A review of 126 studies finds AI cut building energy 14.7–28.1%, mostly in simulation

The gains come mostly from simulation. Only 18% of the academic studies (23 of 126) reported operational deployment, rising to an estimated 26% once industry reports were counted. Among the real-world implementations, performance fell 15–25% below simulation predictions in 43% of cases, and 78% ran into data quality and infrastructure problems.

The authors put payback periods at 2.1 to 5.8 years (median 3.4 years) and implementation costs at $8,000 to $47,000 per facility. Results vary widely by approach (I² of 45–81%), building type and climate, and the evidence leans heavily toward office buildings in temperate climates.

The authors searched Scopus, IEEE Xplore, ScienceDirect and SpringerLink for English-language studies published between January 2010 and March 2024. Two independent reviewers screened records, and savings were normalized as percentage reductions from conventional building management systems. The review was not registered in PROSPERO, and the combined dataset covered over 2.8 million building-hours from 45 countries.

Most of the evidence is simulation-based: 68% of studies were simulation-only, 14% used hybrid validation and 18% involved real-world deployment, typically university campuses and research facilities with validation periods of 3 to 6 months. Reinforcement learning appeared in 89% of studies published from 2020 to 2024, and 83% of studies validated over a single seasonal cycle.

On cost, implementation ranged from $8,000 for small offices to $65,000 for large offices, with annual savings of $2,400 to $47,000. Hardware accounted for 35–45% of costs, and each $0.05/kWh change in energy prices shifted payback by about 0.8 years.

Few studies addressed occupants or ethics: 88% treated comfort as a fixed constraint, only 8% explicitly addressed data anonymization, and only 3% considered fairness. Explainable AI appeared in 12% of studies and cost 8–12% in optimization effectiveness.