Autodesk outlines connected AI Assistant for task execution

AI-generated summary

Autodesk is proposing a shift from software tools toward AI-executed outcomes, with a standalone Autodesk Assistant intended to coordinate work across products, disciplines and projects. Presented at Autodesk University 2026 under the “Project Intelligence” vision, the agent-first Assistant would translate user intent into results for review and refinement, reducing the need to navigate applications and workflows. It remains in development, with launch expected the following year, and will complement—not replace—the Assistants embedded in Revit, AutoCAD, Fusion and other products. The foundation is Autodesk Context, intended to capture not only project changes but why decisions were made, under which constraints, and with whose involvement. Autodesk argues that engineering AI needs geometry, design intent, workflow state and project history rather than text alone. Assistant Builder, currently in preview, lets organisations connect their own and partners’ AI systems. The underlying architecture combines Model Context Protocol (MCP) connections, agents and Autodesk Platform Services (APS). An AI Orchestrator selects among Autodesk models, Claude, OpenAI and open-weight models according to accuracy, speed, security and cost. Autodesk claims specialised smaller models can deliver better task-specific accuracy and substantially lower execution costs. Its “neural CAD” technology comprises multiple generative models already used in Fusion, Revit and Forma; consolidating their capabilities into a broader model is a longer-term research goal. An onstage demonstration illustrated the proposed cross-disciplinary workflow. Assistant prioritised project correspondence and issues, assessed schedule consequences of moving a residential wall three feet, and generated a Gantt chart highlighting risks. Building Layout Explorer then generated an LoD 300 apartment and adapted room layouts to differently shaped units, presented as minutes rather than a day of work. Another workflow extracted facade dimensions from a sketch, generated bracket requirements, and passed work to Fusion, where an archived bracket was adapted and more than 1,000 fascia options were produced before returning results to the architect. These were demonstrations of intended capabilities, not evidence of general production readiness. Autodesk also showed Project Quill, combining iterative sketching and speech to create 3D designs, and discussed future task-specific interfaces generated on demand. Pricing remains unresolved. Autodesk envisages averaging costs across simple and complex tasks to provide predictable charges. Routine AI should be included in subscriptions, while expensive, high-value operations could carry separate charges and rate tables. However, detailed visibility into individual users’ token spending was not confirmed. The article contrasts this managed approach with increasingly practical local AI alternatives, enabled by official, third-party or custom MCP servers connecting models to Autodesk applications. Local demonstrations addressed cost and data control. HP’s ZBook Ultra G3a paired Perplexity orchestration and a post-trained Qwen model with the official Revit 2027 MCP server. Its 192 GB of system memory, including up to 160 GB allocatable to the integrated GPU, supported the CAD and AI stack on one machine. Separately, Greg Schleusner demonstrated local workflows for Revit project setup, real-time clash detection and checking room contents against requirements. Nvidia used a Dell Precision laptop connected to a Dell Pro Max with GB10 running Qwen3.6-27B. Its basic Revit modelling example was a proof of concept, while a Fusion demonstration used Nous Research’s Hermes harness to coordinate specialist agents for modelling, dimensional checks, rendering, engineering drawings and a parts list, with feedback loops and project-manager review. Lenovo demonstrated a more accessible application: an agentic interface simplifying Civil 3D cut-and-fill calculations on a ThinkStation Ultra with a 16 GB Intel Arc Pro B50. Agents could also be distributed across networked workstations. These examples suggest alternatives to escalating cloud-token costs, although hardware requirements and workflow economics remain uncertain. Local open-weight models also offer opportunities for company-specific tuning and keeping intellectual property on-site. Licensing is another unresolved issue. Autodesk confirmed that multiple Revit instances may run on one workstation if properly licensed and users are signed in; its public MCP supports internal business workflows, while commercial developers can use APS. The article cautions that third-party arrangements are less clear and future terms could change if automation reduces seat demand or bypasses usage-based cloud services. Its conclusion is that Autodesk’s strongest differentiation may be connected project context rather than geometry generation alone, but customers’ willingness to adopt that platform will depend on cost, control and competing in-house options. Separately, the source introduces forthcoming Fusion automation, including AutoTimeline, which converts imported geometry into editable parametric parts.