Dr. Maor Farid, Co-Founder & CEO at Leo AI
Dec 16, 2025
Converting a hand-drawn sketch or legacy 2D drawing into a usable 3D CAD model sounds straightforward. At least until design intent gets lost in translation. For mechanical engineers and product designers, the real challenge isn't generating geometry. It's ensuring the converted model accurately reflects dimensional relationships, tolerances, and functional constraints embedded in the original drawing.
The AI hype in this space is loud. Really loud. Every other week there's a new text-to-CAD tool promising to "revolutionize" how engineers work. But here's what nobody wants to talk about: most of these tools produce results that can't survive contact with real manufacturing requirements.
I've spent time digging into what actually works versus what makes good LinkedIn posts. The findings weren't what most marketing materials would have you believe.
The Reality Gap Between Promise and Production
When Xometry, a major manufacturing marketplace, tested seven text-to-CAD tools in August 2025, their findings were sobering:
For simple parts (a 20mm cylinder with specified diameters), some tools performed reasonably well.
For medium complexity (a 24-tooth gear with specific parameters), most tools failed to produce accurate models.
For high complexity (a manifold block with internal channels), the tools either produced simplified approximations or couldn't generate results at all.
Their conclusion? These tools "aren't yet a substitute for professional CAD software." The biggest limitations: lack of control, inconsistent file exports, and minimal support for complex assemblies or functional constraints.
This matches what I hear from mechanical engineers on Reddit and engineering forums: "I don't see this being faster than just doing it yourself. A lot of parts are highly complex and require high precision to work with the other parts."
Why Text-to-CAD Tools Struggle With Real Engineering
The core problem isn't that AI is bad at geometry. It's that manufacturing tolerance is unforgiving.
According to MIT researchers working on VideoCAD, current text-to-CAD models operate on visual inputs around 224x224 pixels. They struggle with the sub-millimeter precision required for aerospace or medical devices. The difference between 10.0mm and 10.05mm is the difference between a working part and scrap metal.
The complexity ceiling is also low. Most training datasets focus on single parts with sequences around 186 steps. A real-world engine assembly has thousands of parts and millions of interaction steps. Current text-to-CAD technology simply cannot maintain context over that horizon.
And then there's the software lock problem. Most of these tools generate meshes or simplified geometry that can't integrate with your existing CAD history tree. You get a "dumb solid" with no parameters, no editability. If you need to change the diameter of a hole, you're often stuck manually patching a mesh.
Comparison Table: Sketch-to-3D CAD Tools
Tool | Enterprise Adoption | Output Type | Simple Parts | Complex Parts | PLM Integration | Engineering Context | Production Ready |
Leo AI | HP, Scania, Intel, Mobileye (60K+ engineers) | 3D mesh + full engineering docs | ✅ Excellent | ✅ Excellent | ✅ Yes | ✅ Understands CAD, specs, tolerances | ✅ Yes |
Zoo.dev | None reported | B-rep CAD (STEP) | ✅ Good | ❌ Failed in testing | ❌ No | ❌ Geometry only | ❌ No |
CADScribe | None reported | STEP, STL | ✅ Good | ❌ Failed in testing | ❌ No | ❌ Geometry only | ❌ No |
Tripo AI | None reported | Mesh (gaming/creative) | ⚠️ Visual only | ❌ Not applicable | ❌ No | ❌ Not for engineering | ❌ No |
AdamCAD | None reported | STL, SCAD | ✅ Good | ⚠️ Simplified only | ❌ No | ❌ Geometry only | ❌ No |
What Each Tool Actually Does (Honest Assessment)
1) Leo AI - The Only Tool With Real Enterprise Adoption
Leo AI takes a fundamentally different approach. Instead of trying to generate geometry from text prompts (which traditional CAD tools already do well), Leo works as an engineering copilot that understands context.
Who's actually using it: HP, Scania, Intel, Mobileye, and over 60,000 engineers across enterprise organizations. It's distributed through value-added resellers in the US, UK, India, France, Germany, Poland, Benelux, and Israel.
What it actually does:
Generates 3D mesh concepts from sketches, specs, text descriptions, and requirements in under 50 seconds
Produces complete engineering documentation: BOM, statement of work, manufacturing methods, quality assurance tests
Answers technical questions grounded in your organization's actual data and 1M+ trusted engineering sources
Finds relevant parts from your PLM and 120M+ vendor parts using natural language
Runs engineering calculations with full traceability to trusted sources
The key difference: Leo is built on a proprietary Large Mechanical Model (LMM) trained specifically on mechanical parts, assemblies, and engineering logic. It understands CAD geometry, tolerances, and constraints in ways that generic AI models fundamentally cannot.
Measurable results from real deployments:
34% reduction in design errors
32% increase in part reuse
12 hours saved per engineer per week
The security piece matters: SOC 2 certified, GDPR compliant, with zero training on customer data. For enterprise engineering teams, IP protection isn't optional.
Best for: Enterprise engineering teams who need production-ready AI that integrates with existing workflows.
2) Zoo.Dev (Text-to-CAD)
What it claims: Generate precise B-rep CAD models from text prompts.
What it actually does: Works well for simple mechanical components like flanges, brackets, and plates. The desktop app (Zoo Design Studio) offers parametric editing through sliders, which is genuinely useful for quick iterations.
The reality check: In Xometry's testing, Zoo delivered accurate models for simple parts but "failed to produce an accurate model" for medium-complexity gear requests. For the high-complexity manifold block, "it could not generate a result at all."
Zoo's own documentation admits: "Traditional, simple mechanical parts work best right now: fasteners, bearings, connectors, etc." They're honest about limitations, which I respect, but this means it's a rapid prototyping tool for basic components, not a production workhorse.
Their error rate for KCL generation dropped from 50% to 16% recently. That's progress, but a 16% error rate in manufacturing is still unacceptable for anything mission-critical.
Best for: Quick concept models of simple geometry. Hobbyists and 3D printing enthusiasts.
3) CADScribe
What it claims: Turn natural language into 3D CAD models.
What it actually does: Generates simple geometric shapes. Their own marketing describes it as useful for "presenting product concepts or 3D mockups for marketing materials."
The reality check: Xometry found CADScribe "performed well" for simple cylinder geometry but results "diverged from expectations" for moderate complexity like a gear with specific parameters. For high-complexity designs, "CADScribe could not generate a model."
The input length restrictions cut off key details from prompts, and there's limited error feedback. Functionally, it's a tool for generating simple shapes quickly, not for engineering work.
Best for: Quick marketing mockups. Simple geometric visualization.
4) Tripo AI
What it claims: Generate 3D models from text, images, or sketches.
What it actually does: Generates mesh models primarily aimed at game development, animation, and creative visualization.
The reality check: Tripo is not a CAD tool. It generates mesh models for gaming and creative applications. The outputs are not dimensionally accurate and are "best viewed as concept art rather than functional designs."
Their own documentation and reviews confirm: "approximately 1 in 10 generations are client-ready without manual cleanup" and it "struggles with complex multi-part assemblies or precise engineering tolerances."
This is a 3D asset tool for artists and game developers. If you're trying to manufacture something, this isn't the tool.
Best for: Game assets. Animation. Creative 3D visualization.
5) AdamCAD
What it claims: Text-to-CAD with parametric controls.
What it actually does: In Xometry testing, AdamCAD "handled simple and medium-complexity mechanical parts well." The 24-tooth gear was generated accurately with editable parameters.
The reality check: This is actually one of the better-performing tools in the text-to-CAD category. It bridges both technical and creative workflows with adjustable parametric models. High-complexity manifold blocks were generated "in a simplified form" with basic features.
Still, the limitation remains: it's a concept tool, not production-ready. Creative mode outputs "may need cleanup for 3D printing" and complex designs "may require more detailed CAD work for accuracy."
Best for: Early-stage engineering ideation. Concept exploration with parametric control.
What Engineers Actually Need (And What Delivers)
Here's what the text-to-CAD hype misses entirely: the bottleneck for mechanical engineers isn't creating geometry from scratch. It's everything else.
McKinsey reports that engineers spend 1.8 hours per day searching for information. They're not hunting for how to draw a circle. They're looking for:
That bracket design from 3 years ago that's similar to what they need now
Whether the material they're considering meets the spec requirements
What tolerances were used on the last similar assembly
Why that particular fastener was chosen in an older project
This is where the real productivity gains hide. Not in generating a basic gear model from a text prompt, but in connecting tribal knowledge across an organization.
Leo AI addresses these problems because it was built by mechanical engineers who understood where the actual pain points are. That's why it's the only AI in this space with real enterprise adoption at companies like HP, Intel, and Scania.
The Honest Decision Framework
If you need to quickly visualize a simple concept: AdamCAD or Zoo might save you time for basic geometry like brackets or plates.
If you're a game developer or 3D artist: Tripo and similar mesh generators are legitimate tools for your workflow.
If your engineering team needs production-ready AI: Leo AI is the only tool with proven enterprise adoption, PLM integration, and measurable results. It's not trying to replace your CAD software. It's making everything before and around the CAD work faster and more accurate.
The Bottom Line
Most text-to-CAD tools are impressive demos that don't survive contact with manufacturing reality. They generate meshes for hobbyists and simple prototypes, not production-ready engineering.
The mechanical engineering workflow has genuine AI opportunities, but they're not about typing "make me a gear" and getting a usable STEP file. They're about:
Making tribal knowledge accessible before it walks out the door
Connecting engineers to relevant past designs they didn't know existed
Automating the tedious documentation and calculation work
Reducing errors by applying organizational best practices automatically
Leo AI addresses these problems with a purpose-built Large Mechanical Model and real enterprise traction. The rest of the text-to-CAD market? It's still figuring out how to be genuinely useful for engineers who need to actually manufacture something.
Want to see how AI can actually improve your engineering workflow? Book a demo with Leo and see why 60,000+ engineers are already using it.
Sources
Xometry Pro - "We Tested 7 Text-to-CAD Tools – Are They Actually Useful for Engineers?" (August 2025) https://xometry.pro/en-eu/articles/text-to-cad-tools-test/
Binary Verse AI - "Text To CAD: 1 Powerful MIT AI Fixing 3D Workflows" - Analysis of VideoCAD and precision limitations (December 2025) https://binaryverseai.com/text-to-cad-vs-videocad-mit-generative-design/
Mike Kalil - "Hate It or Love It: Text-to-CAD Technology Proves Divisive" (July 2025) https://mikekalil.com/blog/hate-it-or-love-it-text-to-cad-technology-proves-divisive/
Pulse 2.0 - "Leo AI: $9.7 Million Raised For Transforming Mechanical Engineering" (September 2025) https://pulse2.com/leo-ai-9-7-million-raised-for-transforming-mechanical-engineering/
Capterra - Leo AI Product Profile and Enterprise Adoption Data https://www.capterra.com/p/10032244/Leo-AI/
Zoo.dev - Text-to-CAD Documentation and Limitations https://zoo.dev/docs/zoo-design-studio/text-to-cad
Zoo.dev - "What's New With Zoo, August Edition" - Error rate improvements from 50% to 16% https://zoo.dev/blog/whats-new-august
Tripo AI Review - "Tripo AI Review 2025: Is It Really A Game-Changer For 3D Modeling?" (October 2025) https://clixsensesuccess.com/tripo-ai-review/
CB Insights - Leo AI Company Profile https://www.cbinsights.com/company/leo-ai
DesignRush - "I Tried 5 Text to CAD AI Tools for Faster Drafting: Here's What Worked" (May 2025)https://www.designrush.com/agency/ai-companies/trends/text-to-cad-ai
