Nashville, TN — If the first keynote at Autodesk University 2025 was about unveiling the company’s three industry clouds, the second was about something timelier: how AI itself is rewriting the rules of design and engineering software.
To a packed ballroom, Raji Arasu, Autodesk’s Chief Technology Officer and EVP of Platform Services, took the stage to declare that “AI breakthroughs are happening in months, not years.” She framed this as both exhilarating and unnerving—and made clear that Autodesk intends not just to keep up, but to lead.
“Our approach is the same as last year,” Raji said. “Supercharge your ingenuity with AI, protect your intellectual property, and act with transparency and trust.”
Building Trust: Standards and Partnerships
Raji announced that Autodesk is among the first companies globally to achieve ISO/IEC 42001, the world’s first international standard for AI management systems. Unlike a check-box compliance scheme, she said, it requires “rigorous risk controls, clear accountability, and continuous monitoring.”
Raji highlighted Autodesk’s practice of co-developing AI tools directly with customers, citing work with Arcadis, the Dutch design and engineering giant. Arcadis is piloting AI that can act like “X-ray vision” for buildings, letting teams see inside walls without demolition. The system enables faster retrofits of existing stock—an urgent sustainability challenge as the built environment is set to double in the next 40 years.
Arcadis executives on a video said the technology lets one person complete in hours what once took teams days. “We’ve gone from being adopters of Autodesk technology to co-innovators,” said a Arcadis executive.
From Point-and-Click to “Describe-and-Do”
Raji traced the history of computing:
- 1940s: Software welded to hardware.
- 1980s: APIs break software out of silos.
- 2000s: Cloud and platforms transform collaboration.
- Today: AI shifts software from executing instructions to reasoning and generating outcomes.
What accelerates this shift, she argued, is the Model Context Protocol (MCP), introduced industry-wide in late 2024. MCP is what allows AI agents to access real-time data and act on APIs—turning them into true “agentic” systems.
“Your intent becomes action,” Raji said. “That’s the future: point-and-click becomes describe-and-do.”
The Capacity Gap: Interoperability and Automation
Raji returned to CEO Andrew Anagnost’s earlier theme: industries face a capacity gap—demand rising faster than available labor. Training alone won’t close it; only AI-driven automation can. She outlined Autodesk’s two strategies:
- Interoperability
Autodesk is tackling the friction caused by groups or software not talking to each other, which experts say wastes six hours per week per person.
Three initiatives are central:
- Granular Data Models: Replacing brittle file-based workflows with semantic, editable data. Already visible in Revit’s Publish features and new manufacturing BOM APIs.
- Data Exchange Connectors: Letting users selectively share only the data they want (stair walls, curtain walls) with partners. Updates sync automatically, no re-exports, no email chains. Connector use has grown 5x in the past year.
- Native Geometry Mapping: Converting geometry seamlessly between design tools. A new partnership with BIMdex now lets a pipe in Plant 3D map directly into Revit as an editable pipe. Companies are saving 70% of time previously spent remapping geometry.
- Automation
Automation, Raji stressed, is not about job loss, but “power—power to focus on what matters.” APIs are the pipes underneath; Autodesk recorded 15.4 billion API calls in the past year, up 43%.
New usage-based pricing is coming: core APIs remain bundled with subscriptions up to generous limits, but heavy automated workflows will incur extra charges.
Meet the New Autodesk Assistant
The biggest reveal was the reimagined Autodesk Assistant. No longer just a contextual chatbot, it is now positioned as a full agentic AI partner that can automate tasks, connect disconnected workflows, and deliver real-time insights.
“It will be the high-performing teammate you’ll want on your side,” Raji said.
Unlike standalone assistants, Autodesk’s is built to call other MCP-enabled agents, and vice versa, allowing cross-company collaboration. It will be available both embedded in Autodesk products and as a standalone agent.
Assistant in Action: IT, Compliance, Collaboration
The keynote then shifted to live demos—Raji narrating scenarios where Assistant converts natural language intent into action across industries.
How it Works — A Construction Play
For purposes of demonstration, we are offered a AEC passion play off sorts, featuring a hospital of “Kaiser Permante size.” A “Flamestone Design Group” has an IT manager who asks Assistant to set up single sign-on (SSO). Assistant collected the domain and metadata, configured SSO, and then assigned Revit seats to three employees—all in minutes.
Later, the manager asked for a software usage report. Assistant surfaced seat utilization data, recommending reallocation and flexible licensing for low-use accounts. A task that once took hours of manual reporting took only seconds.
At “Best Value Construction,” general contractor Luka needed to share stair and railing designs with a manufacturer and cost estimator—but without handing over the full hospital model. Assistant used data exchange connectors to isolate and share only relevant geometry.
Later, Luka ran a compliance check: Assistant scanned every door, ramp, and parking space against code. Noncompliant items were flagged and automatically assigned to owners. Architect Sam in Norway received a notification in Revit, changed the door type, and closed the issue—all tracked end-to-end by Assistant. The cast also includes Andrei, an engineer who designs balconies, Janice, a building designer and Maria, a fabricator.
What would have taken weeks of manual review and emails on this project was done in prompts.
Bringing it home (to Nashville), three small businesses—Andrei and the balcony supplier, Janice the hotel designer, and Maria the fabricator—collaborated seamlessly.
Andrei generated BIM-ready balcony components in Inventor, made discoverable in Revit. Janice found his products via Assistant, verified specs, and inserted them into her hotel design. Maria received the exact balcony specs automatically checked against fabrication rules, then sent them to production.
“Three businesses, three disciplines, connected in minutes instead of weeks,” Raji said. It is a stroke of triumph.
Toward Neural CAD
If Raji laid the foundation of the present and short-term future, the handoff to Mike Haley, Autodesk’s SVP of Research and Generative AI, was about the long-term future: reinventing CAD itself.
Mike reminded us that at AU 2018, Autodesk staged a mock-up demo of AI-driven design via natural language and sketches. At the time, it was aspirational. Today, he said, Autodesk’s AI Lab—publisher of nearly 100 peer-reviewed papers—has built the beginnings of Neural CAD.
Unlike parametric CAD, unchanged for 40 years, neural CAD engines leverage deep learning to reason about geometry, physics, and systems.
Mike explained that while LLMs and image generators reason in text and 2D space, design industries need models that understand 3D geometry, physical behavior, and systems-level logic. Autodesk’s research projects—like Project Bernini (AI for 3D conceptual design)—are steps toward “professional-grade foundation models for CAD.”
Mike demonstrated a power drill generated from a text prompt. Neural CAD output not just surfaces but first-class editable geometry in Fusion. Even details like the handle could be refined via sketch input, and the model carried a history of Fusion commands, making it fully parametric-editable.
Building Systems Demo
For architecture, neural CAD auto-generated floor plans and structural systems in response to massing model edits. Grid lines and cores updated as the building shape changed. Constraints could be locked, while prompts like “switch to timber structure” recomputed column layouts.
“This isn’t a floor plan generator,” Mike said. “It’s a systems-level foundation model reasoning about buildings.”
Research Experiments: Think Aloud and Rubber Ducks
Mike previewed experiments like Project Think Aloud, where designers speak their thought process while sketching. The system uses multimodal AI to both capture design intent and suggest actions.
One playful variant involved a digital “rubber duck” that talked back—sometimes encouraging, sometimes critical—helping designers reflect in real time.
Other demos included automatic naming of unlabeled bodies in Fusion assemblies, based on shape and context, saving hours of manual cleanup.
The Road Ahead
Mike stressed that many of these features are not yet commercialized, but prototypes will flow into products quickly. Fusion’s new Auto Constraint, launched just seven months ago, has already improved significantly.
“Seven years ago we imagined this future,” Mike said. “Now we’re at the doorstep. The advances will come faster than you expect.”
Closing: An AI-Native Autodesk
Raji returned to close, calling this “an AI-native mindset shift.”
- Today: interoperability, automation, Assistant.
- Tomorrow: neural CAD engines.
- Goal: reshape the future of design software as profoundly as the shift from gas to electric cars.
“When we get this right,” Raji said, “you’re not just changing the world. You’re shaping a better one for all of us