Engineering data for AI training: why CAD files and technical drawings are in demand

Many firms keep completed projects as folders on their servers: installation plans, manufacturing drawings, assemblies and plan revisions from many years. These are exactly the documents AI labs are interested in, because every file contains dimensions, tolerances, standards and spatial relationships. This article describes who is looking for such data, what makes it valuable and which rights need to be cleared first.

Why do AI labs look for engineering data?

Because AI models are increasingly expected to handle technical tasks, and there is hardly any professional training data for them.

Models that are meant to generate a CAD model from a description or read a technical drawing have to understand geometry, structure and function. Research has so far worked mainly with freely available models. A well-known example is the ABC dataset with around one million CAD models from the Onshape platform.^[1] Autodesk trained its experimental 3D model Bernini on ten million 3D shapes from publicly available data and writes that the models will become much more useful when trained on larger, higher-quality professional datasets.^[2]

Professional projects differ from such collections: they consist of many related files, go through several plan revisions and are bound by real requirements such as standards, building codes and manufacturing limits. Such datasets sit almost exclusively with the companies that created them.

Who buys such data?

So far only a few programs, and prices are not public. But the demand is becoming visible.

The Who Sells Data directory lists around 30 programs that pay companies for data. Most of them look for communications, ticket and business data. The only one there that names technical drawings and CAD models explicitly is The General Data Company, which buys industrial operating records for a one-off price or a revenue share; Opendatabay lists 3D spatial data as a data type of its own.^[3] A summary of a podcast with Ryan Lock, co-founder of the data broker Polyshares, names PCB CAD, architecture, heavy construction and game 3D as current requests from labs.^[4]

Shutterstock shows that 3D data is licensed for payment: together with NVIDIA, the company trains generative 3D models on its own 3D library and compensates the creators through a fund.^[5] For engineering data from companies, no prices have been published yet. How we value a dataset is described in the article What is company data worth?

What makes engineering data valuable?

Native formats, complete projects, context and long time spans.

Datasets in which a project can be followed across several systems are particularly sought after: from the quote through planning to acceptance, with the related emails and tickets. A model then sees not only the result but also the decisions along the way.

Which rights need to be cleared?

Who created the documents and for whom. With engineering data, the rights often do not lie with the company alone.

• Under German law, technical drawings and plans can be protected as depictions of a technical nature (§ 2(1) no. 7 UrhG).^[6] The Federal Court of Justice considers a low degree of individual character sufficient, with a correspondingly narrow scope of protection.^[7] Only the form of the depiction is protected, not its technical content.^[8]
• Architects' plans can additionally be protected as designs for works of architecture (§ 2(1) no. 4 UrhG). The copyright stays with the planner, and the client receives rights of use only to the extent the purpose of the contract requires (§ 31(5) UrhG).^[9] A licence for AI training is unlikely to be covered without an express agreement.
• The company's own designs, especially for current products, can be trade secrets (§ 2 no. 1 GeschGehG).^[10] It therefore makes sense to license completed projects and discontinued product lines.
• Drawings supplied by clients and projects under a non-disclosure agreement need consent or are excluded.

How we clean title blocks, metadata and site plans is described on the page Selling engineering data.

How do you prepare a dataset?

Take stock, clear the rights and keep the originals unchanged.

1. Take stock: which formats, how many projects, how far back and on which servers.
2. Clear the rights: separate the company's own designs, client drawings, architects' plans and confidentiality agreements.
3. Keep native formats: do not convert to PDF and do not merge plan revisions.
4. Do not delete anything rashly: back up project folders before clearing servers.