Text to CAD: how AI turns a prompt into a 3D model
Text to CAD means typing a description of a part and getting back a 3D model that CAD software can open. The tools work in different ways, and the way they work decides what you can do with the result. This guide explains the approaches and gives you a test to run on a part you make. It ends with the plant’s version, where the order is described in words and the configurator builds it from your rules.
What text to CAD is
Text to CAD is a kind of generative AI tool for mechanical design. You describe a part in words, such as a bracket of a given size with a row of holes, and the tool returns geometry. The useful output is a solid model that opens in CAD software, usually as a STEP file.
Three neighbouring tools are easy to confuse with it. Text-to-3D generators make meshes for games and renderings, which look right but carry no exact dimensions. Chatbots can write code for a scripted CAD tool, which then builds the part. Assistants inside CAD software take instructions in words, but they work on a model you already have.
How text-to-CAD tools turn a prompt into geometry
Under the interface, the tools take one of a few routes. Each route explains a different kind of mistake.
- A model trained on CAD sequences. Research systems learn to output the steps a designer would take: a sketch, an extrusion, then another sketch. Text2CAD, presented at NeurIPS 2024, generates sequences of 2D sketches and extrusions from prompts written at beginner to expert level. The output is exact geometry, limited to the operations the model learned.
- A language model writing CAD code. A general model writes a script for a code-based CAD system, and the CAD kernel runs it to build the solid. The geometry is exact and the script can be read and edited. The script can also run without errors and still build the wrong part.
- A product built around an agent. Commercial tools wrap these methods in a conversation. Zoo described its Text-to-CAD as generating B-rep models from text prompts, and its Text-to-CAD page now presents Zookeeper, a conversational CAD agent.
Whichever route a tool takes, the geometry comes from the prompt plus what the model learned in training. Nothing in that process reads your drawings, your standard parts or your rules unless you put them in the prompt.
Text-to-CAD tools you can try
These are the kinds of tools a manufacturer will meet. The list explains what each does and is not a ranking.
| Tool | What it does | What you get |
|---|---|---|
| Zoo | A conversational agent that creates and edits CAD models from text, with an API for developers | Models exported as STEP, STL, glTF, OBJ or PLY |
| Adam | An AI copilot that models inside Onshape and Autodesk Fusion | Features in a model you own |
| Leo AI | An engineering copilot for mechanical engineers, with sourced answers, part search, concepts and 3D models | Answers and models, with links to SOLIDWORKS, Inventor, Fusion and NX |
| Autodesk Fusion | Autodesk Assistant takes commands in natural language inside Fusion | Actions on your own model, plus automated drawings |
| Research code | Academic models such as Text2CAD, released for experiments | Sketch-and-extrude sequences |
The products change often, so check each vendor’s own page for what it does today. The test below works on any of them.
What a good prompt contains
A text-to-CAD prompt works like a phone call to a drafter who cannot ask questions. Everything the part needs has to be in the words, in an order that is easy to follow.
- The units, stated once at the start.
- The overall shape and its main dimensions.
- Each feature with its size and its position from an edge or a centre line.
- The material and its thickness, if the tool can use them.
- What is symmetric or repeated, such as a hole pattern on a pitch.
Short sentences help, one feature per sentence. If the result is wrong, change one sentence at a time, so you can tell which instruction the tool misread.
A test you can run on your own part
Run the same prompt through each tool and judge the result the way your shop would. Use a part you know well. Here is a prompt for a simple machine guard panel, which you can rewrite with a size you make.
“A flat steel guard panel, 600 mm wide, 900 mm tall and 2 mm thick. It has a 25 mm return flange on all four sides. Put a 10 mm mounting hole 20 mm in from each corner. Fill the centre with 8 mm holes on a 20 mm pitch, leaving a 60 mm border.”
- Open the file in your CAD software, or in the free STEP file viewer, and measure the overall size.
- Check that it is a solid with real flanges and holes, rather than a mesh.
- Count the holes, and measure the pitch, the border and the corner holes.
- Ask for a change, such as 750 mm wide, and check what moved with it.
- Ask for a flat pattern and a drawing, and note what the tool cannot produce.
Keep the files and your notes. The tools change quickly, and the same test tells you next quarter whether anything improved on the parts you make.
Where text to CAD stops
The tools handle shapes that a few sentences can describe completely. They stop where a part depends on knowledge that the prompt does not carry.
- Your standards. Bend radii, hole sizes for your fasteners, preferred gauges and the sections you stock.
- Manufacturability. Whether your press brake can form the flange, or your laser can cut the hole pattern at that pitch.
- Tolerances and fits. A generated model carries nominal sizes. The tolerances that make parts fit together are still a person’s decision.
- Assemblies. A product of many parts needs every part to agree with the others, which is hard to describe in one prompt.
- The drawing and the bill. A shop builds from a drawing and buys from a bill of materials, and many tools stop at the model.
Each of those gaps comes from information that lives in the plant rather than in the prompt. A longer prompt can carry some of it, but a person has to know it and type it every time.
What happens to the files you send
Many text-to-CAD tools run in the vendor’s cloud. A prompt that describes a customer’s part, or a file you upload to edit, leaves your building. Check each tool’s terms for where files are stored, how long they are kept and whether they train the vendor’s models.
For tests, describe your own parts rather than a customer’s. Is ChatGPT safe for confidential information covers the same question for chat tools and drawings.
The plant’s version: describe the order in words
For a product family a plant sells in many sizes, text to CAD can work the other way round. The customer or the sales engineer describes the order in words. A language model maps those words to options in your configurator, and your rules build the rest from your own templates.
The language model’s job is narrow. It turns words into choices from a list you control, and it asks a question when the text is ambiguous. It never invents a dimension, so every number in the model comes from a rule someone can open.
AI CAD calls this the configured route and compares it with the other kinds of tool. CPQ for manufacturing covers how the price comes out of the same configuration as the model.
When to use which
Use a text-to-CAD tool for a one-off shape you would otherwise sketch, such as a spacer, a bracket or a mock-up. Use the configured route for the products you sell again and again in different sizes, where the rules exist and the drawing has to be right. A plant that builds to order usually has work of both kinds.
A one-off part that starts to repeat is worth noticing. When a third customer asks for the same spacer in a new size, stop generating it. Add it to the configurator as a template with its own rules. From then on, every size comes out the same way, with your standards and your drawing.
In both cases, a named person checks the model before anyone cuts metal. The AI CPQ offer describes how the configured route is set up in four weeks on one product family.