Advances in artificial intelligence raise fundamental questions about how technology reshapes human skills. Modern AI systems may crowd out active human decision-making; unlike past technologies, such decisions are a key input for training and improving AI models. We develop a framework based on Markov matrices to study how AI substitutes for and augments different human skills, and how decision authority -- whether the technology is assistive or autonomous -- shapes these effects. We focus on two distinct skills: innovative skill, the ability to generate correct outcomes from incorrect inputs, and executional skill, the ability to reliably produce correct outcomes from correct inputs. We show that the presence of technology crowds out active human decision-making among individuals with low decision quality. Moreover, adoption patterns differ sharply across technologies: autonomous technologies complement innovative skills, whereas assistive technologies complement executional skills. Our model offers empirically testable relations among types and qualities of AI, their sensitivity to training data, human decision patterns, and overall welfare.