Interviews collect information through conversation, requiring interviewers to ask appropriate questions, adapt to respondent behavior, and recover structured information from dialogue. Yet, AI systems’ abilities as interviewers remain under-explored. We present InterviewBench, a benchmark for evaluating LLM-based AI interviewers through both local interviewer decisions and dynamic multi-turn interviews. The static setting abstracts real interviews into controlled local-decision instances, while the dynamic setting evaluates end-to-end interviewing through long-horizon conversational simulations. Built from real interview transcripts, InterviewBench contains 5,500 classification questions, 3,000 generative questions, and 184 dynamic scenarios. We further model interviewer evaluation as contextual decision selection, where the model must choose or generate the next interviewer action conditioned on dialogue context. Evaluating 23 open-source and proprietary LLMs, we find that current models can produce fluent interviewer utterances, but still struggle with strategy selection, explicit evidence elicitation, long-horizon interaction management, and recovery from non-ideal interviewee behavior.