agent behaviour with the procedural constraints that apply in a given domain. The work highlights a gap between general-purpose conversational ability and the need for rule-constrained interaction in high-stakes settings.