The bright line
State producer licensing draws a consistent line: an unlicensed caller — human or AI — may greet, qualify, gather information, and route. Quoting rates, recommending coverage, binding, and enrolling are licensed acts. The same logic gates mortgage under the SAFE Act and Medicare under CMS rules. Any AI-voice design for these verticals starts from that line, not from what the model could say.
Structural, not aspirational
The wrong way to honor the licensing line is a prompt that asks the AI to behave. The right way is architecture: the AI's role ends at qualification and consent by design — plan specifics, pricing, and recommendations are not in its lane, and questions that cross the line route to a licensed human. The handoff carries context forward: who the caller is, what they qualified for, what they consented to, with the trail attached.
Why this is a selling point, not a limitation
A compliance officer's first question about AI calling is "what stops it from selling?" When the honest answer is "the architecture — and every call is QA-scored against that boundary," the conversation changes. The floors that win with AI in regulated verticals are the ones that treat the licensed-human handoff as the product, not the compromise.