How LLMs Are Reshaping CAD: Capabilities, Boundaries, and Selection Criteria
Large language models are finding real applications in CAD workflows, but their capabilities have clear boundaries. This article examines what LLMs can and cannot do in CAD today—including script generation, parametric modifications, standard part automation, 2D/3D collaboration, and cloud architectures—along with data security requirements and a practical framework for evaluating LLM-based CAD solutions before deployment.
What LLMs Can and Cannot Do in CAD
The conclusion first: current large language models are better suited to assisting with script generation, parameter modifications, and standard part generation. They are not yet capable of independently handling complex assemblies, motion simulations, or safety-related design decisions. These tasks require human review and should be carried out under clearly defined acceptance criteria.
Why Does CAD Need LLMs?
The core value of traditional CAD software lies in precise geometric modeling and engineering drawing. But it has a steep learning curve, involves a large amount of repetitive work, and its modeling workflows depend heavily on engineers’ experience. As product structures grow more complex, enterprises want to shorten design cycles and reduce trial-and-error costs—this is exactly where LLMs can step in.
The strengths of LLMs are code generation, multimodal understanding, and reasoning. They can interpret natural language instructions, generate executable modeling scripts, and form a certain semantic understanding of 3D scenes. This gives CAD tools the opportunity to evolve from “humans operating tools” toward “tools understanding human intent.” However, this shift only delivers real value when the model can accurately handle geometric constraints, parametric relationships, and material properties. There is a significant gap between natural language phrases like “smooth outer surface” or “appropriate fit clearance” and the precise constraints in CAD—this is the well-known semantic gap.
How Far Can Script Generation Go?
LLMs can translate design intent into scripts. For example, given the input “create a cylinder with a diameter of 50mm and a height of 80mm,” the model can output a Python script or software-specific macro commands. This lowers the entry barrier for beginners and reduces the time senior engineers spend on repetitive coding.
Modification tasks are harder than creating new models. For instance, if a user asks to “change the gear tooth count from 20 to 24 and automatically adjust the pitch diameter,” the pitch diameter is determined by both the module m and tooth count z, i.e., d = m × z. The model must identify the parametric relationships in the existing model and generate a targeted modification script, rather than rebuilding the entire part from scratch. This requires the model to understand CAx data structures and accurately invoke the software’s API. If the model simply rebuilds a new part, the existing assembly relationships and feature history may be lost.
Material properties are another easily overlooked issue. Take “material: aluminum alloy” as an example—materials are typically maintained through the software’s material library or a PLM system. Material models and APIs differ significantly across CAD kernels, and the same script may fail on different software versions. Therefore, script examples must specify the runtime environment, software version, and required permissions, and should be validated in a test environment first.
Can Standard Parts and Scene Layout Be Automated?
LLMs can batch-generate standard parts such as bolts and bearings, and generate pattern features according to rules. During assembly layout, the model can adjust component positions and some parameters based on interference checks or load analysis results.
Such outputs must be fully compatible with the geometric boundaries and topology of the CAD kernel to be directly usable in simulation and manufacturing workflows. A solid that looks similar but has non-closed topology may cause downstream processes to fail outright. In this role, the LLM acts more like a script engineer familiar with engineering standards—its output must pass geometric checks before use.
How Should 2D, 3D, and Cloud Work Together?
Many manufacturing enterprises still deliver 2D engineering drawings, but the 3D model is the true source of design. A more sensible workflow uses 3D as the core, with 2D drawings automatically derived from it. Models can generate standard-compliant projection views, dimension annotations, and tolerance information, reducing manual drafting time. In addition, datum symbols, surface roughness values, and other information in 2D drawings can be mapped back to the 3D model to keep the two data sets consistent. This reverse mapping requires manual spot checks, because recognition errors will propagate directly into downstream processes.
Cloud adoption is one important direction for reducing hardware dependencies and enabling cross-site collaboration, but it is not suitable for every scenario. CAD operations are latency-sensitive, especially interactions like rotating and zooming. A common approach is to run lightweight preprocessing locally and offload heavy inference to the cloud. Whether such a hybrid architecture is viable should be judged by measured end-to-end latency, validated against users’ actual operating experience. With poor network conditions, or when data cannot leave a specific jurisdiction, a pure cloud solution may not be an option.
How to Handle Data Security and Compliance?
Industrial design data touches the core competitiveness of an enterprise. When introducing LLMs, data isolation, encrypted transmission, and private deployment are baseline requirements. Many enterprises prefer to deploy models in their own private or hybrid clouds rather than uploading data to public SaaS platforms. This requires the service to support containerization and microservices, integrate with existing PLM and ERP systems, and follow the principle of least privilege.
During implementation, use the following checklist:
- Have training and inference data been de-identified, with customer names and project numbers removed from drawings?
- Are model weights stored encrypted, with access controlled through authentication and permission management?
- Is every script execution written to an audit log, including input instructions, generated scripts, execution results, and the operator?
- Is the model service isolated from the production environment, with test data and production data stored separately?
- Do the PLM and ERP integrations expose only the necessary read/write permissions?
Script execution also carries risks. If the model accidentally modifies parameters in a production model, the team needs to be able to roll back quickly. Therefore, all modifications should first run in a sandbox, with version snapshots retained, and be written to production data only after human confirmation.
How to Evaluate an LLM-Based CAD Solution?
When evaluating, start from the following dimensions and set quantifiable acceptance criteria for each.
| Evaluation Dimension | Key Question | Acceptance Metric |
|---|---|---|
| Geometric Understanding | Can it recognize native structures such as B-Rep and CSG and directly manipulate topology, rather than only generating command streams | Topology check pass rate |
| Script Correctness | Can generated scripts pass geometric checks, and do results match design intent | Geometric error rate |
| Rollback Capability | Are complete command logs retained, with an undo mechanism | Rollback success rate and time |
| Ecosystem Compatibility | Can it run on mainstream CAD kernels, or provide a standalone runtime | Integration effort (hours) |
| Private Deployment and Fine-Tuning | Can enterprises fine-tune with their own drawing data to meet industry requirements | Deployment cost and fine-tuning cycle |
| Human Review | How long does an engineer need to review outputs | Time per review |
For acceptance testing, use a set of historical tasks as a test set and compare the time and error counts between manual work and model-assisted work. The test set should cover typical tasks such as standard parts, parametric modifications, and simple patterns—not just cherry-picked success cases.
Which Tasks Are Best for a Pilot?
Start with low-risk tasks, such as standard part generation, parametric modifications, and derived drawing views. Errors in these tasks are easy to detect and cheap to fix. Once the team has formed a judgment about output quality, gradually expand to more complex feature modeling.
