Parametric assemblies from natural language: what is real and what is hype

"Describe your machine to the AI and it will design the whole thing." The phrase sells. And it is hype.

I say it with first-hand knowledge: I built a design assistant that generates parametric parts and assemblies from natural language inside Inventor. Precisely because I built it, I know where the line runs between what works today and what is still a pretty demo.

What actually works today

Asking for a part or an assembly in a chat and having the add-in really build it, through the Inventor API, on top of templates and rules defined by engineering. The result is not an image: it is a native CAD model, with parameters, constraints and a feature tree any designer can open and keep editing.

Resolving components against the Content Center instead of inventing geometry out of thin air. If the assembly needs a profile or a standard part, the assistant looks it up where the company library already lives.

Changing dimensions and configurations by talking. "Make it 200 mm longer" works because the model is parametric: the geometry regenerates from the parameters, the way it should.

The architecture that makes it possible

The assistant has 3 layers: the Inventor API, MCP (Model Context Protocol) and the LLM. The LLM understands intent and holds the conversation. MCP exposes a catalog of concrete tools with defined inputs and outputs. And the Inventor API executes the real modeling operations.

The key point: the LLM never draws. It orchestrates. The geometry is built by the API on engineering rules. That is why the result is a real model and not a hallucination with a render.

What is still hype

Free "creative" geometry with no templates or rules behind it: nobody is generating functional mechanical design from scratch with a prompt. Design without engineering judgment: the AI does not know what load your structure carries or which standard applies. And any demo that never lets you open the model: if all you see is a video of the final result, be suspicious.

How to evaluate a demo in 4 questions

Ask to open the model tree: if there are no features and no constraints, it is an imported mesh, not CAD. Ask to change a parameter: if the geometry does not regenerate, it is not parametric. Ask where the standard components come from. And ask for something impossible: a serious system says no; hype always says yes.

Where does this save hours today? In the same repetitive work as always: variants of the same assembly, configurations that change dimensions and standard components, preliminary models for quoting. There the assistant turns an afternoon of clicks into a minutes long conversation, and the designer keeps control of the result.

AI in CAD is real and it already saves hours. But it is the means, not the promise. Whoever sells you otherwise is selling you the demo.