AI clauses in construction contracts and professional appointments
AI clauses in construction contracts arrived without anyone announcing them. There was no industry consultation and no new standard form. What happened instead is that clients began inserting bespoke amendments into appointments and tender documents, and most practices are signing them without reading closely, because the clause is short and sits near the end.
Some of what is being signed is unremarkable. Some of it is uninsurable.
Why AI clauses in construction contracts are bespoke amendments
The standard suites were not drafted with generative AI in mind. The NEC and JCT families, and the RIAI forms in Ireland, deal with skill and care, confidentiality, intellectual property and data — all of which touch AI use obliquely and none of which addresses it directly.
So the drafting is coming from client legal teams, and the quality varies enormously. A sophisticated developer's clause is usually proportionate. A clause copied from a technology procurement template into a consultant appointment frequently is not, because it was written for a software supplier rather than a design professional.
That difference in provenance explains most of the unreasonable wording currently in circulation.
The four AI clauses in construction contracts you will actually meet
Disclosure. You must notify the client where AI has been used materially in producing a deliverable. Generally reasonable, and easy to comply with if and only if you have a sign-off record. Firms without one find this clause quietly onerous.
Consent. AI may not be used without prior written approval. Workable when the approval is sought once for a category of use, unworkable when it is read as requiring approval per task. Negotiate the mechanism, not the principle.
Confidentiality and no-training. Project information must not be submitted to a tool that retains it or trains on it. This is the most common clause and the most reasonable one. It is also the one most often breached in practice, by someone using a consumer account on a phone.
Prohibition. No AI at all. Usually written broadly enough to catch spellcheck and translation if read literally, which tells you it was not drafted with much thought.
The AI clauses in construction contracts that should stop you signing
Three types of wording are worth escalating rather than accepting.
- A warranty that AI-assisted output is free from error. This imposes a standard higher than reasonable skill and care. It is a contractual liability assumed voluntarily, and it may sit outside your professional indemnity cover — which means the practice, not the insurer, carries it.
- An unlimited indemnity for AI-related loss. Uncapped indemnities are a problem whatever their subject. Attached to a category as loosely defined as "AI-related", they are worse, because the scope of the indemnity is genuinely unclear.
- Assignment of intellectual property in prompts, methods or trained material. Often buried in the IP clause. A practice that has spent a year building its own prompt library and workflow templates should not hand ownership of them to a client as a by-product of one appointment.
None of these are unnegotiable. All of them are worth a conversation, and clients raising them usually have not considered the consequence.
How to negotiate AI clauses in construction contracts
Lead with what you do rather than with what you object to. A practice that can produce a policy, a register and training records is in a strong position to propose alternative wording, because it is offering a control instead of a refusal.
Push disclosure toward categories rather than instances. "We use assistive tools for drafting, research and document review, subject to professional review before issue" is a workable commitment. Notifying every use is not.
Tie the standard back to skill and care explicitly. A sentence confirming that the consultant's obligations are unchanged by the tools used, and that all output remains subject to professional review, resolves a surprising number of badly drafted clauses without further argument.
Check the flow-down. If the clause binds you, it probably needs to bind your sub-consultants, and if your appointments with them are silent you have a gap you are carrying personally.
Answering an outright prohibition
It happens, and the instinct to argue is usually wrong.
Ask what the client is actually protecting. Almost always the answer is confidentiality project data leaving their control rather than an objection to the technology. A confidentiality-framed commitment, backed by tenancy-resident tools and a no-training position, resolves most prohibitions because it addresses the real concern.
Where the prohibition survives that conversation, comply with it properly. That means telling the project team, not just filing the contract, because the failure mode here is a junior engineer who never saw the appointment using a tool the practice promised not to use.
And price it. A genuine prohibition removes efficiencies you may have assumed in the fee. That is a legitimate commercial conversation to have at tender stage rather than a grievance to nurse afterwards.
What AI clauses in construction contracts will not do
They will not tell you whether your use is lawful. Contract and regulation are separate, and a client's permission does not discharge a data protection or AI Act obligation.
They will not stay still. This wording is evolving quickly, and a position agreed on one appointment will be renegotiated on the next. Keep a note of what you have accepted and where, because the first time a client asks what you agreed with someone else, you will want the answer.
They will not enforce themselves. A clause the project team has never read is a liability with no corresponding control, and that gap is where practices get caught.
Where AI clauses in construction contracts fit with everything else
The contractual position sets the outer boundary. Inside it sits the management system that lets you evidence compliance with what you signed, set out in ISO 42001 certification for construction and engineering firms, and the maturity picture that determines whether the practice can actually deliver on the commitment, covered in the AI capability framework. Where agents act with any autonomy, the oversight requirements tighten further, as set out in AI agent governance.
AI Institute works with bid and delivery teams on exactly this gap between what a practice has signed and what its people do the approach is described in how we work, and the scheduled programmes are on the courses page.
