← Back to blog
October 31, 2026·17 min read·AI Tools

Before You Let an AI Agent Touch Your CAD Files, Read Its Terms

We read the live terms, privacy policies, and AI addendums of 58 commercial AI tools for hardware teams. Twenty-five were critical for proprietary design data, twenty-nine caution, four clear. The worst clauses hide in licences, not in sentences about training.

Serge Kadjo
Serge Kadjo
AI toolsCADterms of servicedata ownershipengineering
Before You Let an AI Agent Touch Your CAD Files, Read Its Terms

Before You Let an AI Agent Touch Your CAD Files, Read Its Terms

Your CAD file is not an input. It is the product: the geometry, the tolerances, the supplier choices, the design decisions your company sells. Upload it to an AI tool and you are not just buying a subscription. Depending on the terms you accepted, you may also hand the vendor a permanent right to learn from the design itself. You pay for the seat, then you pay again in data.

We read the public legal documents behind 58 commercial AI tools sold to hardware teams. Twenty-five were classed critical for proprietary design data under the audit's worst-tier rule, twenty-nine more landed in the caution band, and four were clear. The worst clauses were rarely labelled as training. They were licences, written to survive deletion and termination.

How we ran this audit

On October 31, 2026, we read the live Terms of Service, privacy policies, AI addendums, and order form terms of 58 commercial tools across CAD, ECAD, PLM, and manufacturing AI, scored on training use, ownership, retention, subprocessors, deployment, and restrictions, plus a licence audit of the open source tools in our own AI tools directory. Each vendor is counted once, under the most severe classification applying to any widely available tier, default setting, or public surface, because a team can land on that tier before procurement gets involved. A follow-up live-browser check found legal pages for four vendors the first pass missed; their corrected findings are incorporated below, not carried as absences. Every quote below is verbatim from the vendor's own legal page, linked at the quote. Where we could not read a document, we say so. A claim we could not verify on a live page does not appear here as fact.

The critical list: eleven clauses to avoid or negotiate first

Ranked by how hard the clause is to undo, how much it takes, and how easily a working engineer lands on it by default.

1. MecAgent, free tier: your CAD, licensed to third parties, forever

The free tier grants MecAgent a licence to "distribute, make available, transmit, or otherwise provide Customer Data ... to third parties for any purpose, including commercial purposes." Customer Data here includes CAD parts, drawings, and assembly files. Deleting your data does not revoke the licence, and does not require MecAgent to delete or retrain models built from it.

Source: MecAgent Terms, §7.1 and §7.5

2. EasyEDA: uploading is a grant of all IP rights in the content

EasyEDA states that "your publication and uploading of such information is deemed to be a non-exclusive, perpetual and irrevocable grant to us of all legal rights and interests in relation to such information, including all intellectual property rights." No opt-out appears in the text read. The data controller is a Hong Kong company in the JLC group, which shares data with affiliated manufacturing and component businesses, JLCPCB and LCSC. A proprietary board uploaded here is not just stored. It is licensed.

Source: EasyEDA Legal, §VII

3. CircuitMind AI: a perpetual improvement licence over inputs and outputs

The customer keeps title, but grants "a non-exclusive, worldwide, royalty-free, sublicensable, transferable, perpetual, irrevocable licence to use the Input Data and Output Data ... to monitor, assess, analyse and improve the performance of the Circuit Mind Technology (including any algorithm therein)." Inputs are your design files, outputs are the generated schematics and BOMs, and that data is excluded from the erasure duty at termination. Exit download window: up to five days.

Source: CircuitMind Terms, §10.4 and §17.2

4. Altium: uploads to an Altium Website are deemed not confidential

Altium's Website Terms state that "any of Your Data that You send or upload to an Altium Website will be deemed NOT to be confidential or proprietary", backed by a perpetual, irrevocable licence in the section before it. Mitigations exist: the Master Services Agreement takes precedence in a conflict and bars disclosing proprietary models, designs, and schematics to third parties. They protect only customers with an MSA override in writing.

Source: Altium Terms, §8 and §9

5. CloudNC: the standard terms contradict the FAQ

CloudNC's FAQ for CAM Assist says it does not train its AI on customer data. Its Standard Terms grant a "perpetual licence to use, copy, adapt, prepare derivative works from, and translate the Customer Data and the Results for the purpose of operating and improving the Platform and/or the Services for other users and/or the underlying technologies." Customer Data includes parts and tooling data, and Results are the generated toolpaths and G-code. A separate Regulated Environments version is much narrower. Which one you sign is not determinable publicly.

Source: CloudNC Terms, Standard §1.6

6. Duro Labs: a perpetual licence over PLM data, under a $100 liability floor

Duro's Netlify-hosted Terms copy grants "a perpetual, irrevocable, non-terminable, transferrable, worldwide, royalty-free, sublicensable, fully paid-up, non-exclusive and transferable license to (i) use, reproduce, display, and publicly perform Submitted Information." A protective ownership clause in the same document sits in tension with it. That copy is dated February 20, 2020 and says nothing about AI. The live durolabs.co Terms of Use is dated May 4, 2021, but the verification read could not render Section 3.5 there; the live Privacy Policy is dated June 9, 2026. The Netlify copy caps total liability at the greater of fees paid or $100.

Source: Duro Terms of Use, §3.5(a)

7. Vizcom: on the free tier, the vendor retains title to your designs

Vizcom's free tier takes the designs themselves: "Vizcom retains all right, title, and interest in Designs created under a Starter Account." Starter and Education content may also be used to train. Paid Premium accounts reverse both points. The hazard is a designer testing it on a real concept with a free account.

Source: Vizcom Terms, §9 and §10

8. siliXon: free projects are public by default, under a licence that cannot be taken back

siliXon's terms state: "On the Free plan, new projects are public: a board is published to Community once it passes every manufacturing check." Published boards are licensed to other users under the CERN Open Hardware Licence, permissive version, and copies made before you switch a board to private stay theirs, because the licence cannot be revoked.

Source: siliXon Terms, §4

9. ProtoFlow: training is on by default, and the records do not expire

For consumer accounts, ProtoFlow's improvement setting is on by default, and while it is on "you grant ProtoFlow LLC a worldwide, non-exclusive, royalty-free license to use, reproduce, and create derivative works from the designs, prompts, and other content and activity you generate in the Software ...". Before and after snapshots of the active design upload automatically. Those records have no automatic expiration, and opting out does not delete records already received.

Source: ProtoFlow Terms, §3 and Privacy Policy, §6

10. Siemens: the version of the AI terms in your order decides

Siemens AI Terms version 1.1, dated May 2025, state that "Siemens and its affiliates and their subcontractors may collect, store, and analyze conversational experiences with the AI Services (input / and Output) to train and improve its products or services ...". Versions 1.2 and 1.3 remove that sentence and narrow use to aggregated systems information. Which version governs depends on your order. Pin it in writing.

Source: Siemens AI Terms of Use v1.1

11. Foundation EGI: AI input and output are reserved for improvement

Foundation EGI's Terms cover the SaaS product and define User Contributions to include CAD files. Its AI Features section states: "We reserve the right to utilize your Input and Output to improve upon our Website and all related services." It also states: "You acknowledge that your Input and Output may be shared with GenAI Providers in order to deliver you the AI Features." Those providers are not named. The risk allocation is explicit: "You submit any confidential information, personal information, or otherwise sensitive information in your use of AI Features at your own risk." No opt-out from Foundation's own improvement reservation was found in the Terms read. The mitigation is narrower and should be stated fairly: its Privacy Policy says "we do not use Customer Content submitted through paid SaaS accounts to train or improve third-party foundation models except as authorized by the customer in writing (for example, via an enterprise agreement or data processing addendum)". That protects paid SaaS content from third-party foundation model training. It does not withdraw Foundation's own improvement right.

Source: Foundation EGI Terms of Use and Privacy Policy

Also critical: Haitch turns quality retention and model training on by default, with dataset data retained indefinitely until revoked. Leo AI may use all Content to train and to market the service, no opt-out found. ToolPath takes a transferable, sublicensable licence over content including drawings and models, liability capped at $500. Quilter's free tier uses board metadata and input files to generate training puzzles; Quilter had no findable standalone product Terms or Privacy Policy behind that page.

Pay attention: the patterns behind the clauses

The promise lives in marketing, not in the contract. PTC's line, "We never use customer data to train or fine-tune AI models," is verified in a blog post; we could not verify its binding status in the MSA or DPA. Oracle's no-train commitment sits in a Q&A, while its binding AI Terms contain no equivalent sentence and allow third party AI to process your content. Cadence's promise exists in a product guide, while the binding SaaS Addendum licenses usage data collection for improvement. DeepPCB, Flux, Flow, and Propel repeat it. If the sentence is not in the Terms, the AI Addendum, or a signed DPA, it is not a commitment.

Deletion does not end a licence. Retention clauses govern copies; licence clauses govern rights, and the rights outlive the copies. Haitch retains dataset data indefinitely until revoked. Werk24's opt-out is prospective only: what was kept stays kept. Drafter keeps a post-term licence to train on aggregated customer materials, with no opt-out found. Ansys SimAI may delete content and trained models within 30 days of termination, and the model right ends with the agreement.

Check the output ban in both directions. Ansys, Siemens, and Dassault Systemes bar using outputs or generated content to train AI systems or competing software. Werk24 and CloudNC bar using output or Results as training data. That ban does not always pair with a vendor-side training right: Ansys, for example, bars its own product-improvement use in the SimAI contract. Where the two do pair, as in CloudNC's Standard Terms, the asymmetry is direct: your data can improve their platform, while their Results cannot be used to train yours.

Three shorter patterns. Benchmark and competitive-testing restrictions recur across the documents read for PTC, Cadence, Altium, Ansys, Jama, Zoo, Trace Space, Werk24, and CloudNC; for Dassault Systemes, the ban was verified in its evaluation agreement, while the commercial EULA equivalent was not fully verified. Staff access is real: HeyPCB discloses that a small number of staff can open complete projects, including chats, without notifying the user, and MecAgent allows tasks to be corrected by contractors, possibly outside your country. Export controlled data is excluded or your problem: Ansys bans ITAR and EAR data outright, Autodesk prohibits ITAR content, PTC excludes it from SaaS by default, and Altium makes classification solely the customer's responsibility.

A protective ownership clause can still leave AI use unanswered. Bild and Tandem are caution cases, not missing-terms cases. Bild's Terms state: "The Terms do not grant Bild any ownership of your Content". Its operating licence is limited to what is reasonably necessary to provide the Services. But neither Bild document contains any training, AI, or Meru data-use clause at all, even though Meru works from geometry, dimensions, and metadata. After termination, Section 12(b) gives the customer fourteen (14) days to request Content back. Tandem's Terms state: "You retain ownership of content you provide. You give Tandem permission to process that content only as needed to provide, secure, support, and improve the services, subject to any applicable written agreement." No training sentence appears in Tandem's documents, but "improve the services" is not bounded, subprocessors are not named, and the material product specifics are left to a written agreement. Ownership is answered in both cases. AI use is not.

Source: Bild Terms of Service, Bild Privacy Policy, Tandem Terms and Conditions, and Tandem Privacy Policy

The clear list: vendors that put the protection in the contract

Credit where it is due.

CoLab Software: "Input will not be (a) used to train the AI Features in general or (b) used to train any models provided by the Third Party AI Providers or (c) commingled with other customers' data to provide results to other customers or Third Party AI Providers." In contract in its AI Addendum, alongside customer ownership and return within 30 days of termination. Watch item: audit log personal data is retained indefinitely. AI Addendum

Trace Space: "Trace.Space does not use Customer Data or Customer Personal Data to train, fine-tune, retrain, or improve AI Systems, except where a Customer has expressly opted in to improve their AI in writing via a separate addendum, as set out in the MSA." Opt-in only, revocable, with self-hosted software and bring your own LLM keys. Condition: no ITAR or EAR data in the cloud version. AI Policy

Bananaz: "Company shall not use Customer IPR to train, develop, improve or teach any Model." In its platform terms, with customer IPR ownership and output assigned to the customer on payment. Caution: a perpetual anonymized analytics licence. Platform Terms, Section 5.2

McNeel: "Rhino does not collect geometry or other data in files you create or open." Structural, not contractual: desktop software, internet access can be disabled at install, no cloud account needed. The software does not collect the file. Rhino usage statistics

Embedder (paid): "Accepting these Terms does not itself authorize model training or development." Training needs a separate opt-in, withdrawal ends further use including pending datasets, and paid output is assigned to the customer. Terms of Service, Sections 1 and 5.2

Ansys: "Ansys shall not collect or use Customer Content or the Trained model as part of the product improvement program." The incumbent exception: the bar is in the SimAI contract itself. It stays in caution for its exit terms and its ITAR and EAR ban. SimAI clickwrap, Exhibit B, Section 6(b)

The open source side: the licence is the contract there too

Our directory listed 602 tools. Five hundred were labelled open source (411) or source available (89). We could recover the actual licence family from our own data for 47 of the 500. For 453 of 500, about 91 percent, no licence family is recorded anywhere in our data. That gap is ours, and we are fixing it.

Permissive is safe to build on. MIT and Apache-2.0 cover most of the KiCad MCP and agent ecosystem, build123d, eda-agent, PlatformIO, and ESP-IDF: closed commercial products are fine, keeping the notices.

Copyleft is safe to drive, not to absorb. FreeCAD and OCCT sit under LGPL-2.1, KiCad, Arduino CLI, KLayout, and FreeRouting under GPL-3.0. Running them, including over a process boundary, and using their outputs is unrestricted. Linking their code into a closed core is not.

AGPL is the network trap. Konnect, FreeGAD, and KiCad Agentic MCP are named in our directory notes under AGPL-3.0. Offer a modified AGPL component to users over a network and you owe those users its source.

Source available is not open source. The nurb repo sits under FSL-1.1-MIT, which permits internal use and bans Competing Use: offering substantially similar functionality commercially. Evaluation is fine, shipping a competitor is blocked.

No licence file is the most restricted option of all. Default copyright applies, all rights reserved. Our sample found repos labelled source available with no licence detectable, including Seeed Studio's kicad-mcp-server. Visibility is not permission.

What we could not find

Some vendors publish nothing a buyer can read before handing over design data. In the corrected count, 18 of 58 had no public product terms covering CAD or design data. The examples below are the cases the final live-browser verification settled, stated as absence, not as inference:

  • DraftAid: no Terms or Privacy Policy findable, and no legal link in the homepage footer.
  • Dimension AI: no Terms or Privacy Policy findable, and no legal link on the homepage, for a tool that rebuilds CAD from legacy drawings and STEP files.
  • Skyfall (SKY-001): the footer shows Privacy Policy and Terms of Service labels, but both links are dead placeholders that go nowhere.
  • Bench: legal pages exist, but the Terms are website-only by their own Section 1 wording: "Any purchase or use of Bench's SaaS products will be governed by a separate written agreement entered into after a sales discussion." No public term governs Bench's product or engineering data.
  • CADDi: narrowed on recheck. A Privacy Policy and a Security page exist. We could find no product Terms governing CADDi Drawer customer CAD data.
  • Flow Engineering: a Privacy Policy exists, and its Security page states that prompts, documents, and outputs are never used to train any model. We could find no public SaaS Terms, MSA, or DPA behind that statement.

Source: Bench Website Terms and Conditions

If nothing enforceable can be read before you sign or upload, every protection has to come from a private agreement you negotiate yourself, before the first upload.

Before you upload: a checklist

Run this against any AI tool before a proprietary design touches it, including ours.

  1. Which document governs? Name the exact document and version: Terms, AI Addendum, MSA, or DPA. A protection in a blog post or FAQ is not a commitment.
  2. Read the licence grant before the training sentence. Search for "perpetual" and "irrevocable" before "train." The worst clauses here never use that word.
  3. Check your tier. Free, starter, evaluation, and pilot terms are where title transfers, public by default projects, and default-on training live.
  4. Find the training toggle. Is training on by default? Is the opt-out tier limited? Does turning it off delete what was already collected?
  5. Pin the exit terms. What is the retrieval window after termination? Does deletion cover licences and derivative data, or only copies?
  6. Name the subprocessors. Which AI providers receive your prompts and design context, and do they commit not to train? Can staff or contractors open your projects?
  7. Confirm deployment for controlled data. Is ITAR or EAR data banned, excluded by default, or solely your responsibility? Is self-hosted or air-gapped in the contract, or only on a web page?
  8. Check the liability cap. If the cap is $100 or $500, compare it with what one failed board spin would cost your program before you price the tool.

A note on who wrote this

ProductFlo sells in this market, and we are not a neutral party to this audit. That is why every claim above is tied to a quote on the vendor's own legal page, and why the clear list names competitors whose contract language beats the industry norm. Our own position is published separately, as the second article in this series, and should be held to the same checklist.

How to cite and sources

Cite this article as: ProductFlo, "Before You Let an AI Agent Touch Your CAD Files, Read Its Terms," AI tools terms audit, October 31, 2026. The legal pages relied on most: MecAgent Terms; EasyEDA Legal; CircuitMind Terms; Altium Terms; CloudNC Terms; Duro Labs Terms of Use (Netlify hosted copy, where §3.5(a) was read; the live durolabs.co page did not render that section); ProtoFlow Terms; Siemens AI Terms v1.1; CoLab AI Addendum; Trace Space AI Policy; Bananaz Platform Terms; Embedder Terms; Ansys SimAI clickwrap; Foundation EGI Terms of Use; Bild Terms of Service; Tandem Terms and Conditions; Bench Website Terms and Conditions.

Browse the full list behind this audit: AI tools for hardware engineering, every tool in the directory with its source and licence notes.

Stop bad hardware changes before they ship.

30-day pilot on one active product. Fixed $8k. Live on your own BOM in week one.