AI for CAD Tools
See what an AI assistant can answer about a large Onshape assembly after it syncs, and what still needs checking in Onshape itself.
·
⏱
8 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Heavy Onshape assemblies make simple questions slow, because every answer used to require loading the full model. Once the assembly is synced into an assistant, structure, repeat and part location questions can be answered without a seat. Leo's late September 2026 release made large CAD and STEP files, including those synced from Onshape, sync more reliably, and added automatic retries after temporary server problems. Ask for counts and groupings first, confirm anything that feeds a purchase or a design change in Onshape, and pilot the tool on an assembly you already know before trusting it on one you do not.
A heavy assembly changes how an engineer works long before it changes the design. The model takes a while to load, orbiting gets sluggish, and a quick question such as "how many of these brackets are there?" turns into a few minutes of waiting and clicking. Engineers who search for how to cope with Onshape and heavy assemblies are rarely asking about editing. They want an answer from a model they do not want to open.
This guide looks at what becomes possible once that assembly is synced into an AI assistant. It covers what the sync has to get right, which questions about structure, repeated parts and part locations are worth asking, and what you should still confirm in Onshape itself. In late September 2026, Leo's release notes added two fixes that matter here: large CAD and STEP files, including those synced from Onshape, now open and sync more reliably, and Onshape sync retries automatically after temporary server problems.
It is written for mechanical engineers, design leads and reviewers who work with big assemblies and need answers without holding a seat open.
Why heavy assemblies slow down the simple questions
Most questions asked of a large assembly are read-only. A reviewer wants to know which parts make up a subassembly. A purchasing contact wants to know how many identical fasteners appear. A newer engineer wants to know where a particular bracket sits and what it is attached to. None of these need the model to change, yet each one traditionally needs the full model loaded, a seat occupied and an engineer who knows where to look.
Size makes this worse in a predictable way. The more instances an assembly contains, the more the same part appears again and again, and the harder it is to tell one instance from another by eye. A frame with dozens of identical bolted brackets looks the same from most angles. Finding the one that matters means walking the feature tree or the instance list, and that walk scales with the size of the assembly, not with the difficulty of the question.
There is also a people cost. The engineer who knows the assembly best becomes the default answer desk, and every interruption pulls them out of their own design work. Capturing structure in a form that can be queried in plain language moves part of that load away from a person. Our look at what an agent actually receives when it reads an Onshape document explains why structure is available to an assistant even when geometry is not, and that distinction is the foundation for everything below.
IN PRACTICE
We're less dependent on outsourced engineers. We do it all in-house. We get answers in a few minutes instead of a few days.
- Harel Oberman, CEO, Oberman Industrial Designs
What the sync has to get right
An assistant can only answer about an assembly that reached it intact. For heavy models that is the hard part, because large files are exactly the ones that fail on the way in. A timeout, a temporary server problem or an oversized STEP file can end a sync that works fine on a small part.
The Sept 29 to Oct 4, 2026 release notes for Leo describe two changes aimed at this. Large CAD and STEP files, including those synced from Onshape, now open and sync more reliably instead of failing outright. And Onshape sync now retries automatically after temporary server problems, which should reduce failed syncs. Neither change adds a new question you can ask. They remove the reason a heavy assembly never became askable in the first place.
Leo is an AI assistant built for mechanical engineers, trained on more than one million pages of standards, books and articles, and able to connect to an organization's knowledge base across CAD, PDM, PLM, local and network directories and enterprise systems. It sits on top of those systems as an intelligence layer rather than replacing any of them. For a team on Onshape, the practical value is that the synced assembly becomes one more source an engineer can question in everyday language, next to the specifications and past designs that already live in the same place.
One caution applies to any synced copy. It is a snapshot of the model at the time of the sync. If the design changed this morning, ask when the sync last ran before trusting an answer for a decision that depends on the latest revision.
Questions worth asking once the assembly is in
The brief for a good question is simple: it should be answerable from structure, and wrong answers should be easy to spot. Four kinds of question fit that description well for a heavy Onshape assembly.
Structure questions ask what the assembly is made of. How many distinct parts does it contain, what are the main subassemblies, and which parts belong to the drive side versus the frame? Repeat questions ask what appears more than once. Which parts are used in multiples, how many instances of a given bracket are there, and are two similar-looking parts actually the same part or two different ones? Location questions ask where something sits and what it touches. Where is this bracket, which crossmember is it mounted to, and what else is fastened at the same point? Comparison questions ask how this assembly relates to others. Does the same bracket appear in the previous generation of the machine, and could it be reused here?
Used this way, a short list of habits keeps the answers dependable:
Ask for counts and groupings first. A number is easy to check against a quick look at the instance list.
Name the part the way your team names it, and say which subassembly you mean when the name could apply to several parts.
Ask where a part sits in words, then confirm in Onshape when the answer feeds a purchase or a design change.
Treat any answer about two parts that look alike as a prompt to compare their part numbers, not as a conclusion.
Keep a record of questions that were answered well. They become the pilot set for judging the tool.
The repeat questions deserve special attention, because they connect to cost. Duplicate and near-duplicate parts are a quiet source of procurement and inventory overhead, a theme we cover in how nested assemblies and cost rollups behave in a multi-level BOM. Being able to ask "how many of these are there?" without loading the model makes that kind of review far cheaper to start.
Using the answers in a design review
Reviews are where the no-seat workflow pays off most. A review usually involves several people, and only one of them typically has the CAD session open. The others wait, ask for views, and lose the thread of what was asked. If part of the group can ask the assistant about structure and part locations in parallel, the person driving the model can stay on the question that really needs geometry.
The same release also added something relevant to review work: the measure tool can show a single part's properties, or measure the distance between two parts. That makes a clearance question something an engineer can often settle in the middle of a discussion, rather than at the end of it. Treat the number as a fast first answer, and confirm it in the live model before it appears in a sign-off.
Clash and fit questions need the most care. Structure tells you what is connected, but whether two parts physically interfere depends on geometry and tolerances. For that category, see how AI catches assembly clashes before the shop floor for what automated checks can and cannot establish. An assistant that answers confidently about interference without showing its basis deserves suspicion.
Answers that cite where they came from are easier to trust in a meeting. Leo provides citations, so a reviewer can open the source behind a statement instead of taking it on faith.
What it does not replace, and how to test it
A synced assembly in an assistant does not replace Onshape. It cannot edit the model, it does not run your release process, and it does not certify that a design is correct. Onshape's own AI features work inside the environment and have their own strengths; what Onshape's built-in AI does and where an outside assistant still helps draws that line. If you are comparing the wider field, the roundup of AI tools for Onshape lists what else is available and what each needs from your team.
The safest way to adopt any of this is a small pilot. Pick one heavy assembly your team knows thoroughly. Write down ten questions whose answers you already trust: part counts, a few locations, two or three repeated parts. Run them through the assistant and record how many answers match, how long each took, and whether the sources behind them can be opened. Then try three questions where you do not know the answer and check those by hand in Onshape.
Security belongs in the same pilot. Leo is SOC-2 certified and GDPR compliant, no AI is trained on customer data, and customer intellectual property is protected. Ask every vendor on your shortlist for equivalent statements in writing before any assembly is synced.
FAQ
Ask your assembly questions
Structure, repeats and part locations, answered from the synced model.
Leo is an AI assistant for mechanical engineers. It connects to your CAD, PDM and PLM, and every answer cites its source so you can verify it.
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
