Autodesk is building a new AI architecture for design software that moves beyond simple co-pilots, centering on geometry-aware foundation models and a persistent knowledge layer the company calls "project intelligence." The strategy, outlined by CEO Andrew Anagnost and SVP of research Mike Haley, signals a fundamental shift in how engineering data is captured, retained, and reused across a product's lifecycle - with implications for platform dependency and intellectual property control.
The project intelligence layer
Anagnost's core argument is that too much engineering knowledge evaporates as work moves between people, applications, and development stages. The CAD model records what was designed, but not why a tolerance changed or which manufacturing constraint drove the final geometry. Project intelligence aims to capture those decisions, assumptions, and lessons and keep them connected to the product permanently.
"AI potentially makes this surrounding engineering knowledge far more valuable, because machines can retrieve and reason over it rather than leaving it buried in documents, meetings and engineers' memories," Anagnost said. The vision applies across AEC and product development alike. AI should bridge the gap between an idea and a fully developed model, carrying design intent into simulation, engineering, and manufacturing without replacing designers.
Neural CAD and geometry-aware AI
At Autodesk University 2025, the company unveiled Neural CAD, a foundation model trained on professional CAD data that reasons directly about geometry, topology, and engineering relationships. Most current AI tools in CAD use an LLM to interpret instructions and call APIs in an existing modeling system. Neural CAD places geometry understanding inside the model itself.
The technology generates editable boundary representation (B-rep) CAD geometry from text prompts, sketches, images, and voice commands. Autodesk positions it as complementary to parametric modeling, not a replacement. Parametric CAD handles precision and deterministic control; Neural CAD supports conceptual exploration and more natural interaction. Mike Haley described it as "the first fundamental change in CAD interaction for more than four decades."
Haley confirmed that Neural CAD can produce multiple design alternatives simultaneously and, for some tasks, generate a parametric construction sequence. Future interfaces will combine prompts, sketches, reference documents, images, and voice commands. Organizations may eventually fine-tune Autodesk's foundation models on their own historical data, creating company-specific AI that reflects proprietary engineering expertise. This aligns with broader trends in Generative AI and LLM adoption across technical industries.
The lock-in question
If project intelligence becomes the repository for CAD models, design rules, workflows, and ontologies, it represents a deeper form of platform dependency than proprietary file formats ever did. The industry spent two decades trying to escape those formats. AI raises the prospect of proprietary knowledge lock-in, where intellectual property shifts from manually created design files to the accumulated knowledge surrounding them.
Haley said some models use customer-authored design data in aggregated, de-identified form, with opt-outs available for advanced AI features. Autodesk employs transparency cards and anti-parroting checks to manage trust and IP risk. Still, many firms will hesitate to place commercially sensitive information - client briefs, performance requirements, design intent - inside a vendor-managed platform, regardless of safeguards.
Economics add another layer of concern. Companies already pay for authoring software, cloud collaboration, and data storage. AI compute and project intelligence services create another recurring cost tier.
Why this matters for product development professionals
Neural CAD is not a magic button for perfect geometry. Haley acknowledged it cannot yet guarantee precision. Autodesk's approach pairs probabilistic AI generation with deterministic parametric engines that heal, constrain, and bring geometry into tolerance. For now, the technology is best suited for concepting and bounded design tasks where AI generates options and engineers remain in control. The context window problem - large projects are semantically dense, not just large - demands compact representations and retrieval strategies that align with the project intelligence framework. The next few years will determine whether AI for Product Development consolidates around managed cloud services or whether firms insist on running AI on their own infrastructure, retaining ownership of the engineering intelligence that powers it.
Your membership also unlocks:
