We investigate a phenomenon we call micro-productivity decrease, which is expected to be found in most development or maintenance projects and has a specific profile that depends on the project, the development model and the team. Micro-productivity decrease refers to the observation that the cumulative effort to implement a series of changes is larger than the effort that would be needed if we made the same modification in only one step. The reason for the difference is that the same sections of code are usually modified more than once in the series of (sometimes imperfect) atomic changes. Hence, we suggest that effort estimation methods based on atomic change estimations should incorporate these profiles when being applied to larger modification tasks. We verify the concept on industrial development projects with our metrics-based machine learning models extended with statistical data. We show the calculated micro-productivity profile for these projects could be used for effort estimation of larger tasks with more accuracy than a naive atomic change oriented estimation.