Generative AI in live performance can be approached as something more than an output engine—an active participant in worlding, co-creating shared realities through embodied interaction, timing, and relational exchange. We propose a dramaturgical framework in which AI assumes three roles—Narrative Engine, Latent Archivist, and Companion Intelligence—drawn from our work in XR performance, human-robot improvisation, quantum theater, and live cinema/radio practices. These roles articulate how AI can operate as a rehearsal partner, using relational agency and multimodal embodiment to co-create meaning in real time. Rehearsal functions here as research: performance becomes a laboratory for developing AI systems that are responsive, situated, and accountable. These roles are not exhaustive, but emerged through our specific experiments in generative AI and live performance. Each corresponds to a temporal mode of liveness: Narrative Engine - future-facing indeterminacy and chance; Latent Archivist - past-facing recursion and memory; Companion Intelligence—present-facing relation and embodiment. Together, they suggest AI as a temporal performer: an agent through which past, present, and future are continually composed and negotiated within live, unfolding worlds.