Can You License Client Engineering Data to AI Companies?

Can Engineering Firms Legally License Client Data to AI Buyers?

Your firm’s decades-old drawings and site reports could be valuable to AI developers, but owning the copyright doesn’t mean you’re free to sell that data. What’s actually hiding in your confidentiality clauses might surprise you.

Key Takeaways

  • Holding copyright over a design does not automatically give a firm the right to disclose or license the underlying project data for AI use
  • The original client contract, not the firm’s own preference, decides whether project records can be shared, especially its confidentiality, permitted-use, and survival clauses
  • A client’s sign-off on using documents for construction or maintenance rarely stretches to cover AI training or other commercial reuse
  • Stripping names from CAD or BIM files may still leave behind coordinates, revision histories, usernames, and distinctive geometry that identify a project
  • Archive Trove can help firms screen and organise records before any sensitive files are shared, while the firm decides whether to proceed

Engineering firms sit on decades of drawings, specifications, and site reports that AI developers increasingly want for training and testing their systems. Before chasing that opportunity, a firm’s principal needs to work through a tangle of contract law, professional ethics, and data hygiene that most project agreements never anticipated.

A closer look at how this plays out for engineering practices is set out in a guide on licensing client engineering data, worth reading before any archive gets anywhere near a buyer’s inbox.

Copyright Isn’t Permission to Share Data

A common assumption trips up even experienced principals: holding the copyright on a design means the firm can do whatever it likes with the underlying data. Copyright confirms authorship and control over reproduction of the instruments of service, such as drawings, specifications, and computational models, but it says nothing about whether that information can be disclosed to a third party. Under standard agreements like the AIA’s B104-2017, the architect or engineer retains copyright while the owner receives only a project-limited licence to use those documents for construction, operation, and maintenance.

That gap between ownership and disclosure rights is exactly where licensing deals to AI buyers can go wrong. A firm might rightly own the copyright in a structural calculation sheet, yet still be bound by a confidentiality clause that blocks sharing the client’s site data with anyone outside the project team. Owning the copyright answers a different question than having permission to disclose, and the two should never be confused when a buyer comes calling.

This distinction matters more as AI buyers specifically seek real-world technical and operational data rather than generic design templates. The value they are after, decades of site inspection logs, QA files, and decision trails, is precisely the material most likely to carry confidentiality obligations tied to a named client or project.

What the Original Contract Actually Allows

Every licensing conversation with an AI buyer should start with a plain reading of the contract that created the records in the first place. That document, not general industry practice, decides what can be shared, with whom, and for how long.

Confidentiality, Survival, and Third-Party Access Clauses

Most professional services agreements include a confidentiality clause preventing either party from disclosing project information to outsiders without written consent. These obligations frequently survive well past project completion, sometimes for five, ten, or twenty years, so an archived project from several years ago is not automatically free to use just because the client relationship has gone quiet.

Third-party access provisions add another layer, often requiring that any outside party receiving project data meet specific security standards or be vetted by the client before access is granted. Handing a dataset to an AI company introduces exactly this kind of third-party processor, which means the original contract’s access rules apply in full.

Canadian professional engineering guidance reinforces this point directly: engineers should not reveal facts, data, or information obtained professionally without the owner’s prior consent, aside from a narrow public-safety exception. The same guidance warns against using a former client’s privileged or proprietary information without consent, which applies just as much to a licensing deal with a technology buyer as it does to a new design project.

Project-Use Licences Rarely Cover AI Training

A client’s licence to use project documents for construction, operation, or maintenance is defined by purpose, and that purpose rarely stretches to cover machine learning. Under agreements such as B101-2017, the owner’s permitted use is described specifically in relation to the project, not as a blanket right to repurpose instruments of service for unrelated secondary uses. If a contract does not list research, product development, or database inclusion as permitted uses, training an AI model on that material sits outside the agreed scope entirely. Silence on AI-specific use should be read as a gap requiring fresh permission, not as an opening to proceed.

Whose Data Is It Really?

A project archive is rarely the product of one party alone, which makes ownership and permission questions more layered than they first appear.

Clients, Sub-Consultants, and Mixed Rights

Project files often blend contributions from the client, the lead engineering firm, and various sub-consultants, each of whom may hold a different claim over their portion of the dataset. A structural engineer’s calculations, an architect’s floor plans, and a specialist mechanical consultant’s schedules might all sit in the same project folder, yet none of them share identical rights or restrictions.

In an Australian engineering consultancy, for instance, copyright in work created by an employee during their employment generally belongs to the employer, but that ownership does not by itself establish permission to disclose client information. Before any licensing conversation goes further, it helps to treat an archive as a collection of distinct rights holders rather than a single undifferentiated dataset, keeping a record of which licence conditions attach to each source.

Why Scrubbing Files Isn’t Enough

Removing a client’s name from a title block feels like a reasonable first step toward anonymising a dataset, but it rarely goes far enough to satisfy confidentiality obligations or protect sensitive information.

Hidden Identifiers in CAD and BIM Files

Modern CAD and BIM files carry far more identifying detail than the visible drawing. Revision histories can contain engineer usernames, timestamps, and internal notes about client-requested changes, while hidden layers sometimes include original site survey data with precise geospatial coordinates. Distinctive equipment layouts, custom-coded design blocks, or a building’s unique geometric footprint can all func

tion as fingerprints, pointing back to a specific client or project even once obvious l

abels are gone. For sensitive sites, this kind of residual detail is also a physical security risk, since removing a name never addresses what the geometry itself reveals.

Reformatting Doesn’t Erase Original Restrictions

Converting a detailed BIM model into a simplified mesh, or exporting a drawing into a different file format, might feel like it resets the data’s legal status, but the obligations attached to the original source generally travel with any derivative version. If a contract restricts commercial use of the underlying instrument of service, a simplified or reformatted copy typically remains just as restricted, because the restriction is tied to the information itself rather than its container.

The same logic extends to synthetic data generated from a protected dataset: if the generative process draws exclusively on a client’s proprietary files, the resulting output may still carry an echo of those original restrictions. Reformatting changes the file, not the legal obligations attached to it.

Matching Permissions to the AI Use Case

Not every AI buyer wants the same thing from an archive, and the specific use case should shape exactly what permission is sought and granted.

Training, Evaluation, RAG, and Synthetic Data Differ

AI companies generally fall into a few different patterns of data use, and each one carries its own risk profile:

  • Training absorbs patterns directly into a model’s weights, making the data’s influence effectively permanent within that system.
  • Evaluation or benchmarking tests a model’s performance against the data without necessarily baking it into the model itself.
  • Retrieval-augmented generation (RAG) queries an archive for specific references at the point of use, rather than training on it wholesale.
  • Synthetic data generation uses existing records as a basis for creating new, simulated datasets that mimic real-world patterns without directly exposing the originals.

A sound approach documents data sources, collection methods, applicable licences, and preprocessing steps for each of these pathways, along with clear rules on who can access the data and under what conditions it can be reused. Treating all four uses as interchangeable risks granting permissions far broader than the original contract or the client relationship can support.

Sensitive Sectors May Carry Extra Restrictions

Government, infrastructure, and defence-related projects often carry restrictions well beyond a standard confidentiality clause. Data produced under certain government contracts can be subject to specific government rights rules, meaning a contractor should check the governing terms of that particular contract before assuming the records can be licensed elsewhere.

Technical-data rights in this space often distinguish between unlimited rights, government-purpose rights, and limited rights, depending partly on how the original work was funded. For engineering firms whose archives include public infrastructure, utilities, or security-adjacent facilities, this is a sector where a routine licensing conversation can quickly surface jurisdiction-specific or security-driven limits that a purely commercial project would never raise.

License Client Engineering Data

Building a Clear Licensing Agreement

Once the rights picture is clear, a well-drafted licensing agreement becomes the tool that keeps a deal within the boundaries both parties actually intend.

Defining Included Records and File Versions

A strong agreement names exactly which records are covered, right down to the file version. A firm might, for example, limit a licence to CAD files from projects completed before a specific year, excluding more recent or more sensitive work. Specifying file types, project date ranges, and revision status prevents an AI buyer from assuming that access to a server folder implies access to everything inside it.

Setting Rules on Reuse, Sharing, and Derived Data

Beyond defining what is included, the agreement needs to spell out what the buyer can actually do with it. Useful terms to pin down include:

  • Which specific models, systems, or affiliated entities are permitted to use the data
  • Whether the permission covers training, testing, product development, or commercial deployment, since these are treated as distinct uses
  • Rules on sublicensing or onward sharing with the buyer’s own partners or successors
  • How rights in derived outputs, such as embeddings or model improvements, are allocated between the parties
  • Retention, deletion, security, audit, and incident-response obligations once the data has been shared

An agreement that leaves these points vague invites the kind of scope creep that turns a narrow, controlled licence into an open-ended transfer of value the firm never intended to give away.

Rights Checks Must Come Before Any Licence

Licensing a project archive to an AI buyer can be a legitimate way to put dormant records to new use, once the rights picture is genuinely clear. That means reading the original contract closely, accounting for every contributor whose work sits inside the archive, treating file scrubbing as a starting point rather than a finish line, and matching permissions precisely to the buyer’s intended use.

Archive Trove works as a neutral introducer in this space, helping firms map records, spot readiness gaps, and prepare clearer submissions before any sensitive file changes hands, while leaving the decision to proceed entirely with the business.

For engineering firms weighing up whether an archive is ready for this kind of conversation, a sensible first move is running a low-disclosure fit check for AI data licensing before sharing any sensitive files.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top