Artificial Intelligence (AI) on Construction Projects: Regulatory Position and Governance Gaps

Abstract

Construction organizations are adopting Artificial Intelligence (AI) software faster than governance practice has developed to control its use, and the products now reaching the sector differ in kind from the prediction models preceding them. Large Language Models (LLMs) return the most probable continuation of a text rather than the result of a lookup, and they fail fluently and confidently, in ways difficult to detect on the face of the output. Deployed as multiple interacting agents, they also fail at the hand-off, where no single owner observes the error. This paper establishes where the law binds such deployments. Under Regulation (EU) 2024/1689 as amended by the 2026 Digital Omnibus, three features decide the question: Annex III point 2 is a closed list that does not reach structures, so software assessing the condition of public infrastructure falls outside the high-risk tier; Annex III point 4(b) with Article 6(3) captures any function building individual behavioral profiles and forecloses the available exemptions; and Article 5(1)(f) prohibits inference of emotion in the workplace outright. United States (U.S.) instruments differ in kind rather than degree: the National Bridge Inspection Standards (NBIS), the model law of the National Council of Examiners for Engineering and Surveying and Policy Statement 573 of the American Society of Civil Engineers (ASCE) constrain the software only through what a qualified person must personally do. Both traditions converge on human oversight as the principal control, and four decades of automation bias research show the control failing wherever it is specified without being sized. Applied to a five-agent platform procured under one commercial decision, the analysis yields five distinct regulatory positions, from a prohibited practice question to a function no reviewed instrument covers. The unit of governance is therefore the function, not the product.