Review of 66 Studies Maps Where AI Agents Are Ready for HVAC Deployment

A systematic review of 66 peer-reviewed studies has assessed the readiness of large language models for use in HVAC and building automation systems. The research categorized studies across five application areas—energy modeling, fault detection, control, load forecasting, and occupant interaction—finding none had reached full operational deployment, with only four achieving pilot-level evidence. Point-name normalization and fault detection were identified as the nearest-term viable use cases, while direct control and load forecasting remain research-stage applications. A key finding is that LLMs are not yet suitable for issuing direct control commands, due to latency constraints and the risk of hallucinated outputs causing physical failures such as frozen pipes or overheated zones. Instead, the review recommends keeping LLMs in a semantic or advisory layer, with physics-based controllers handling actual system commands and human operators retaining oversight.