Both on-device and cloud AI can seem plausible regarding battery life and efficiency, but these are not measured claims. When deciding AI placement, consider the four distinct budgets: user wait time, network transfer costs, provider expenses, and device energy consumption. Each of these requires its own specific measurement and evidence.The execution path must be clearly identified to make accurate comparisons. For instance, the reviewed MonkeyCode mobile code uses server-supported streaming for tasks and speech-to-text, indicating cloud-based inference, not on-device. A fair study would compare a mobile client using remote services against a separate prototype demonstrating on-device capabilities.A comprehensive measurement envelope should include fields like sample ID, type, placement, device, OS, framework, model, network type, token counts, latency, data transfer in bytes, energy in joules, and cost in USD. These details are crucial for interpreting results and understanding workload size and network behavior. Battery percentage is an insufficient metric for short runs due to numerous external influences.Comparisons must utilize matched user flows, ensuring the same tasks are tested across different placements. This includes short prompts, voice turns, offline scenarios, background/resume behavior, and thermal loops. Warm-up periods should be reported separately, and tests should be randomized, repeated, and failures recorded.An analyzer should prevent false energy conclusions by requiring measured joules for each data point. Synthetic data is useful for testing parsing but does not represent actual performance. In a real pipeline, data provenance should be strong, including profiler exports and raw file preservation.Release decisions should be explicit, based on meeting targets for P95 interaction latency, network bytes, provider spend, energy and thermal behavior, privacy, and quality. On-device AI introduces download size and RAM pressure, while cloud AI relies on network connectivity and service dependency. Using clear units ensures an honest assessment of these tradeoffs.