AI for CAD Tools
Text-to-CAD tools can generate single parts, but assemblies break them. Here's why multi-part relationships are the hardest unsolved problem in AI-driven CAD.
·
⏱
9 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Text-to-CAD has real limitations with assemblies, and these are not limitations that more compute or bigger datasets will easily fix. The multi-part problem involves constraint propagation, tolerancing, fit classes, and engineering relationships that current AI architectures cannot reason about. For engineers building real assemblies, the more practical AI approach in 2026 is finding existing validated parts from your organization's design history. Leo AI makes that possible across every major PLM platform, giving you production-proven components instead of AI-generated geometry that does not actually assemble.
Single parts are hard enough. Ask a text-to-CAD tool to generate a "flanged bearing housing with four M8 mounting holes," and you might get something that looks reasonable. Maybe the geometry is mesh instead of parametric. Maybe the dimensions are approximate. But the shape is recognizable, and for conceptual purposes, it works.
Now ask the same tool to generate an assembly. A gearbox with an input shaft, two intermediate shafts, six gears with proper mesh engagement, bearing supports, a split housing with bolt flanges, seals, and a mounting foot. Suddenly everything falls apart. The gears do not mesh correctly. The shaft diameters do not match the bearing bores. The housing halves do not align at the split line. Bolt holes in mating parts are in different locations.
This is not a solvable-with-more-training-data problem. The multi-part challenge exposes fundamental limitations in how current text-to-CAD systems represent and reason about engineering relationships. Understanding why assemblies break these tools helps engineers set realistic expectations for where AI can and cannot help in their design workflow.
Why Single-Part Generation Is a Different Problem Than Assembly Generation
Generating a single part from a text description is fundamentally a geometry prediction task. Given a text input, produce a 3D shape that matches the description. The AI model learns statistical associations between text descriptions and geometric features. "Bracket" correlates with L-shapes and mounting holes. "Shaft" correlates with cylindrical geometry and stepped diameters. These associations are learnable from datasets of parts and their descriptions.
Assembly generation is a different category of problem entirely. An assembly is not a collection of shapes. It is a network of relationships. Every part in an assembly exists in context with every other part. The shaft diameter determines the bearing bore, which determines the housing pocket, which determines the housing wall thickness, which determines the bolt pattern spacing. Change one dimension and it cascades through every connected component.
Text-to-CAD models that generate parts independently and then try to combine them into assemblies produce collections of parts that do not actually work together. The shaft does not fit in the bearing. The bolt holes do not line up. The gear teeth do not mesh. The interference fits are not interference fits. Each part might look correct in isolation, but the assembly as a whole is engineering nonsense.
This is why the Dassault AURA demo at 3DExperience World 2025 was so striking to engineers who watched it, and why the fact that it never shipped tells an important story. Generating assemblies that actually work requires understanding engineering relationships at a level that current AI architectures have not achieved.
IN PRACTICE
The search in Teamcenter has always been a weak point for us. If you don't know the exact part number or file name, you're basically not finding it. Leo changed that.
"The search in Teamcenter has always been a weak point for us. If you don't know the exact part number or file name, you're basically not finding it. Leo changed that."
- Verified User, Defense and Space Enterprise
The Constraint Propagation Problem
In a real SolidWorks or CATIA assembly, parts are related through mates and constraints. A coincident mate locks two faces together. A concentric mate aligns two cylindrical features. A distance mate sets a specific gap between components. These constraints form a system of equations that the CAD solver resolves simultaneously to position every part correctly.
Text-to-CAD systems have no equivalent mechanism. When you describe an assembly in text, you are implicitly referencing hundreds of constraints that you do not spell out. Saying "gearbox with an input shaft and output shaft" implies that the shafts are parallel (or perpendicular, depending on the gear type), that the gear center distances match the gear module and tooth count, that the bearings supporting each shaft are coaxial with the shaft axis, and that the housing provides clearance around all rotating components.
No text prompt captures all of this. And even if you wrote a 2,000-word description specifying every constraint, the AI model would need to resolve all those constraints simultaneously while generating geometry that satisfies each one. This is essentially asking a statistical model to be a constraint solver, which is not what neural networks are designed to do.
The result is assemblies where individual parts look plausible but do not actually fit together. Holes that are close but not aligned. Clearances that are sometimes too tight and sometimes too loose. Mating surfaces that almost match but have slight angular misalignment. The parts pass a visual inspection but fail the moment you try to assemble them in CAD software with real constraints applied.
Tolerancing, Fits, and the Details That Make Assemblies Work
Even if a text-to-CAD system could generate parts that geometrically fit together, assemblies require far more than geometric compatibility. They require engineering compatibility, which involves tolerancing, fits, material interactions, and assembly sequence considerations.
Take a simple shaft-and-bearing assembly. The shaft diameter is not just a number. It is a nominal dimension with a tolerance band that creates a specific fit class with the bearing inner race. An H7/g6 fit on a 25mm shaft means the shaft outer diameter is between 24.987mm and 24.998mm, and the bearing bore is between 25.000mm and 25.021mm. This creates a clearance fit that allows rotation while limiting radial play.
Text-to-CAD systems generate nominal geometry without fit classes, tolerance stacks, or assembly-level dimensional analysis. The shaft might be 25.000mm and the bearing bore might be 25.000mm, which in reality means the shaft will not fit inside the bearing without force, and even with force, there is no running clearance.
Multiply this by every mating interface in a complex assembly: every fastened joint, every sealed surface, every sliding fit, every pressed bearing. Each interface has tolerancing requirements that depend on the function of the joint and the manufacturing processes used to produce each part. This is engineering knowledge that cannot be captured in a text prompt and is not learned by training on geometry datasets.
Real assemblies also have assembly sequence dependencies. Some parts must be installed before others. Some fasteners require access clearance that constrains the order of operations. Seals must be compressed to specific percentages. Shims may be needed to achieve axial preload on bearing stacks. None of this information exists in the geometry alone.
What Would It Take to Solve the Multi-Part Problem?
Solving text-to-CAD for assemblies would require advances on several fronts simultaneously. The AI model would need a fundamentally different architecture than what is used for single-part generation, one that reasons about relationships between parts rather than generating parts independently.
It would need an integrated constraint solver that maintains geometric consistency across all components. When the AI decides the input shaft should be 30mm diameter, every bearing, seal, gear bore, and housing pocket that interfaces with that shaft must update accordingly.
It would need a tolerance analysis engine that assigns appropriate fits based on functional requirements. An AI that generates an assembly where a shaft and bearing have the same nominal diameter has not produced an assembly. It has produced a collision.
It would need domain-specific manufacturing knowledge to ensure that each part can actually be produced. A topology-optimized housing that minimizes weight is useless if it cannot be cast, machined, or printed in the required material with the required surface finishes at the required tolerance.
And it would need access to standard component libraries. Real assemblies use hundreds of off-the-shelf components: bearings, fasteners, seals, retaining rings, dowel pins. These components have specific dimensions defined by standards (ISO, ANSI, DIN). An AI generating an assembly needs to reference these standards, select appropriate components, and integrate them correctly.
None of these capabilities exist in current text-to-CAD offerings. They represent years of development, and some of them may require fundamental architectural innovations that have not yet been proposed.
What Actually Works for Assembly Design in 2026
While text-to-CAD struggles with assemblies, there are AI capabilities that genuinely help engineers working on multi-part designs today.
The most impactful is intelligent part search across your existing design library. When you are designing a new assembly, a huge portion of the work is selecting and integrating components that already exist. Bearings, brackets, housings, shafts, and connectors that were designed and validated for previous products. If you can find the right existing parts quickly, you eliminate the need to design them from scratch, and you avoid the risks that come with untested new designs.
Leo AI enables exactly this workflow. Describe what you need in plain language and Leo searches your organization's full PDM and PLM systems using patented geometry-aware search. It does not generate new parts. It finds existing proven parts that match your requirements, complete with drawings, BOMs, revision history, and manufacturing data.
For assembly design, this is far more valuable than generative AI. When you find an existing validated bearing housing instead of generating a new one, you get a part that already fits the bearings you are using, already has the correct bolt pattern, already passed design review, and already has an established supply chain. You spend your time on the genuinely new design challenges in your assembly, not reinventing components that your organization already perfected.
Leo AI offers integrations with leading PDM and PLM platforms including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM. SOC-2 certified and GDPR compliant, with zero training on customer data, so your IP stays protected.
FAQ
Build Assemblies From Proven Parts
Search your full design history with plain language
Leo AI finds existing validated components from your PDM vault using natural language or uploaded geometry. Build better assemblies faster with parts that already work.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
Build Assemblies From Proven Parts
Search your full design history with plain language
Leo AI finds existing validated components from your PDM vault using natural language or uploaded geometry. Build better assemblies faster with parts that already work.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams