For nearly 40 years, parametric CAD engines have remained fundamentally the same. The technology itself is accessible through menus, GUI-based tools, and human mouse movement and data inputs. As CAD users, we all fundamentally understand this.
However, there is one significant problem with very mature parametric CAD engines, and that is they were not designed for the AI era. They only understand computer-human input from the pre-AI era and cannot comprehend new types of inputs associated with AI and soon-to-be Agentic AI. Furthermore, parametric CAD limits the user to a preset degree of detail (from conceptual to highly detailed and dimensional) and cannot easily transition between these levels.
The first AEC-related AI foundation model sits in Forma Building Design today. The neural CAD engine enables users to rapidly move from design concept models to detailed building layouts and structural systems quickly and to iterate. (click for larger view)
Imagine if a new tool enabled rapid and iterative workflows that allowed users to explore and test ideas quickly at scale. It could then scale up these iterative tests from conceptual, rough, and approximate forms to highly detailed, dimensional versions, all within the same tool. And imagine if users could input a generative CAD system with voice-text commands, hand sketches, and images simultaneously?
Neural CAD Engines
All of this is possible, but only with neural CAD engines. Autodesk announced the existence of two neural CAD engines, one with a geometry focus aimed at the D&M workflow with Autodesk Fusion and the other with a building logic, structural, and spatial focus aimed at the AEC workflow with Autodesk Forma Building Design.
Now Autodesk asks us to imagine the following scenario:
Picture a scenario where your computer comprehends spoken language, sketches, three-dimensional design data, and industry-specific workflows. Now, enrich this scenario with decades of your team’s project knowledge.
Autodesk demonstrated the early workings of such a scenario. Using voice for input, Autodesk’s new AI foundation models (neural CAD engines) were able to ingest spoken language instruction and free-hand sketch input for the design of a chair, also pairing instructions that came in the form of a picture of a chair that had aspects the user wanted to see in their neural CAD engine-generated (generative AI) based design.
Beyond LLMs
This is moving far beyond just pairing large language models (LLMs) with existing solutions. Autodesk’s approach for the next level of AI innovation is a complete re-architecting of the traditional CAD engines behind tools like Forma, Revit, Fusion, Civil3D, AutoCAD, and more.
These new neural CAD engines are capable of creating CAD geometry from a multi-input system-level approach. At the same time, unlike the 3D graphics that today’s AI systems, such as ChatGPT, can produce, Autodesk’s new neural CAD engines generate CAD geometry that is fully editable using traditional parametric CAD inputs as well.
In terms of architecture, Autodesk’s new neural CAD AI foundation models (aka: neural CAD engines) currently consist of one dedicated to the future of Forma. It is already forming the basis of Forma Building Design, which is why the system can allow text-box instructions to generatively produce alternative building design layouts.
This view of the geometry-oriented AI Foundational Model (neural CAD engine) can create designs spontaneously from a text prompt. It is an entirely new machine learning approach to generating CAD objects, in contrast to classical parametric CAD engines that have existed for over 40 years. (click for larger view)
Forma Building Design can rapidly produce alternative designs based on changing the structural grid, selecting alternative structural systems, and more. You can ask Forma Building Design to produce a single-loaded corridor design and a double-loaded corridor design, and then the software can let you compare them side-by-side.
But in speaking to the folks at the geometry-oriented neural CAD engine about architecture—and in describing the process whereby architects today use physical trace paper in a recursive iterative process during design at both the plan, elevation, and 3D level—the importance of sketch input (something Autodesk demonstrated) with the geometry-0riented neural CAD engine seemed relevant to architecture in their eyes as well as my own. In other words, I was able to describe and somewhat convince the neural CAD guys that the AI foundation model for Fusion has capabilities that are useful for architects.
These two AI foundation models will likely share common aspects over time. The Fusion-oriented model can produce shapes that Fusion can make now. The Forma Building Design-oriented model seemed to only demonstrate box-shaped architecture. Granted, the vast majority of buildings are rectangular in nature and do not feature curved surfaced bodies or facades, but the common work of “maquee design firms” does.
Reasoning with CAD Geometry
Autodesk AI researchers have been working to teach their AI foundation models to reason with CAD geometry, systems design, and the real world. This, of course, complements their ability to reason with LLMs.
The future of these AI foundation models lies in their availability to customers, who can personalize them by tuning them to their organization’s proprietary data and processes. That is the big picture view of the future of Autodesk’s neural CAD engine models, and it offers an enticing view of the future of design.
Architosh Analysis and Commentary
While Autodesk has shown its hand with its generative AI technologies, the software leader in AEC isn’t entirely in its own space with this capability. There were third-party Autodesk partners at AU25 that had similar or related generative AI technologies, including ones that could do sophisticated automation of multi-step operations inside of tools like Revit. Outside of the Autodesk world, there are competitors that are also offering similar generative AI functions. (see: Architosh, “Architosh 11th ‘BEST of SHOW’ honors for digital technologies at AIA25 Boston,” 2 Jul 2025)
However, what is perhaps most unique about Autodesk’s neural CAD engines is the demonstration of multi-modal input methods, ranging from speech to hand sketches to images to slider user-interface widgets, and the various combinations that are possible. This, combined with the company’s intention of enabling organizations to customize their AI foundational models with their own proprietary data and operate on granular data across the entire Forma cloud platform, is what sets Autodesk apart within its own ecosystem. It’s the vision of the total lifecycle of a building and its data, and where AI technologies can actively participate in various workflow processes.