The uses we will not support, stated as rules rather than as guidance. Deceiving a caller about material facts, outbound contact without recorded consent, giving regulated advice without a licensed human in the loop, and any use designed to harass or defraud the person on the other end.
An agent may open a call without announcing it is AI, which is legal in most contexts and common practice. It may not deny being AI when a caller asks directly. That line is enforced in the system prompt and tested in the eval suite, not left to model discretion.
A first breach gets a written notice and a remediation window. A breach involving unconsented outbound or a direct denial of AI status is grounds for immediate suspension, because the exposure is legal rather than reputational and it accrues to both parties.