Debt collection calls do not reward conversational fluency but the ability to follow the right sequence. Before mentioning the debt, the agent must confirm it speaks with the debtor, identify itself and the creditor, state the amount, and avoid prohibited language. Researchers asked whether a small self-hosted policy could be trained once and then follow regulated procedures it has never seen. The answer comes from a 9-billion-parameter model fine-tuned on 2,506 conversations generated from 297 randomized procedures and evaluated on a Thai outbound-call agent.

This is not the usual conversational quality evaluation. The system uses FlowSpec, a machine-readable specification for each deployment. FlowSpec generates the policy instructions, gates the tools, provides supervision, and defines the tests. The agent can speak only approved catalogue sentences. This constraint changes the problem: the system does not need an LLM that improvises elegantly, but a deterministic and auditable component that respects a law-driven flow.

The results on four unseen deployment procedures — three debt collection and one non-debt, one from a fully withheld lineage — show documented checks fixed by the procedure rising from 67-71 percent to 93-100 percent. Grounded checks, determined by the conversation, already passed at 97-100 percent before training. The scenario pass rate rises from 15 out of 43 to 40 out of 43.

One detail prevents reading these numbers as a full endorsement: an audit of the shared specification finds 11 defects, three of which the system's own metric did not expose. In other words, passing the tests does not guarantee that the specification is correct. Those adopting a similar approach must separate two risks: the model can violate a procedure, or the specification can describe the procedure badly. The second risk is more insidious because it does not appear as a conversational error.

For teams evaluating on-premise deployment, this is not academic. A debt collection agent handles personal data and sensitive conversations: moving everything to an external cloud LLM creates compliance and data sovereignty constraints. The policy described in the research is small and self-hosted, so it suits a local architecture where procedure variability does not force retraining a model for each creditor, but rewriting a specification. This changes TCO: the recurring cost is not training but managing specifications and audits. The bottleneck shifts from VRAM and compute to the quality of regulatory documentation.

The structural lesson is that regulation is not a context to add downstream: it is an operational grammar that can become supervision and test data. Catalogue constraints, tool gates, and lineage isolation reduce the space of allowed responses. If the specification is unreliable, compliance becomes a metric illusion. The three defects not caught by the metric are the most important signal: governance cannot outsource everything to a benchmark.