Artificial Intelligence (AI) Agents in Construction: A Conceptual Framework Across Autonomy Types and Application Domains

Abstract

Construction is one of the least digitized and most hazard-prone industries, and signs of risk often go unconnected to timely action, as the 2025 NIST findings on the Champlain Towers South collapse illustrate. Artificial intelligence (AI) agents have been proposed as a response, but the term spans computer vision monitors, physical robots, and large language model (LLM) reasoning systems studied in separate communities. This paper develops a conceptual framework connecting agent autonomy to the construction problem it addresses. The study is a structured review and synthesis of peer-reviewed and industry literature published mainly between 2014 and 2026, drawn from construction engineering, robotics, and computer science. The framework organizes agents along two axes: autonomy type (perceptual, cognitive, embodied) and application domain (safety monitoring, design and modeling, project and contract management, progress monitoring, on-site automation), rating the maturity of each pairing by weight of evidence. Perceptual agents are mature where the task is to see and report, such as safety and progress monitoring. Embodied agents are mature only for narrow physical tasks such as aerial survey. Cognitive agents reach every domain but are mature in none, and remain unreliable for safety-critical calculation. The most capable systems combine agent types rather than one alone. The framework provides the first two-axis view that unifies three separate agent literatures and makes explicit where construction capabilities are mature, overlapping, or absent. It gives transportation agencies a tool for matching an agent type to a construction need and prioritizing investment, and identifies three deployment requirements, namely human oversight, system integration, and trust, that practitioners must meet for AI agents to move from pilots to practice.