The State of AI-Generated CAD in Industry

In 1984, a group of engineers at ICAD, Inc. developed the ICAD system, a knowledge-based engineering (KBE) system built to run on the Symbolics LISP Machine.^[1] Their aim was for computer aided design to be automated by a centralised "brain" that had the knowledge of an expert engineer.

Now four decades later, AI is currently entering the engineering industry and it's achieving results in three main areas:

1. Generative AI applied to topology optimisation
2. Startups building text and image-to-CAD foundation models
3. Established CAD vendors building in "AI assistance" into existing products

Through history, AI and CAD have had a fairly intimate relationship.

CAD was one of the most popular (and successful) applications of the expert design systems of the '80s with KBE products that were valued at hundreds of thousands and used by the likes of Boeing and General Motors. Then, we had AI built by AutoDesk to search through massive databases of designs by their geometry, rather than just keyword tags. But now, established companies and startups are trying to use AI to generate designs directly from user queries, rather than just act as a nice search tool.

All together, industry estimates put the AI-for-CAD market at $2.3B in 2023 and project it to reach $12.6B by 2033, growing at an 18.5% CAGR.^[2] What follows is a look at what's actually working in each tier, what's still early, and an analysis of the bottlenecks of generative-AI in CAD.

Using AI to Make CAD Faster

Topology optimisation is a method in computer aided design that has been around for over 30 years and is well-established in industry.^[3]

At a high level, topology optimisation works by designing a new shape from a boundary condition and objective (such as minimising compliance) and then forming a shape in a reductive manner.

One of the problems of this technique is that it requires finite-element analysis (FEA) at every step of the optimisation, which makes it very expensive and slow.

But in the past few years, we're now seeing the rise of AI-generated methods, specifically, to speed up this optimisation algorithm and make it cheaper. YC-backed startup Topological is claiming a 1930x speed up over traditional methods with under 5% compliance error.^[4]

But interestingly, Ole Sigmund, one of the leading figures in topology optimisation and creator of the SIMP method, has argued that AI as a TO substitute is a dead-end: the model "interpolates based on data or observations and is bad at extrapolating."^[5]

This criticism might suggest the creative and generative part of generative AI will really come from the next technology.

The Rise of Text-to-CAD and Image-to-CAD

Several AI startups today are rushing to make "Claude Code for CAD" a reality.
We saw the explosion of ideas and opportunity that came from the birth of agentic coding. Now, can we make make the same thing happen for physical products?

It initially started with image-to-CAD models, where a user provides a 2D image of their design and receives a 3D model back. We're now seeing text-to-CAD models as the next paradigm.

Zoo (formerly KittyCAD) share their vision as "providing the first API for hardware designers." They are building a programming language (KCL) to define B-Rep geometry as a program: a set of instructions for how to construct a given geometry in a CAD engine.

Their agent (the Zookeeper) then automates the process of writing these programs, allowing users to give a text prompt and get a CAD model back where the underlying code behind the design is interpretable and editable.

Autodesk is making a similar bet, albeit at a much larger scale. They have a research project called Project Bernini which has now become a foundation model trained on 10 million 3D shapes that produces fully editable, parametric geometry from a single prompt. This has been termed "neural CAD" and has recently generated a functional air-fryer from one prompt.

While earlier (mostly academic) research has been doing text to 3D for a long time, the 3D output has never been directly editable and so it must be re-interpreted from a mesh or voxel to editable B-Rep geometry. This is what makes Autodesk's "neural CAD" interesting: we now generate the geometry and its interpretation directly from the text prompt.

The Successes

So, as with almost all generative AI of the past few years, image-to-CAD and text-to-CAD models have been getting better on benchmarks.

Modifications such as introducing a VLM critic in the loop have resulted in considerable improvements in geometric accuracy on benchmarks like CADPrompt.^[6] Recent work, such as by Berger et al., have also included FEA as part of the agentic loop, so the agent can partially verify the physical performance of the part it designs, like the safety factor (how much stronger a system, part, or structure is than it needs to be for normal use).^[7]

Another example: AgentCAD is an MCP server and CLI for coding agents to design CAD models. The models can specify constraints (e.g. that there must be "2 circular holes of diameter 5mm") and design their CAD models in CADQuery or build123d. The tool then gives the agent an interface to view their own design in 3D from different angles and get constructive error feedback to help them iterate.

All of these modifications can be described as adding more infrastructure for the model to verify its outputs.

Unlike in coding, it is currently impossible to formally specify and verify a CAD model like a program. At the same time, coding agents are building CAD models in python and the ways we are representing software and hardware, are slowly converging.

This gap is what has resulted in ad-hoc verification approaches such as those described above.

The Key Limitations

While text-to-CAD and image-to-CAD have had significant investment over the past few years, and the market is predicted to grow by a factor of 10, the successes we've seen are mostly academic benchmarks or company claims measured against each other and they lack clear, measurable success in industry.

Berger et al. tested four frontier models: Claude Opus 4.5, Claude Sonnet 4.5, Gemini 3 Pro, Gemini 3 Flash.^[7:1]

Their test involved a four agent pipeline for designing and evaluating designs, including an agent for planning, an agent for engineering, and VLMs for reviewing geometry and physical robustness.

They then tested how well each model made it through different stages of evaluation:

R1: Execution Success - Did the generated code run without a geometry error?
R2: Meshing Success - Could the resulting shape be meshed at all?
R3: FEA Success - Did the physics simulation run without erroring?
Target safety factor - Did the result of that simulation land in the safe range?

What they found was counter-intuitive: the model with the lowest failure rates was the smallest model. All models succeeded at syntax but failed consistently at one thing: geometric validity.

Gemini 3 Flash (the smallest model) produced geometrically or physically invalid designs only 8.7% of the time, while Sonnet 4.5 failed 85.7% of the time. And on average, for every 10 designs the models produced, 4 of those designs would be invalid, even with a VLM in the loop.

The most common failure mode by a wide margin was violating the design space and producing disconnected geometry, a problem consistent across every model.

So, even though models could generate correct syntax and successfully mesh their design and run it through FEA, the result would still be physically and geometrically invalid most of the time.

This would be like if Claude Code spent 5 minutes building an app, it ended without testing the code, AND the result was wrong 4 out of 10 times. It would be a near-pointless tool to use and, for the user, a very quick path to bankruptcy.

To produce more correct results more often, the problem comes down to how a frontier model can accurately test its output and iterate on test results.

When SWE-bench was first introduced, models acting in a zero-shot, single-turn capacity scored less than 4%, while being able to verify and iterate on their outputs, on average, tripled that score.^[8] Similar boosts to performance can be observed in almost any software engineering benchmark.

Another, potentially much more powerful consequence is that more reliable verification leads to better RL.

Reinforcement learning from verifiable rewards (RLVR) improvements are dramatically larger than any improvements that come from being able to verify and iterate in an agent loop.

And the performance improvements we are seeing are specifically high in verifiable domains: frontier models are now consistently achieving gold in the IMO and top scores in programming competitions.^[9] This was not conceivable more than 2 years ago.

If CAD can be made more verifiable, the improvements will mean better inference-time results and significantly better models.

Even better: if AI is able to reason purely about specifications the engineer provides and we can verify against that specification, there may be no longer a need to even look at the geometry.

Conclusion

AI is achieving results in three categorically different areas. On one end, topology optimisation is well-established and trusted, and the AI being added to it is mostly making a classical algorithm cheaper and faster.

On the opposite end, text-to-CAD is the category that is attempting something new, and yet to be established in the same way agentic coding has become established in software engineering.

Its dominant points of weakness aren't problems with functional correctness (yet), but with geometrical weakness, like designs with incorrect dimensions or disconnected geometry. Verifying these problems costs just a few lines of code to implement, but that's not the point. They show us that we still have some way to go before we have reliable agentic coding for CAD. Furthermore, the development of LLMs over the past year has taught us that verification is critical. Not only for agents to iterate better and faster, but also for training models, specifically by RLVR.

In 1984, the answer to automating design was to write down everything the expert knew. Today's models arguably have that knowledge already and the missing piece for generative CAD might be a lot less glamorous than a brain. It might just be a checker.

Click here for an outline of what I think this CAD checker will look like and how better verification might change the way we think about design.