Productive Failure (PF) engages students in problem-solving before instruction but designing effective PF problems is challenging. To address this, we developed ProductiveMath, an AI-powered tool to support teachers in generating PF problems. Across three studies, we explored five research questions related to problem quality, AI-generation, assessment accuracy, and teacher perceptions. In Study 1, we conducted a literature review to define high-quality PF problems, created a rubric, and used GPT-4o to generate and evaluate 30 problems alongside human raters. Study 2 replicated this process with 60 additional problems, showing strong correlations between AI and human ratings, confirming GPT-4o’s capability to produce high-quality problems. In Study 3, seven math teachers evaluated human- and AI-generated problems through surveys and interviews. They rated AI-generated algebra problems as high-quality, reported positive perceptions of ProductiveMath’s usability, and expressed intentions to use it. Key teacher feedback included suggestions to adjust problem difficulty, simplify text, enhance visuals, and provide additional PF support.