Why is CAD Hard for AI?

10 min read

For professional designers/engineers and businesses relying on CAD as the backbone of their projects, 3D modeling isn’t just about generating a shape. AI tools today can produce impressive 3D results, but most of them output non-editable meshes or point clouds. What teams actually need is AI that understands design logic and generates geometry that can be modified, iterated, and built.

Large Language Models (LLMs) were originally and are still primarily trained on massive amounts of text. While modern multimodal models now also learn from images, audio, and video, CAD remains fundamentally different. It operates on procedural geometry, constraints, and interdependent sequences of operations rather than perceptual or linguistic patterns.

In this article, we’ll explore why CAD remains one of the most challenging domains for AI, from its reliance on design history and structured logic to the lack of usable datasets and the need for step-by-step reasoning. We’ll also explain why translating CAD workflows into structured natural language unlocks more reliable results when working with AI inside tools like Rhino.

CAD Is a Sequence of Design Decisions

what CAD fundamentally is: history, intent, operations.

Most AI-generated 3D (point clouds, meshes, level-set fields) represents finished geometry - static, non-editable shapes. But CAD is fundamentally different: every shape is built through a history-based sequence of operations (sketch → constraint → extrude → trim → boolean → fillet, etc.). AI struggles because geometry alone is not enough. To modify or regenerate a CAD model, the agent must understand:

The order of operations
The intent behind each sketch
The constraints and parametric relationships
The dependencies between features

This is why describing steps and intent leads to better AI results. The model needs a view into the construction history, not just the final shape.

The Limits of LLM Reasoning in CAD

why this process is hard for AI: long-horizon, chain-of-thought, dependency reasoning.

CAD is fundamentally different from the domains where LLMs tend to learn well - fields where knowledge exists in large volumes of text, such as writing, law, medicine, or code. CAD’s knowledge lives in geometry, topology, and parametric logic instead of text.

Where this becomes critical is in how LLMs actually reason. Large language models are autoregressive: they generate one step at a time by predicting what should come next. This works well in writing because small drifts are acceptable, and it works reasonably in coding because the symbolic structure provides guardrails. But CAD workflows have almost no tolerance for drift. Even a slight deviation can break a face, misalign a trim, or collapse a solid. This is fundamentally a long-horizon reasoning problem. Without a CAD agent orchestrating each step, autoregressive inference quickly leads to hallucinated actions and structurally invalid geometry.

The Structural Incompatibility Between LLMs and B-rep Validity

why even if AI reasons correctly, it still breaks geometry.

Even if an AI reasoned correctly, CAD presents a second, deeper challenge: B-rep geometry itself is incompatible with token-based models. Almost all industry-standard CAD systems rely on boundary representations (B-reps) to describe geometry. A B-rep encodes a precise network of trimmed NURBS surfaces, edges, vertices, tolerances, and adjacency relationships. This representation is mathematical, topological, and highly interdependent - not textual, not visual, and not linear. LLMs operate on linear token sequences, not geometric structures, so they cannot reliably maintain a valid B-rep. Even tiny mistakes can break how surfaces connect, how trims align, or whether a solid can close. This makes direct LLM manipulation of B-rep geometry extremely fragile.

In practice, geometry editing requires a level of structure and validation that an LLM alone cannot provide. A reliable CAD agent must issue changes through structured APIs, use constraint-aware operations, and validate each modification as it happens. In simple terms: it needs a guaranteed, trustworthy Ctrl+Z.

Lack of Existing CAD Datasets

why CAD is a blank frontier for AI: no training data, no public models, no design history.

So, how and where is a model supposed to learn construction history? One of the biggest challenges is simply that CAD data doesn’t exist at scale. Large language models learn from trillions of tokens of text and millions of images, but CAD models are rarely shared publicly. Most CAD files are proprietary, locked inside companies, or bound by IP restrictions. Mathematical and topological structures where CAD operates on - Manifold surfaces, NURBS, parametric constraints, and geometric kernels - never appear in LLM pretraining data. Even when datasets do exist, they’re often exported as meshes or point clouds, losing the design history, constraints, parametric dimensions, and feature trees that AI actually needs to learn from. Without large, diverse datasets of process-level CAD modeling, it’s nearly impossible for AI to generalize the structure and logic behind real-world geometry.

Why This Matters to All of Us

Most of us don’t need to care about every technical detail, but these challenges directly shape what becomes possible in day-to-day CAD work: what breaks, what stays editable, and why many AI CAD tools still feel unpredictable today. Understanding these constraints helps us see past the hype and identify tools that actually fit the realities of CAD work.

This is why Reer Agent is designed as a structured, context-aware system - one that is genuinely grounded in how CAD actually works. Our goal is to amplify workflows by reducing the technical friction that interrupts them. When the underlying mechanics become reliable, practitioners across design and engineering can focus on the decisions that matter most.

Further Reading and Research Behind This Post

If you’d like to dive deeper into the academic work behind AI × CAD, here are some of the papers that informed our perspective.

Wu, R., Xiao, C., & Zheng, C. (2021, October). DeepCAD: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 6772–6782). Retrieved from https://arxiv.org/abs/2105.09492
Alam, M. F., & Ahmed, F. (2025). GenCAD: Image-conditioned computer-aided design generation with transformer-based contrastive representation and diffusion priors [Preprint]. arXiv. https://arxiv.org/abs/2409.16294