How do you take the precision, scale, and digital-first mindset of the automotive world and apply it to modern construction?

Find out live on June 3rd in London, where we sit down with special guest Chetan Kotur, Head of Technology and Innovation at Laing O'Rourke (and formerly of Polestar).

Please note: Spaces for this exclusive event are strictly limited. Registrations are subject to approval to ensure a curated room of industry peers.

When an AI-driven scheduling tool produces a faulty critical-path forecast and the project slips into liquidated damages, the first call goes to legal. The GC points at the software vendor. The vendor's terms of service cap liability at a nominal sum and explicitly disclaim errors in AI-generated outputs. The owner reviews the contract, finds no mention of AI anywhere, and the dispute becomes a maze of competing indemnities with no obvious exit.

This is a real-life scenario, not a fictitious one. In February 2026, ConsensusDocs published guidance on managing AI risk in construction contracts, framing it around what happens "when your scheduler hallucinates." The fact that one of the industry's two dominant contract families had to address this as an urgent operational concern in a standalone publication tells you how far ahead the tools have moved from the frameworks that govern them.

In most sectors, a bad AI output can be corrected before it causes lasting harm. A bad AI recommendation embedded in a structural design, a safety monitoring system, or a procurement specification cannot be recalled after the building is occupied. Built assets have operational lifespans of 40 to 60 years.

The data used to train a construction AI tool may be outdated within a decade, the outputs embedded in project records may be inscrutable once the original model is superseded, and the professional who validated those outputs may be long gone. A building informed by AI decisions made in 2026 will still be occupied in 2066, and by then the liability chain stretches across dissolved vendor agreements and long-superseded contract forms.

The parallel to Grenfell Tower is direct. That disaster exposed what happens when accountability gaps exist across complex, multi-party systems, where each party holds partial responsibility for safety-critical decisions and no single thread of documentation holds them together. The Building Safety Act 2022 responded with the "golden thread" concept: a continuous, auditable record of safety information throughout a building's lifecycle. The AI equivalent, in most construction firms today, does not exist.

The data on how people actually use AI at work doesn't flatter anyone. A 2025 KPMG global study of 48,000 workers across 47 countries found that 58% of employees admit to relying on AI to complete work without properly evaluating the outputs, while half the US workforce uses AI tools without knowing whether their employer even permits it.

Separately, a KPMG boardroom survey found that while 53% of boards have issued responsible-use policies, only 8% have established AI ethics boards, and just 5% have GenAI expertise represented at board level. In construction, where a decision embedded in a design or contract document can carry legal weight for the life of the asset, those numbers describe a very specific kind of exposure.

The AIA and ConsensusDocs contract families cover the overwhelming majority of US construction projects, and neither contains AI-specific provisions. Jane Kutepova, counsel at Michelman & Robinson, LLP, who advises on high-stakes construction litigation, has identified the core gaps: the contracts don't allocate responsibility for selecting or configuring AI tools, don't establish ownership of AI-generated outputs, don't address whether AI outages or cyberattacks qualify as force majeure events, and don't require human oversight or verification before AI outputs are acted upon. As Kutepova puts it: "Construction contracts are overdue for a software upgrade."

The mismatch runs deeper than omission. ConsensusDocs 200 includes written notice requirements that presume human observation and deliberate communication. Those requirements weren't drafted with the assumption that an AI dashboard might generate an alert, the alert might go unreviewed by qualified personnel, and the unreviewed alert might later appear as evidence in a dispute. Whether a dashboard notification qualifies as "written notice" under ConsensusDocs is genuinely contested ground, and it's the kind of question that tends to get answered expensively in arbitration.

Every project call now has an AI note-taker in the room. That feels like progress. Here's where it gets complicated.

AI note-taking agents produce an interpretation. A summary of what the agent decided was important, filtered through a model that has no understanding of construction contracts, scope boundaries, or the difference between a discussion about Block A and a decision about Block B. Construction attorney Michael Vardaro, 30 years in construction law, has seen cases where the minutes were, in his words, "flat out wrong": two separate project discussions compressed into a single entry that no participant would recognise as accurate.

The liability sequence that follows is straightforward and underappreciated:

"All of a sudden this tool that we thought was really going to help with communications actually made it worse," Michael observes. The fix is simpler than you might think: for any call where a decision of contractual consequence gets made, run a full transcription. It will be long and imperfect, but what it captures is what was actually said.

We covered this in full with Michael on the podcast; the full conversation is worth your time.

Intellectual property adds a further layer of uncertainty.

AI-generated designs, models, and project specifications may carry ambiguous ownership, particularly on projects where multiple parties contributed to their creation. Browne Jacobson's 2026 construction sector legal analysis noted that some AI tools explicitly retain rights to generated designs within their license terms, a fact that most firms aren't scrutinizing during vendor procurement.

The contractual exposure here is not abstract anymore. Michael shared a case where a counterparty used ChatGPT to redline a contract, and the AI-generated language ended up arguing against the party who submitted it. Nobody caught it until a human lawyer reviewed the document. Those are the disputes being seeded today. The AAA's new AI arbitrator, active since November 2025, is already beginning to reshape how they get resolved.

Construction's multi-party structure makes liability harder to assign than in almost any other sector, and AI deepens that complexity considerably. A vendor's algorithm, a firm's input data, the professional who approved the output, and the owner who relied on it in a contract decision can each carry exposure, and none of the standard forms directs the inquiry toward any of them.

Vendor liability caps make this particularly acute. AI tool agreements, following patterns established by major platforms including OpenAI, routinely limit the vendor's financial exposure for errors in AI-generated content. A contractor who relies on a faulty AI cost estimate cannot treat that reliance as a recoverable cost if the vendor's terms explicitly disclaim liability for outputs. The exposure stays with the construction party unless contract language assigns it elsewhere, and almost no construction contracts currently do.

Professional indemnity insurance is moving faster than contracts. In its 2025 construction PI market review, Miller Insurance noted that proposals now routinely include AI-related questions, with insurers closely examining how firms use generative AI in design, planning, and forecasting. The concern is structural:

The AI-related PI claims that will eventually test the market haven't materialized yet, but the tools generating them are deployed today.

RICS addressed accountability within the surveying profession in the standard that took effect on March 9, 2026 for all members and regulated firms worldwide: "AI tools cannot replace the surveyor's professional judgement, skill and scepticism. A qualified named surveyor must own every AI-assisted determination." That's a clear standard for one slice of the project team. For the broader liability chain across owners, GCs, designers, and subcontractors, ownership of AI-assisted decisions remains fully open.

Two regulatory deadlines matter for construction executives right now.

The EU AI Act's high-risk AI system obligations become generally applicable on August 2, 2026. Construction firms operating in or supplying the EU market need to assess whether tools used in safety-critical applications, worker monitoring, or automated decision-making fall within the high-risk category. Most haven't started that assessment. In the US, there is no equivalent federal framework, and the void is being filled unevenly by project-level bespoke clauses and vendor-initiated terms.

For firms regardless of geography, the practical action is contractual:

As we emphasized in the March Executive Briefing, firms that use AI only to accelerate existing workflows are leaving the real value on the table. The liability gap described in this piece is where that same pattern turns into a balance sheet problem.

Construction has always grounded accountability in documentation: who specified what, who approved it, and who signed off. AI changes how that work gets done but doesn't transfer accountability away from the humans in the chain.

When an AI-influenced decision turns out to be wrong and the question arises of who validated it, using what process, against what standard, most firms today cannot answer cleanly.

That's the gap that will define the first wave of AI-related construction disputes. The contracts being signed right now are the evidence that will either support or undermine those answers.

POWERED BY: