Construction AI BriefSubscribe →

Guide · updated 2026.11.08

Who is liable when construction AI makes a mistake?

The contractor or design professional who approved the AI's output is liable, not the vendor — vendor contracts cap damages at a few months of fees. Here's how liability actually splits by mistake type, and what your insurance and contracts say about it in 2026.

The contractor, engineer, or design professional who approved the AI-assisted work is liable when it turns out wrong — not the AI vendor. Vendor agreements are written specifically to prevent the alternative: liability caps that limit what you can recover from the software maker to a few months of license fees, no matter how large the resulting rework, delay, or injury claim actually is. If your firm is using AI anywhere in design, estimating, submittal review, or scheduling, you are the party holding the exposure, and the standard-form contracts your industry runs on don't say so explicitly yet.

That gap — real liability, no contract language addressing it — is why this question keeps coming up in 2026. Here's how liability actually splits by the type of mistake, what your insurance says right now, and what to put in writing before your next AI-touched deliverable becomes a dispute.

Why can't I just blame the AI vendor?

Because you agreed not to when you clicked through the license. AI vendor contracts routinely cap the vendor's liability at a small fixed dollar amount or the cost of the software subscription itself — legal analysis of these agreements points to caps commonly set at roughly one to three months of fees — which creates a gap between what a bad AI output actually costs your project and what you can recover from the company that built the tool (Dan Cumberland Labs, "The 8-Point AI Clause Audit for EPC Contracts"). Under the liability framework courts are applying, the "integrator" — the party that deployed the AI and benefited from its output — is the first party a claim reaches, ahead of the model developer, unless you can show the developer owed you a specific duty of care that most commercial license agreements are drafted to disclaim (HFW, "Legal Liability for AI Decisions: Who Is Responsible When AI Fails?"). In practice, on almost every AI tool your firm runs, you are the integrator.

Does liability change based on what kind of mistake the AI made?

Yes — legal exposure on a construction project sorts into three distinct categories, and AI can now be the root cause of any of them (AWCI, "AI and Automation at Construction Sites"):

  • Personal injury or property damage — an AI-enabled machine or jobsite camera system misreads a condition and someone gets hurt. This can pull in products-liability claims against the equipment manufacturer and the AI software vendor directly, on top of a standard negligent-supervision claim against whoever was operating or overseeing the equipment.
  • Construction defects — an AI tool drafts a spec update, a takeoff, or a submittal comparison that's wrong, and it gets built. Liability lands on whoever approved the AI's output without catching the error, litigated as ordinary professional negligence or breach of the standard of care.
  • Delay damages — an AI scheduling tool or document-review agent produces a bad output that costs time rather than causing physical harm, and the delay claim follows the contract's normal delay-damages allocation, with the AI's role as the contributing cause rather than a separate liability theory.

The category matters because it determines which policy, if any, responds — a products-liability claim, a professional-negligence claim, and a delay claim can each trigger different coverage, or none.

What does "hallucinated" AI output actually look like in a construction document?

It's not hypothetical. AI tools used to update specification sections or harmonize requirements across projects can generate "phantom codes" — plausible-sounding standard designations that don't exist in any actual standards catalog, confident enough in tone that a rushed reviewer waves them through (Got Law STL, "When 'Smart' Tools Make Dumb Mistakes"). The same failure mode has already reached construction litigation directly: in A2K Inc. v. Fluor Federal Services, a staffing subcontractor's attorneys used generative AI to draft and cite-check a legal brief in a construction-adjacent contracts dispute, and the filing allegedly included a fabricated quote attributed to a Texas Supreme Court decision that doesn't exist (Jones Walker, "Fake AI Citations Strike in a Construction Lawsuit"). The lesson generalizes past legal filings: any AI-drafted document that cites a standard, a code section, or a manufacturer spec needs a named human checking that the citation is real before it becomes a project record.

Does insurance actually cover an AI-caused mistake?

Less than it did over the summer, and the gap is widening by carrier. Most professional liability and CGL policies written before 2026 still treat AI-assisted work as ordinary negligence with no special exclusion. That's ending on the CGL side specifically: ISO issued three generative-AI exclusion endorsements effective January 2026 — CG 40 47 strips generative-AI-related bodily injury and property damage from Coverage A and personal/advertising injury from Coverage B, CG 40 48 is the narrower Coverage-B-only version, and CG 35 08 extends the exclusion to products and completed-operations coverage, which matters most for a defect that shows up after substantial completion (Gallagher, "ISO Introduces Generative AI Exclusion in Commercial General Liability Policies"). By April 2026, W.R. Berkley, Chubb, Travelers, and Berkshire Hathaway, among others, had filed to adopt ISO's language or their own version. The forms are optional for each insurer, so whether one is on your policy depends on your carrier — and the wording is broad enough that generative AI anywhere in the causal chain can be used to challenge coverage (Insurance Business, "ISO's generative AI exclusion is already on thousands of CGL policies"). On the professional-liability side, the same shift started earlier: Berkley introduced an absolute AI exclusion across its D&O, E&O, and fiduciary-liability products that names ChatGPT, Bard, Midjourney, and DALL-E specifically, and Philadelphia Insurance and Hamilton Select went further and excluded AI-related claims from E&O entirely (2026 E&O Renewal Audit Guide). A policy that covered your AI-assisted submittal review last year may not cover the same workflow at your next renewal — ask your broker to point to the specific endorsement form, not just confirm "you're covered."

Comparison: who's actually on the hook, by scenario

ScenarioWho's typically liableWhat limits or shifts it
AI drafts a bad RFI response or spec update that gets builtThe engineer, PE, or PM who approved itNamed human sign-off on every AI-assisted item, documented in the log
AI vendor's tool is objectively defectiveStill mostly you, up to a pointVendor's contractual liability cap (often ~1-3 months of fees) limits recovery from the vendor
Sub uses an AI tool the GC didn't specifyThe sub, under standard indemnification languageAn explicit AI carve-out in the subcontract, negotiated in advance
Sub uploads project documents to a public AI tool on a confidential projectThe sub for breach of confidentiality, possibly the GC to the ownerA subcontract clause requiring AI-tool disclosure, a no-training/no-retention warranty, and written consent
AI-enabled equipment causes a jobsite injuryContractor (negligent supervision) + possibly equipment maker and AI vendor (products liability)Documented training, maintenance, and human oversight of the AI-enabled system
Insurance claim traces to an AI-assisted deliverableThe firm carrying the policy, if the claim is deniedWritten AI governance policy; confirming which CGL/E&O exclusion endorsement is on your current policy before a claim, not after

What should a GC or sub actually do about this?

Four things, in order. First, put a named human sign-off on every AI-assisted deliverable that leaves your office — an RFI response, a submittal comparison, a spec update — so you have a documented standard-of-care trail if a dispute ever asks who approved what. Second, ask your broker directly which CGL and E&O exclusion endorsements are actually on your current policy — ISO's CG 40 47, CG 40 48, and CG 35 08 by name — and get the answer in writing before your next renewal, not during the renewal call. Third, on any project with real confidentiality exposure — data centers, medical or scientific facilities, energy sites, anything with an owner NDA — add a subcontract clause requiring subs to disclose which AI tools touch project documents and confirm those tools don't train on or retain the data (Snell & Wilmer, "Under Construction — August 2026") — the same account-level-access question xAI's Grok Bot rollout for enterprise raises for any AI agent your own team runs. Fourth, don't wait for AIA or ConsensusDocs to update their standard forms — AIA's January 2026 position statement is guidance on professional judgment, not a liability clause, and neither AIA nor ConsensusDocs currently addresses AI-generated work product or vendor data rights, so the fix is still a one-page side letter on your next contract specifying who verifies AI-assisted output and how an AI-caused dispute gets resolved (ConsensusDocs, "Handling Responsibility for A.I. in Construction").

None of this means slowing down AI adoption — it means making sure the paper trail exists before you need it. Federal work may soon add its own layer: a bipartisan bill introduced in the House in September 2026, the Stop Rogue AI Act, would make inventorying and logging every AI agent on your network a condition of winning new USACE, NAVFAC, GSA, or VA construction work if it passes — worth tracking now even though nothing is required yet. For the fuller risk picture, including prompt injection, vendor shutdown risk, and CUI restrictions on federal work, see our guide to AI risk, liability, and security in construction.

Construction AI Brief covers what's actually shipping for GCs, trade subs, and estimators — three times a week. Subscribe at constructionaibrief.com.

Frequently asked questions

Can I sue the AI vendor if its tool causes a construction error?

You can try, but the contract you clicked through when you bought the tool almost certainly caps what you can recover. Vendor agreements commonly cap AI-related damages at a fixed dollar figure or the cost of the software license — reporting points to a common range of one to three months of fees — regardless of how large the resulting rework or delay actually is.

Does my company's E&O or CGL policy still cover an AI-caused error?

Less than it did in July 2026. ISO issued three generative-AI exclusion endorsements for CGL policies effective January 2026 — CG 40 47 (excludes bodily injury and property damage tied to generative AI), CG 40 48 (a narrower version limited to personal and advertising injury), and CG 35 08 (applies the exclusion to products and completed operations). By April 2026, W.R. Berkley, Chubb, Travelers, and Berkshire Hathaway, among others, had filed to adopt ISO's language or their own version — the forms are optional for each insurer, so whether you carry one depends on your carrier. Check your renewal language now, not at the next claim.

Is a general contractor liable for a subcontractor's AI mistake?

It depends entirely on what your subcontract says, and by late 2026 more of them are starting to say something. GCs on projects with strict confidentiality obligations — data centers, medical and scientific facilities, energy sites — are now writing clauses that require subs to disclose which AI tools they use, confirm the tool doesn't retain or train on project data, and get written approval before uploading any project document to it. Where that clause is missing, a sub's standard indemnification language still applies to defects an AI tool helped cause, which is why negotiating the AI-disclosure clause before signing matters more than arguing over who's at fault after a claim.

What's the difference between an AI mistake that causes a defect versus one that causes an injury?

The liability theory changes even if the underlying failure looks similar. A defect or delay caused by AI is typically litigated as a contract or professional-negligence claim against the party who approved the work. A personal injury involving AI-enabled equipment can pull in products-liability claims against the equipment manufacturer and the AI software vendor directly, alongside the standard negligent-supervision claim against the contractor operating the equipment.

Do standard AIA and ConsensusDocs contracts already cover this?

Still no, even after AIA's first formal move on the topic. AIA published a Position Statement and Responsible Use Guidance in January 2026 stating that AI should support, not replace, professional judgment — but it's guidance, not contract language, and the 2026 editions of AIA B102 and EJCDC E-500 shipped without a dedicated AI-liability clause. ConsensusDocs has an AI drafting aid for dispute-resolution language, also not a published liability clause. The working fix industry-wide is still a side letter or addendum, not a wait for the next form revision.

Project
Construction AI Brief
Dated
2026.11.22
Sheet
1 / 1
Rev
A
Published independently · constructionaibrief.com · © 2026Facebook·Privacy·About