Casting brief, identity sheet, trained identity, scene plates, motion, voice. Four days from brief to nine finished variants, after which variant forty costs what variant ten did, because the cast already exists.
The gates are the point: likeness, hands, consent, disclosure. Nothing renders until a signed release for the voice and the likeness reference sits in the folder, and nothing releases without a synthetic-media label that satisfies both platform policy and client legal.
How do you make a consistent AI presenter for ads?
Write the persona as a casting brief, shoot or generate twenty-plus stills at varied angles under even light, train an identity from that sheet, and build every scene from the identity. Never from a fresh prompt.