A year ago, asking an AI to help you build a 3D CAD model felt a bit like asking a very enthusiastic intern who had read every engineering textbook but never actually touched a part. They knew all the right vocabulary, they'd give you a confident answer, and about half the time that answer would be subtly wrong in ways that weren't obvious until you were already committed.

That's still partially true. But the gap has closed faster than I expected, and the engineers getting real value out of AI today aren't doing anything exotic. They've just figured out where to lean on it and where not to.

This is what we've learned from watching that process play out.

The blank canvas problem is real, and AI actually solves it

Every designer knows the feeling: you've got a brief, a deadline, and a completely empty viewport staring back at you. The hard part isn't the complex geometry you'll build later. It's finding the starting point. What proportions? What topology? Where do the interfaces go?

AI is surprisingly good at this. Describe a rough enclosure, a mounting bracket, a heat sink concept, something general, and you get a workable starting point in seconds. It won't be right. The wall thicknesses might be off, the fastener pattern might ignore your standard library, and the fillets might make a machinist raise an eyebrow. But you have something to react to, which is worth a lot more than a blank screen.

The blank canvas problem is genuinely underestimated as a productivity drain. If you've spent any time watching designers work, you know that the first ten minutes of a new part are often the least productive: lots of staring, second-guessing, false starts. Eliminating that with a rough AI-generated starting point isn't magic, but it consistently shaves meaningful time off early-stage work.

Where AI reliably earns its keep

Beyond that first-draft problem, a few use cases have proven consistently valuable:

  • Rapid variation testing: When you need to evaluate five slightly different versions of a snap-fit tab, a rib pattern, or a boss configuration, generating those variants manually is tedious. AI can produce them quickly enough that you can actually compare options rather than just committing to whatever seems most plausible in your head.
  • Prompt-to-parameter sanity checks: Describe a part and its process, and AI can suggest reasonable starting parameters (wall thicknesses, draft angles, fillet radii) that are at least in the right ballpark. Think of it less as an oracle and more as a knowledgeable colleague who's seen a lot of parts and can tell you when something is clearly out of range.
  • Documentation that nobody wants to write: Design intent notes, first-pass BOMs, inspection criteria, review summaries: all of this is work that needs doing and that most engineers find deeply tedious. AI handles the drafting competently. You still review and correct it, but you're reviewing rather than writing from scratch, which is a different and faster mental mode.
  • Exploring the unfamiliar: Working in a material or process you don't know well? AI can give you a reasonable orientation: typical design rules for aluminium die casting, common pitfalls with PEEK, what to watch for with selective laser sintering. It's not a substitute for domain expertise, but it's a useful starting point for building it.

Where AI consistently bites people

This part matters more than the benefits section, because the failures are less obvious and more expensive.

Wall thicknesses are frequently wrong for the stated process. An AI prompted for an injection-moulded part will sometimes generate walls that are 8 to 10 mm thick. That is technically valid geometry, but a nightmare for sinking and warping. It knows the vocabulary of DFM, but it doesn't always connect the vocabulary to the physics. Always check wall thickness against your process constraints, not just against "does this look reasonable in the viewport."

Hole patterns look right but often aren't. AI will place holes somewhere sensible-looking without knowing your fastener library, your supplier's preferred drill sizes, or the minimum edge distances for your specific material and load case. M4.5 holes where M5s were expected, holes too close to a thin wall. These are common problems, easy to miss if you're moving fast.

Internal fillets and machining accessibility. This is the one that catches people most often. AI-generated geometry sometimes includes internal corners that are geometrically valid but practically impossible to machine without specialist tooling. It's worth asking yourself, for every internal feature: how does a cutter get in here? If the answer isn't obvious, it probably needs redesigning.

Tolerance and fit decisions require context AI doesn't have. Clearances, interference fits, GD&T call-outs: these depend on your production process, your inspection capabilities, your supplier relationships, your assembly sequence. AI can give you nominal dimensions; it can't give you the judgment that comes from having parts come back out of tolerance three times in a row.

None of this is fatal. It just means AI output is a starting point, not a deliverable. The engineers who get burned are the ones who stop reviewing geometry once AI is involved, as if "the AI checked it" is a substitute for "an engineer checked it."

Prompting well is an actual skill

The quality gap between vague prompts and specific ones is enormous, more than most people appreciate until they've spent time on both ends.

"A bracket" produces generic geometry that will probably need to be thrown away. "An aluminium 6061 mounting bracket, 80 × 40 mm base footprint, 3 mm wall thickness, two M6 clearance holes at 60 mm centres, 4 mm radius corners, designed for CNC machining" produces something you can actually work with.

A few things that consistently improve output:

  • Name the manufacturing process explicitly. Injection-moulded, die-cast, machined, and 3D-printed parts all have different design rules. If you don't say which one, you'll get geometry that implicitly assumes something, and it might not be what you need.
  • State the material. Not just "aluminium" but which alloy, if it matters. Not just "plastic" but PA66 or ABS or PEEK. The more specific you are, the more appropriate the output.
  • Describe function, not just shape. "A cover that clips onto a rail and prevents dust ingress" will outperform "a rectangular cover with clips" because the first gives the AI useful context about what the part needs to do.
  • Iterate, don't regenerate. The fastest path to good geometry is usually refining an existing draft rather than generating new ones from scratch. Get something on screen, identify what's wrong, give specific correction instructions, repeat.

The workflow that's actually working

The engineers getting consistent value from AI-assisted CAD have converged on something similar: use AI to eliminate the tedious early stages, then apply full engineering rigour to everything that comes out of it.

Concretely, that looks like: let AI generate a first draft to escape the blank canvas, review it critically the way you'd review a junior engineer's submission, refine with specific prompts until the geometry is close, then do a proper DFM and tolerance analysis before it goes anywhere near a drawing or a purchase order.

The last step matters. It's easy to let the momentum of having a "finished-looking" model compress the rigour that should come after it. Good geometry on screen is not the same as a manufacturable, reliable part.

One habit worth building: log the AI-assisted steps in your design history. Not because anyone's likely to audit them, but because "we used AI to generate the initial bracket geometry, then reviewed and modified based on our standard fastener library and DFM guidelines" is a genuinely useful note when someone is tracing a decision six months later. It costs thirty seconds and occasionally saves hours.

An honest take on where this is heading

The capabilities have improved a lot in twelve months. The rate of improvement hasn't slowed. There will almost certainly be things AI-assisted CAD can do in another year that it genuinely can't do today.

But the fundamentals of engineering judgment aren't going anywhere: knowing your supplier, understanding the physics, owning the design decisions, being accountable for what gets built. The value of AI here isn't that it replaces that judgment. It's that it removes the parts of the workflow that were consuming time without requiring it.

That's not a small thing. If you get an extra hour of genuine design work out of every day because AI is handling the rough drafts and the documentation, that compounds. The engineers who figure out how to use these tools well aren't working less hard. They're working on harder problems, which is where good engineers want to be anyway.