AI for Engineering Knowledge Management
What a no-training AI guarantee must cover for CAD, drawings, indexes, and logs, plus five checks to run before an AI reads your PDM vault.
·
⏱
8 min read
Dr. Maor Farid
Maor Farid is the Co-Founder and CEO of Leo AI, the first AI platform purpose-built for mechanical engineers. He holds a PhD in Mechanical Engineering and completed postdoctoral research at MIT as a Fulbright fellow. A Forbes 30 Under 30 honoree and former AI researcher and Mechanical Engineer in an elite military intelligence, Maor leads Leo AI's mission to transform how engineering teams design better products faster.
BOTTOM LINE
A no-training guarantee is only as good as its coverage. Check that it names content, metadata, interactions, derived data, and every form of model improvement. Ask who else can see your files, what happens to the index on exit, and whether the promise lives in a signed document rather than a web page. If the vendor can answer each of those in writing, you can let an AI read your drawings with confidence. If not, keep the vault closed until they can.
Every AI vendor selling into engineering says some version of the same sentence: we do not train on your data. It is a good sentence. It is also doing far more work than it can carry, because a mechanical engineer's data is not one thing. It is geometry, drawings, revision history, supplier quotes, tolerance decisions, review comments, and the prompts typed against all of it.
A guarantee that covers only one of those layers leaves the others open. This post walks through what a no-training promise has to cover before an AI is allowed near your vault, where the gaps usually hide, and how to test the answer before you sign anything.
Why "we do not train on your data" is not one claim
Training is the headline risk, but it is only one of several ways data can outlive the moment it was used. When a vendor says they do not train on customer data, the useful follow-up is: which data, which kind of training, and what happens to everything else?
Consider a single released part. The native CAD file holds the geometry. The drawing holds the tolerances and notes. The PDM record holds who approved it and why the revision changed. An AI assistant that answers questions about that part may touch all three, plus the question you asked and the answer it returned. Each of those is a separate surface with a separate retention story.
Training also has more than one meaning. Pretraining a base model on customer files is the case everyone pictures. Fine-tuning a model for one customer, adjusting ranking behavior from usage, and using conversations to evaluate quality are quieter versions. A guarantee that names only pretraining can be literally true while the quieter versions continue.
This matters more in mechanical engineering than in most fields. Your drawings encode manufacturing know-how, supplier relationships, and design intent that took years to build. Nobody wants a competitor's assistant to produce a suspiciously familiar bracket because your geometry once passed through a shared model.
There is also a regulatory angle. Aerospace, defense, and medical teams often work under export, customer, or quality-system rules that restrict where technical data may flow. For them, a vague promise is not a procurement-grade answer. They need to point to a clause, a certification, and an architecture diagram, and be able to hand all three to an auditor.
IN PRACTICE
It surfaces the relevant internal material, previous design decisions, past calculations, and backs everything with a cited source I can actually click on and verify.
- Yuval F., Clalit
The five things a guarantee has to cover
When we review a vendor's data language, we look for explicit coverage of five areas. If any one is missing or vague, treat the guarantee as partial.
Content. Native CAD files, drawings, BOMs, markups, and attached documents, not just "documents."
Metadata. Part numbers, revision history, approval records, and folder structure. These reveal as much as the geometry does.
Interactions. The prompts your team writes and the answers returned, including thumbs-up signals and feedback text.
Derived data. Embeddings, search indexes, caches, and summaries built from your content.
Purpose. Every form of model improvement, including fine-tuning, evaluation sets, and ranking adjustments, not only pretraining.
A strong guarantee names all five, in a contract or a security document you can attach to a purchase order. A marketing page alone is not enough, because a marketing page can change on Tuesday without anyone signing off.
It also helps to separate what the vendor promises from what the architecture enforces. A promise can be broken by accident, by a new product team, or by an acquisition. An architecture where customer data is never routed into a training pipeline in the first place holds up better, and you can ask to see how it is built.
Derived data is where promises quietly leak
The most common gap sits in the fourth item above. To answer questions about your vault quickly, any AI system builds some kind of index. It might store text chunks, vector embeddings of drawings, or a graph of how parts relate to assemblies. We covered what such an index holds in what AI stores when it indexes your PDM vault, and the short version is that the index is a transformed copy of your data, not a harmless fingerprint.
Embeddings can be partially inverted. Summaries restate your content in new words. Caches keep recent answers. If a guarantee says your files are not used for training but is silent about the index, ask three questions. Is the index built per customer, or pooled across customers? Is it deleted when you leave? Can it ever influence answers shown to another customer?
The right answers are per customer, deleted on exit, and never. Anything softer deserves a written clarification. This is the same discipline you already apply to a supplier who keeps your tooling drawings after a contract ends: you want to know what they hold, where, and for how long.
Who else sits between your drawings and the model
Most AI products call a model hosted by someone else, which means your guarantee is only as strong as the weakest link in that chain. Ask for the full list of sub-processors that can see your content, and ask what each one is contractually barred from doing with it.
There are three links worth checking. The first is the model provider: do its terms prohibit training on the traffic your vendor sends it, and is that a default or an opt-out? The second is logging: do prompts and responses get stored for abuse monitoring, for how long, and who can read them? The third is human review: can an employee read your conversations to improve quality, and under what controls?
Security certifications help here because they turn promises into audited controls. We explain what an independent audit does and does not prove in our piece on SOC-2 for engineering AI. Leo is SOC-2 certified and GDPR compliant, no AI is trained on customer data, and customer intellectual property is protected. Leo connects to an organization's PDM, PLM, and file shares as an intelligence layer on top of them, with offered integrations for leading platforms such as SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM. Your vault stays the system of record.
For a broader view of why the source of a model's knowledge matters at all, see why AI training data matters for engineering. The short form is that a model trained on vetted standards and books does not need your drawings to be useful, which is the cleanest way to make a no-training promise easy to keep.
How to test a guarantee before you sign
Reading the language is step one. Step two is making the vendor demonstrate it. These checks take an afternoon and tend to separate durable guarantees from decorative ones.
Ask for the data processing terms and find the exact sentence about model improvement. Check it names fine-tuning and evaluation, not only training.
Request the sub-processor list and confirm each entry has no right to retain or learn from your content.
Ask what happens on deletion: the index, the caches, the logs, and the backups. Get a timeline in days.
Run a canary test. Put a distinctive, harmless part name into a pilot vault, and later check that it never surfaces for a different account.
Confirm that answers carry citations back to your own sources, so you can see where each claim came from.
Permissions deserve their own review, since a model that respects the guarantee can still show the wrong person the wrong file. We look at that side in how AI connects to PDM and PLM, and the knowledge-structure angle in engineering knowledge graphs for product data. Together they show why the guarantee, the index, and the access model have to be reviewed as one system rather than three separate checkboxes.
One last habit worth building is a yearly re-check. Vendors ship new features, add model providers, and update terms. A guarantee that was complete at signing can develop gaps when a new summarization feature or integration quietly adds a new place for your data to land. Put a recurring review on the calendar and ask for notice, in writing, before any sub-processor or data flow changes.
FAQ
See Leo on your own vault
Book a demo and ask us the hard data questions.
Leo is SOC-2 certified, GDPR compliant, and trains no AI on customer data. See how it works on top of your PDM and PLM, with every answer citing its source.
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
