Judgment Day On The Jobsite: Accounting For AI In Construction — From Contract To The Field

When we last checked in on our friend the Terminator (Artificial Intelligence: Accounting for AI in Construction Documents), we left off at the front end — accounting for AI at the time of contract. But the machines did not stop at the contract. As James Cameron’s 1991 sequel taught us, the real story is what happens after the paperwork, when the technology walks off the page and onto the site. This is that sequel.

AI has steadily become a part of nearly every phase of a project, extending well beyond the contract itself. What began with technologies such as CAD and BIM has expanded to include drones, robotics, and now analytical and generative AI (GenAI). This next wave of innovation may be the most transformative yet, raising important questions about how risk should be allocated and managed throughout the construction process. Below, I discuss several legal concerns these developments present, from contract drafting and project management to on-site construction activities.

At Contract

• Permitted use and disclosure. Define where AI may be used and require disclosure. Model clauses call for legends on AI-generated deliverables and, critically, state which party must review the output for errors, because that allocation determines whose warranties answer for inaccuracies and hallucinations. Moreover, a growing number of statutes and rules now mandate disclosure of AI use with some requiring AI-generated output be identified as such and certain AI-driven recommendations undergo human review before implementation. These statutory and regulatory requirements are increasingly being incorporated into contract terms and public project requirements.
• Data, consent, and confidentiality. Confirm consent and licensing rights before feeding project data into an AI tool, and prohibit the use of that data for AI training or GenAI prompts without written consent. Vet any tool using non-confidential content first, and independently verify all outputs.

In the Field

• Guard against “garbage in, arguments out.” Project-administration teams are increasingly running day-to-day tasks through publicly available AI tools, often pasting potentially confidential or proprietary project information into open systems without consent or licensing in place. Worse, teams are then issuing project direction and administering the work based on that output, even though the tool may be operating on incomplete information, poorly framed prompts, or data it has simply misinterpreted, and AI cannot transparently show the analysis behind its answer. Keeping the human element in administration remains essential to curb this. When that direction turns out to be wrong, liability is genuinely unsettled, which is escalating arguments among owners, general contractors, and subcontractors.
• Standard of care/human-in-the-loop. The design professional’s standard of care, the skill and care ordinarily used on a similar project in the same locality, exists independently of the contract and does not require perfection. Whether routine AI use will reshape it is an open question; until then, keep a human in the loop, because AI lacks judgment and human verification remains essential. Professional judgment cannot be delegated to a machine.
• Uncertain liability. When an AI tool errs, liability is genuinely unsettled which is muddied by software disclaimers, disputes over inputs, and which party chose the tool. That is exactly why AI risk belongs in the contract.
• Autonomous tools, drones, and robotics. Where fault cannot be traced to a human, courts must determine whether the AI itself is the “actor” and what standard of care should apply (one court has already treated software controlling a robotic production line as a “product” for design-defect purposes following a worker’s death). Build in failsafes, human-intervention triggers, warnings, and vendor safety commitments.
• Jobsite privacy. Camera-based safety and productivity monitoring can capture workers’ facial scans, implicating biometric-privacy statutes that require notice, consent, and data protection.
• IP and insurance. Purely AI-generated content is not copyrightable, though human contribution can be, so document prompts and human edits. And do not assume existing insurance will cover AI issues. Traditional policies increasingly exclude cyber and data losses, and Errors & Omissions (E&O) and Commercial General Liability (CGL) policies may leave AI incidents uncovered.

Bottom line

Decide in writing who owns the AI risk and who verifies AI output, lock down project data, keep a human on the controls, vet autonomous equipment, mind biometric law, and recheck your insurance. Whether AI is the T-800 protecting your project or the one you wish had been sent back depends less on the technology than on how carefully the parties account for it — at contract and in the field.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.