Engineering and architecture
AI clauses in owner-architect agreements: what clients now ask for
By SourceX Editorial · Reviewed by Noah Loul ·
Short answer
AI clauses in owner-architect agreements can be grouped into five types: bans on training third-party models with project data, disclosure or approval of AI use, ownership of AI-assisted work, indemnities and warranties, and audit or data-handling rights. Each may limit future licensing, so tag every project that carries one in the project database when the contract is signed.
Key takeaways
- Training prohibitions are the clause type most likely to exclude a project from any future data license.
- The defined term a clause attaches to, such as project data or confidential information, decides how far it reaches.
- Disclosure and approval clauses mainly govern tool use during the project, but broad wording can reach later uses.
- Tagging AI clauses at signing means future rights reviews rely on records, not memory.
What AI clauses are owners adding to design contracts?#
Owners are adding AI clauses to design contracts that control how the architect uses AI tools and what may happen to project information. Some arrive as riders to standard AIA forms; others are built into the owner's own form of agreement, its data security exhibit or its master services agreement.
Clients, their counsel and industry groups have started circulating model language, so similar wording now shows up across very different owners. For a general counsel or risk manager the task is the same either way: understand what each clause controls, negotiate where it overreaches, and record it where future decision-makers will find it.
Clause types and what they mean for future licensing#
Each AI clause type affects future licensing differently, and some barely touch it. The table summarizes the main types, their usual aim and the tagging action that keeps the archive reviewable years later.
| Clause type | Usual aim | Effect on future licensing | Tag as |
|---|---|---|---|
| Training prohibition | Keep project data out of third-party model training | Often excludes the project entirely | Exclude unless counsel clears a carve-out |
| Disclosure or approval of AI use | Owner knows about or approves AI tools on the project | Usually limited, but may require notice for later uses | Review |
| Ownership of AI-assisted work | Owner owns or receives rights in AI-assisted outputs | Can move deliverables out of the firm's control | Review deliverables |
| Indemnity and warranty | Architect stands behind AI-assisted work | Indirect; raises the stakes of any later dispute | Refer to insurance review |
| Audit and data handling | Owner can audit tools, storage, return and deletion | May require deletion or return of project data | Check retention and deletion terms |
Training prohibitions: read the defined terms#
A training prohibition's reach depends on the defined term it attaches to. A clause that bars training on the owner's confidential information reaches less than one that bars training on all project data, and both differ from a clause that covers the architect's instruments of service.
A firm negotiating new work can often narrow these clauses without dismissing the owner's concern. Proposing a carve-out for records de-identified under a stated standard, or limiting the ban to owner-furnished information, protects the owner while keeping the firm's internal process records usable.
Example carve-out wording to discuss with counsel: This restriction does not apply to the Architect's internal design review, quality control and process records after they have been de-identified so that they do not identify the Owner, the Project or its location, provided that no Owner-furnished information is included.
- Which defined term does the prohibition use, and how is that term defined elsewhere in the agreement?
- Does it cover the firm's internal records about the project, or only information the owner provided?
- Does it bar only third-party models, or also the firm's own internal tools?
- Is there a carve-out for de-identified or aggregated information, and is the de-identification standard stated?
- Does the restriction survive completion or termination, and is there an end point?
Disclosure, ownership and indemnity clauses#
Disclosure, ownership and indemnity clauses mostly govern how the project is delivered, but each can reach the archive later. Disclosure clauses may require notice before AI tools touch project information; if worded broadly, a later licensing program could count as a use that needs notice.
Ownership clauses for AI-assisted work can shift deliverables to the owner, which removes them from what the firm can license. Indemnities and warranties about AI-assisted work raise the stakes of any error and should go to the firm's professional liability broker, since coverage for AI use varies by policy.
Audit and data-handling clauses deserve a separate read. A right to audit the firm's tools and storage, or a duty to return or delete project data at closeout, can leave nothing to include in a later package, so retention and deletion terms belong in the project tag.
Where AI language hides in the contract set#
AI language often sits outside the main agreement, so reading the signed owner-architect form alone will miss it. Risk managers should search every document that forms part of the contract, including those issued after signing.
Search each document for terms such as artificial intelligence, machine learning, model training, automated tools and generative, and check the definitions section for any of those words. A defined term buried in an exhibit can quietly widen a short clause in the agreement.
- Owner riders and supplementary conditions attached to a standard form.
- Data security, information technology or confidentiality exhibits.
- Master services agreements and the task orders issued under them.
- Owner design standards and BIM execution plans, which can set rules for model data.
- Amendments and change orders issued after the original signing.
Which projects to tag, and how#
Projects to tag are every project whose agreement, rider, exhibit or amendment mentions AI, machine learning, model training or automated tools, plus projects with broad confidentiality or ownership-transfer terms. Tagging at signing takes minutes; reconstructing the same facts from scanned contracts later takes far longer.
These fields can usually live in the project record in Deltek Vantagepoint, BQE Core or a contract management system, as custom fields or a linked table. Public-sector, security-sensitive and critical infrastructure projects deserve their own flag regardless of AI language.
| Field in the project record | Example values |
|---|---|
| AI clause present | Yes, no, unknown |
| Clause types | Training prohibition, disclosure, ownership, indemnity, audit |
| Scope of restriction | Owner information only, all project data, instruments of service |
| De-identification carve-out | Yes with a stated standard, yes without a standard, none |
| Survival | Ends at completion, survives, unclear |
| Reviewed by | Counsel or risk manager, with review date |
Illustrative: a risk manager meets a developer's AI rider#
Illustrative: a fictional architecture firm received a rider from a multifamily developer client that barred use of all project data to train any AI model. The firm's risk manager compared it with the firm's internal AI policy and with its interest in licensing de-identified review records.
The risk manager proposed narrower wording: no training on owner-furnished information or identifiable project data, with a carve-out for records de-identified under a stated standard and reviewed by the firm. The developer accepted the carve-out but kept a requirement for written notice before any outside use.
The project was tagged with a training restriction, a de-identification carve-out and a notice obligation. When the firm later assembled a licensing package, the rights review found the tag at once and routed the project to the notice step instead of excluding it.
How SourceX uses contract tags in rights review#
SourceX relies on contract tags during the Rights step of the SourceX five-step transaction. Projects with training prohibitions are excluded by default. Projects with carve-outs or notice duties move forward only when the firm and its counsel confirm the conditions are met, and silent contracts are reviewed one by one.
The SourceX Evidence Packet for each package records the licensing rights and permitted use behind every included project, so the firm can later show why a project was in or out.
Frequently asked questions
Do new AI clauses apply to projects signed earlier?
Generally a new clause governs the agreement it appears in, not earlier projects. Watch for master agreements, task orders and amendments that update terms across existing work, since an owner's new standard terms can reach ongoing projects that way. Counsel should confirm how each instrument applies.
Does a confidentiality clause alone prevent licensing?
Not necessarily, but it is often the deciding clause. Much depends on how confidential information is defined, whether de-identified records still fall within it, and what exceptions exist. Read it together with the ownership terms and any AI rider before deciding.
Can an owner's AI clause bind our consultants?
Owner agreements often require the architect to flow terms down to consultants. Where consultant agreements include those flow-downs, consultants carry the same restrictions on their own records. Check the consultant agreements as well as the prime contract before assuming either way.
Should we propose our own AI clause first?
Many firms find it easier to offer balanced language than to react to an owner's draft. A short clause describing how the firm uses AI tools, how it protects owner information and how de-identified records are handled sets expectations early. Have counsel prepare the wording.
Do these clauses affect professional liability insurance?
They can. Indemnities and warranties about AI-assisted work may go beyond what a policy covers. Share new AI clauses with your broker, and ask how the policy treats AI use and any data licensing program the firm is considering.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.