Can AI Do CAD? What It Can and Cannot Model

Last updated: August 2026

“Can AI do CAD” sounds like a yes-or-no question, but the honest answer depends on what you mean by “do.” If you mean can AI help create and modify a 3D model, yes, in several specific and genuinely useful ways. If you mean can it take a prompt and hand you a finished, editable, production-ready parametric model with no engineer driving the work, no. The interesting part is the space in between, where AI assists real CAD work without replacing the person at the keyboard.

This guide answers the capability question directly, breaks “AI doing CAD” into the four things it actually means today, and shows where each one helps and where it stops. It is the capability companion to our guides on whether AI CAD tools are worth it and the problems with AI CAD software, and it sits under our roundup of AI tools for mechanical engineering.

Short answer: AI assists CAD in four real ways and autonomously replaces none of them. Generative design and topology optimization explore many manufacturable options from constraints and loads you define. In-CAD copilots automate discrete steps such as constraints, toolpaths, and drawings. Text-to-CAD turns a prompt into geometry, though quality depends heavily on the prompt and degrades on complexity. And general language models can write CAD scripts you run and verify. In every case the engineer still owns design intent, parametric editability at complexity, and the final result. Research shows autonomous first-pass generation is unreliable and improves mainly with human refinement, so treat AI as a powerful assistant in CAD, not the modeler.

A person viewing a 3D model on a desktop monitor in modeling software — AI can accelerate parts of CAD, but a person still drives the modeler. Photo: Pexels.

What “doing CAD” actually means

Set the bar first, because it is what separates hype from capability. Doing CAD means creating and modifying editable, parametric 3D models: geometry that carries design intent, that another engineer can open and change, and that is fit to drive manufacturing. A one-off mesh generated from a prompt is not the same thing as a parametric model with a feature history you can edit. Keep that distinction in mind through the four sub-capabilities below, because most of the honest limits come down to editability and design intent.

Sub-capability 1: generative design and topology optimization

This is the most mature form and the most misunderstood. Generative design in Autodesk Fusion, and implicit-modeling platforms such as nTop, explore many design variants from goals, loads, materials, and manufacturing constraints that the engineer defines. nTop describes its own approach as exploring hundreds of design variants systematically, with manufacturing constraints and performance requirements built in (nTop). That is constrained optimization, not a model that designs for you. It is powerful, but you set up the problem, and the organic output often needs mesh-to-solid cleanup before it is production CAD.

Sub-capability 2: AI copilots inside CAD

The second form is the assistant built into the CAD tool. Autodesk has announced an in-context assistant for Fusion that turns natural-language prompts into geometry, runs commands, writes custom code, and automates sketch constraints and toolpaths, with a newly announced capability to generate editable geometry from a single prompt that is rolling out rather than proven mature (Autodesk News, September 2025, vendor claim). The clearest statement of the boundary comes from a competitor: Onshape says its AI Advisor is a knowledge assistant that “does not generate designs” or make complex engineering decisions (Onshape). Copilots accelerate steps; the human still drives the modeler.

Text-to-CAD turns a prompt into geometry, but the output and its editability are the real test. Video: 3D Comparison via YouTube.

Sub-capability 3: text-to-CAD

The newest and most eye-catching form generates a 3D model from a text prompt. Zoo Text-to-CAD produces models from a single prompt, with output as code in its own language plus exportable CAD formats, and the vendor claims fully editable models (Zoo, vendor claim). The honest caveat comes from the same company: in its research, Zoo notes that most generative 3D work fails to produce tools useful to working engineers because prior approaches lack coherent design intent and intent-preserving edits (Zoo research). Quality is highly dependent on a detailed, spec-rich prompt and degrades as parts get complex. Text-to-CAD is real and improving, but it is not yet a substitute for a modeler on anything intricate.

Sub-capability 4: general LLMs writing CAD scripts

The fourth form is the one most engineers reach for: asking ChatGPT or Claude to write a script for a scriptable CAD system such as OpenSCAD, CadQuery, or FreeCAD Python. CadQuery is literally a parametric scripting language for building CAD models (CadQuery), so a model can emit a script that a human runs and verifies. The reliability ceiling is documented. In the Query2CAD study, GPT-4 Turbo generating CAD from natural language succeeded about 53.6 percent of the time on the first attempt, rising to roughly 76.7 percent only after iterative refinement with a human in the loop (Query2CAD, arXiv 2024). And a NASA Glenn report found ChatGPT could help build a plotting tool and an optimization method but failed to automate an analysis workflow through scripting (NASA NTRS, 2024). It writes scripts; it does not autonomously model.

What AI still cannot do

Three limits run through all four forms. It does not originate design intent: it optimizes or generates against goals a human sets, and it does not know why the part exists. It struggles with parametric editability at complexity: a generated mesh or a fragile script is not the same as a clean feature history another engineer can safely edit. And it does not reliably automate a real end-to-end workflow on its own, as the NASA scripting failure showed. Those are exactly the parts of CAD that carry engineering responsibility, which is why the human stays in charge. For where this bites in practice, see our guide to the problems with AI CAD software.

Three engineers reviewing work on a laptop together in an engineering lab — AI accelerates steps, but design intent and the final model stay with the engineers. Photo: Pexels.

How engineers should actually use AI in CAD today

Used well, AI is a real time-saver in CAD. A practical approach.

Use generative design for exploration. Set the loads, materials, and constraints, and let it surface options you can evaluate, then clean up the winner into production geometry.
Let copilots handle the busywork. Constraints, repetitive features, drawings, and code snippets are good targets; keep the modeling decisions yours.
Treat text-to-CAD as a starting point. Use it for simple parts and first drafts, write detailed prompts, and expect to rebuild or heavily edit anything complex.
Have AI write scripts, then run and verify them. A CadQuery or OpenSCAD script is easy to check; never trust generated geometry you have not confirmed.
Keep the engineer accountable. Design intent, parametric quality, and the final model are yours; AI assists, it does not sign off. Before uploading proprietary models, check how the tool handles your data, as we cover in can I upload CAD files to ChatGPT.

Frequently asked questions

Can ChatGPT do CAD?

Not by modeling directly, but it can write scripts for scriptable CAD systems such as OpenSCAD, CadQuery, or FreeCAD Python that you run and verify. Research puts first-attempt success for natural-language CAD generation around 54 percent, improving mainly with human refinement, so treat its output as a draft to check rather than a finished model. It is genuinely useful for scripts, boilerplate, and explanation, not for autonomous, production-ready modeling.

Is text-to-CAD parametric and editable?

Sometimes, but with caveats. Some text-to-CAD tools output code and exportable CAD formats and claim editability, yet quality depends heavily on a detailed prompt and degrades on complex parts, and generated geometry can lack the clean design intent that makes a model safe to edit. For simple parts it can be a real head start; for intricate ones, expect to rebuild. Treat vendor editability claims as claims to test on your own work.

Will AI replace CAD engineers?

Not on current evidence. AI automates discrete steps and explores options, but it does not originate design intent, guarantee parametric editability at complexity, or reliably automate a full workflow on its own. Every mature capability, from generative design to copilots, is set up and finished by an engineer. The realistic outcome is that AI shifts where engineers spend time, not that it removes the engineer from CAD.

Sources

Written by the CognitiveFuture editorial team. We build our guidance from official product documentation and published research, and we label vendor statements as such. We do not independently benchmark any tool, and we do not treat AI-generated geometry or scripts as reliable without verification. Design intent, parametric quality, and the final model belong to a qualified engineer, and generated output must be checked before it drives manufacturing.

About the author

Richard Johnson

Richard Johnson is an AI specialist at one of the world's largest technology companies, where he has spent the past three years helping organizations adopt AI. CognitiveFuture extends that work publicly: gathering the available evidence on each tool, from vendor documentation to independent reviews and user feedback, and cutting a crowded market down to the right choice for the job in front of you.