The energy consumption of AI and especially individual AI tasks is complex to measure. A critical aspect of the energy evaluation of AI systems is the precise definition of both the scope and methodology. It is not evident if the differentiation should occur at the task level or model level. Here it is argued that full task is the best entity for functional unit setting for LCA of AI systems. An example of data analysis is provided to show the usefulness and reasonability of the conceptual and analytical framework which helps identify hidden drivers. The proposed framework reveals that time-extended service phases are energy drivers which remain invisible in both interference-only and average LCA approaches. Main contributions are interaction-level energy accounting, theoretical expansion of existing LCA and scaling approaches and identification of dominate non-compute energy drivers.