Problem decomposition is an important skill in programming, allowing learners to break down complex tasks into manageable subgoals. However, translating these subgoals into executable code poses a significant challenge for novice programmers. In this study conducted in an introductory programming course, learners received instruction in stepwise refinement and integration of AI-generated code within their assignments. Throughout the course, learners were permitted to rely on AI code generators, following opportunities to receive feedback on their ability to read and write code without AI assistance. Our findings show that learners frequently relied on AI-generated code when working on assignments outside the classroom, but that the frequency of reliance varied from one assignment to another. The reliance on AI-generated code was not correlated with the learners’ year in their degree, nor whether they were enrolled in a CS degree or not. Instead, it was associated with their prior knowledge, as learners who were less proficient in reading and writing code were more likely to seek AI assistance. AI tools were primarily used to translate subgoals into code, fix errors, and explain algorithmic concepts. Few learners encountered difficulties in understanding or integrating AI generated code into their solutions. Overall, learner performance in meeting assignment requirements was relatively high, regardless of their prior knowledge or reliance on AI code generators. We conclude that leveraging the capabilities of generative AI can effectively bridge the gap between problem-solving and implementation, enabling learners to engage in skills that might otherwise be beyond their reach.