The International Energy Agency's Energy and AI report, published in April 2025, examines both the electricity consumed by AI infrastructure and the potential for AI to improve efficiency across energy systems including buildings. The residential section of that story is more constrained than the headline figures suggest. For occupied homes, the AI applications with clearest near-term evidence involve: predictive temperature control that avoids heating or cooling unoccupied periods; fault detection in HVAC systems that catches degraded performance before it accumulates; tariff-responsive scheduling of water heaters, dishwashers, and EV chargers; and forecasting to reduce battery charge/discharge cycles in systems with time-of-use electricity pricing. Each of these depends on sensors, commissioning, a stable internet connection, and building physics that allow response without comfort penalty. The DOE's Grid-Interactive Efficient Buildings programme describes how buildings can provide flexibility services to electricity networks. The NIST AI Risk Management Framework, while not building-specific, is a useful lens for evaluating reliability, transparency, and fallback behaviour in systems that control thermal comfort, appliances, or safety-adjacent functions. Lawrence Berkeley National Laboratory's Buildings group publishes research on measured performance of advanced controls and demand response in real buildings. Editorial caution: Treat projected savings as scenario- and application-specific. Require disclosed baselines, climate, building type, comfort constraints, and whether results come from occupied buildings or simulations only.