Your Designs, Your Models, Your Infrastructure

The question to ask before you upload a single CAD file

Before you let an AI agent work on a board, a part, or an assembly, ask one question: after this tool has seen my design, who gets better because of it?

Every hardware team understands the obvious risk: a proprietary schematic leaves the building and sits on a vendor server. That risk is real, and it is not the whole risk. The quieter one is written into the terms. In the worst clauses we found, your files, your prompts, and the outputs the tool generates become raw material for the vendor: to train its models and improve its platform. The improved platform is then sold to other users, potentially including the competitor who will quote against you next quarter.

AI is going to keep getting better at engineering work. We build ProductFlo on that bet, and we want agents doing real work inside real tools. The question is not whether to use them. The question is whose asset your design history becomes while you do. With too many AI tools on the market, your design history can improve the vendor's model for everyone. With ProductFlo, you can keep that data in your environment and use it to improve your next product. That difference is architectural, not rhetorical.

What we found in the field, in brief

This is the second article from our terms audit. In Article 1, we read the live Terms, Privacy Policies, and AI terms of 58 commercial AI tools for hardware teams and scored what we found: 25 were classed critical for proprietary design data under the audit's worst-tier rule, 29 were caution, and only 4 were clear. Three findings matter for what follows.

First, training rights hide in licences, not in sentences about training. EasyEDA treats an upload as, in its words, a "non-exclusive, perpetual and irrevocable grant" of all intellectual property rights in the content. CircuitMind, an enterprise product sold on order forms, takes a perpetual and irrevocable licence over both uploaded files and the schematics and BOMs it generates, to improve its algorithms, expressly carved out of the duty to erase that data at termination. CloudNC's standard terms take a perpetual licence to use customer data and results to improve the platform "for other users," which contradicts its own product FAQ.

Second, the dangerous setting is often the default, and the dangerous tier is often the free one. MecAgent's free tier grants a perpetual licence over CAD parts, drawings, and assemblies, including providing the data to third parties for any purpose, and deleting your data does not revoke it. ProtoFlow turns its improvement setting on by default for consumer accounts and keeps the resulting records with no automatic expiration. Vizcom's free Starter tier goes further than training: the vendor retains title to designs created under a Starter account.

Third, some tools publish nothing enforceable at all. DraftAid and Dimension AI have no findable Terms or Privacy Policy. Skyfall has footer links labelled Privacy Policy and Terms of Service that are dead placeholders and go nowhere. Nothing to read before you upload, nothing enforceable about training, ownership, or retention. And where a reassuring no-training promise does exist, it often lives in a blog post, a Q and A document, or a product guide rather than in the binding terms. That is the pattern at PTC, Oracle, and Cadence. A sentence that is not in the Terms, the AI addendum, or a signed DPA is not a commitment. It is marketing.

We also found vendors who get this right. Article 1's clear list names CoLab, Trace Space, and Bananaz, plus McNeel for CAD-file confidentiality and Embedder on paid terms. It also credits Ansys, which remains caution overall for exit and export-control terms, because its no-use clause sits in the SimAI contract itself. The bar exists. Most of the field does not meet it.

The ProductFlo model, point by point

Deployment: the system runs where your data already lives

Data ownership starts with geography. If your files have to leave your network before any agent can read them, every protection after that point is a promise about someone else's building. Fusion, CircuitMind, and ToolPath are cloud-only in the documents we read.

ProductFlo's deployment options are stated on our site in these terms:

"Cloud. Single-tenant. US, EU, or APAC region. Customer-managed KMS."

"Self-hosted. Helm chart. Air-gapped mode. Your Kubernetes, your network."

For enterprise customers, the FAQ states that ProductFlo can be deployed "on their own infrastructure (AWS, Azure, GCP, or on-premise)." And in self-host mode: "Your artefacts never leave your environment in self-host mode."

In our founder's terms, the same point goes one step further: everything is local, or on a VPS you own, in your cloud environment, locked down. We bring the system to the data. For controlled work, our site is blunt: ITAR and EAR data is self-host only. Controlled data is a deployment question, not a feature tier.

Ownership: your data, full stop

Our FAQ states the ownership and training rule directly:

"Your engineering data belongs to you. We never use customer data to train AI models."

The homepage says the same thing from the product side: "Your IP stays yours. Every change leaves a trail." And on exit and portability: "You can export all your data at any time in standard formats."

Contrast that with what retaining title means elsewhere. CircuitMind's customers keep title, and still grant the perpetual improvement licence described above, with a download window of up to five days after termination. Dassault Systemes states that customer data remains the customer's property, then qualifies AI output ownership as subject to pre-existing vendor and third-party rights, and may destroy customer data on termination. Ansys hands the customer a trained model that lives inside its platform: the right to use it ends with the agreement, and the model may be deleted within 30 days of termination.

Our position, stated plainly by our founder, is that ProductFlo owns none of the customer's data. Not the inputs, not the outputs, not the agent history. Ownership that ends at termination, or that survives only inside the vendor's platform, is a rental. We are not interested in renting your own designs back to you.

Training: our models learn from public knowledge, not from your board

The site states the rule twice, so there is no room to misread it:

"Customer data is never used for model training."

"Our AI models are trained on public engineering knowledge and synthetic datasets. Your proprietary designs stay private."

Your board does not become a training puzzle. On the site's stated rule, there is no improvement use to hunt down and toggle off. Compare the field, where the audit found four separate mechanisms: express training rights, perpetual improvement licences that never use the word train, aggregated and anonymized carve-outs sitting next to a no-training promise, and silence. A buyer who searches the terms for the word "train" will miss at least two of those four.

Models: ours built in, or yours by key

The site states the model choice directly: "Bring your own models, or use ours." Frontier models handle the agents that benefit from long context. Smaller, local models handle the agents that run on every commit. In a hybrid deployment, agents reach out only to the models you approve.

Our founder's position adds two further options. ProductFlo includes a built-in LLM that can do the work out of the box. And if you would rather use your own API key, you can bring it: your key, your model account, your billing relationship, your data terms with that provider. Both points are our position as a company. The controls around any model, ours or yours, are the part already described on the site, and they matter more than the model itself: agents only propose changes and never commit them, dry-run mode is on by default, scoped tokens limit an agent's reach to one product or one revision, and, in the site's words, "Every agent call is audit-logged with the prompt, the model, and the diff produced."

Set that against the field. Oracle's AI terms make the customer solely responsible for what its agentic features do. HeyPCB discloses that a small number of its staff can open complete projects, including chats, without notifying the user. Ours leaves a trail by default.

Your data trains your models, not ours

The asymmetry repeats in several of the 58 tools: a vendor reserves a right to learn from customer work while its terms deny the customer the mirror-image right. Ansys, Siemens, Dassault Systemes, Trace Space, Werk24, and CloudNC all carry verified bans on using outputs to train or improve an AI model. That restriction is sharpest where the same vendor also reserves improvement rights over customer inputs, as Siemens and CloudNC do in clauses covered in Article 1: the vendor can use your work to improve its platform, while your use of its outputs to train a model of your own is barred.

ProductFlo inverts it. Everything we collect, process, and orchestrate for a customer is that customer's data, and our founder's position is that the customer can use it down the road for their own purposes: to train a better model of their own, to tighten design review, to raise quality, and to make the next generation or the next variation of the product better than the last. Your design history is a compounding asset. Reviews your senior engineers approved, changes that survived DFM, the respin and the fix that closed it: kept in your environment, that record is the material a team needs to get faster over years. On ProductFlo, that value accrues to the team that created it.

The honest gap on our own side

We held ourselves to the same standard we applied to everyone else in this audit, and here is what we found when we started it. Our data promises lived in FAQ and homepage copy, and our legal pages were generic website templates. That is the exact weakness this audit criticizes at PTC, Oracle, and Cadence, and we are not going to pretend otherwise. So we rewrote them and published them alongside this series: our Privacy Policy, Terms of Service, and Security pages now carry the commitments in this article, written where counsel can read them. Hold us to them.

The checklist: questions to ask any AI tool, including us

The middle column compresses the patterns in Article 1. It is not a claim about every vendor.

Question	Typical cloud AI tool	ProductFlo
Where does the software run?	The vendor's cloud. Your files travel to it.	Your Kubernetes, your network, including air-gapped mode, or single-tenant cloud with customer-managed keys.
Who owns the inputs and outputs?	You keep title, often beside a perpetual licence back to the vendor.	Your engineering data belongs to you. We own none of it.
Is your data used to train the vendor's models?	It can be: by licence, by default-on toggle, or on the free tier.	Never. Our models train on public engineering knowledge and synthetic datasets.
Can you use your own data to train your own models?	Banned for outputs in several sets of terms.	Yes. That is the point. Your processed data is yours to build on.
Which models process your work?	The vendor's choice, sometimes unnamed providers.	Ours built in, or yours by model or API key. Hybrid agents reach only approved models.
What can an agent change on its own?	Varies. Oracle's AI terms, for one, make the customer solely responsible for agentic AI actions.	Nothing by default. Agents propose, dry-run is default, tokens are scoped, every call is logged.
What happens when you leave?	It can be a retrieval window measured in days, deletion on the vendor's clock, and licences that survive.	Export anytime in standard formats. In self-host mode, artefacts never left your environment.

The ProductFlo column combines site-stated deployment, ownership, training, export, and agent controls with the founder-stated ownership and model positions identified above. It is not a claim that the founder-stated BYOK or built-in LLM wording appears on the website or in a contract.

Ask these seven questions of every tool on your shortlist, and read the answers in the terms, not the brochure.

Your designs are the most valuable training data in your company. Make sure they are training your company.

The tools scored in this audit, and the rest of the field, are listed in our AI tools directory.