Why Text-to-CAD Doesn't Work (Yet)
“Design me a phone case with a camera cutout.” Sounds simple. But which phone? How thick are the walls? What's the corner radius? Where exactly is the camera module? How much clearance around the lens? Is it snap-fit or slide-on? Does it need a lip for screen protection?
A single sentence hides dozens of critical dimensions. Text-to-CAD tools have to guess every one of them. And in mechanical design, guessing means the part doesn't fit.
The Dimensionality Problem
A simple bracket has a dozen dimensions. A PCB enclosure has 50–100. A multi-part assembly can have thousands. Each dimension is a degree of freedom. Text-to-CAD has to map a short string of words onto this high-dimensional space.
Think of it this way: language is low-bandwidth, geometry is high-bandwidth. A 20-word prompt might specify 3–4 constraints. The model fills in the other 96 with defaults, training data, or noise.
Example: “a box with a hinged lid”
- › Shape: box
- › Lid: yes, hinged
- › Length, width, height
- › Wall thickness
- › Corner radii
- › Hinge type (living, pin, barrel)
- › Hinge placement and clearance
- › Lid overlap / lip depth
- › Latch mechanism
- › Draft angles
- › Material thickness
- › Mounting features
2 constraints specified. 10+ left to the model's imagination.
Lost in Translation
Even detailed prompts lose information. Language is inherently ambiguous for spatial relationships. Try describing the exact position of a USB-C port on an enclosure using only words:
Near the bottom — how near? 3mm from the edge? 8mm? Centered horizontally on the side face — but relative to the inner wall or the outer wall? These ambiguities compound. A part with 5 ports, 4 screw bosses, and ventilation slots has hundreds of spatial relationships that are trivial to define in a sketch but nearly impossible to fully specify in prose.
This isn't a model quality problem. It's an information density problem. Words are the wrong encoding for geometry.
The Output Problem
Even when text-to-CAD tools produce something, the output is usually a mesh (STL/OBJ) or a script (OpenSCAD). Neither gives you what engineers actually need:
Meshes aren't parametric
You can't change a dimension on a mesh. You can't add a fillet. You can't shell it. If any dimension is wrong, you start over. For manufacturing, you need B-Rep geometry with a feature tree.
Scripts are fragile
Generated OpenSCAD or Python CAD scripts look editable, but they're typically a wall of magic numbers with no meaningful variable names. Changing one value breaks the geometry elsewhere.
No design intent
Real CAD models encode relationships: this hole is concentric with that boss. This wall is offset from that face. Text-to-CAD output has no constraints, no references, no history — just raw shapes at absolute coordinates.
What Actually Works
The bottleneck isn't AI capability. It's the input modality. Text is too lossy for geometry. The approaches that work use higher-bandwidth inputs:
Image → geometry
A photo carries thousands of spatial constraints implicitly — proportions, positions, shapes. Converting an image to a DXF sketch preserves all of them. You don't need to describe the shape; the shape is right there in the pixels.
Text + existing CAD model
“Add four M3 screw bosses to the corners” works when the AI can see the current design. It knows where the corners are, how thick the walls are, and how deep the cavity is. The existing model provides the missing context that text alone can't.
Parametric tools, not raw generation
Instead of generating geometry from scratch, call specific CAD operations (extrude, fillet, shell) with specific parameters. The AI decides what to do; the CAD kernel handles how. The output is a real feature tree, not a mesh blob.
The Bottom Line
Text-to-CAD is a compelling demo. Type a sentence, get a 3D model. But for real engineering work — where tolerances matter, where parts have to fit together, where designs iterate — it produces output that needs so much rework it would have been faster to model from scratch.
The better path: use AI where it adds leverage (interpreting images, choosing operations, calculating parameters) and let proven CAD tools handle the geometry. Meet engineers in the tools they already use, with output they can actually modify.
AI that works with your CAD tool, not instead of it
Bevell connects to Fusion 360 via MCP. 79 tools, parametric output, real feature trees.