Skip to main content
Roviral

Autodesk’s AI Strategy Puts Project Context and Local Workflows in Focus

Autodesk is building an agent-led Assistant around project context, while workstation demos show how firms can run design AI locally and manage cost and data.

Autodesk’s AI Strategy Puts Project Context and Local Workflows in Focus

Autodesk’s AI plans reach beyond adding faster commands to Revit, Fusion or AutoCAD. At Autodesk University 2026, the company described a system intended to interpret a project task, choose the tools and models it needs, and return work for people to review. At the same event, workstation demos showed a different path: running models close to the design data, with firms retaining more control over processing and cost.

Autodesk CEO Andrew Anagnost called the broader idea “Project Intelligence”: a connected system that brings project data and decisions together and builds context as work advances. The practical test will be whether that context makes cross product tasks easier without making costs, model choices or data handling harder to understand.

A standalone Assistant for work across products

The main announcement was a standalone, agent first version of Autodesk Assistant, which Anagnost said would reach across Autodesk products. Mike Haley, senior vice president of research, described the need for work that does not fit neatly inside a single product such as Revit, AutoCAD or Fusion. The software remains in development and is expected to launch next year.

The standalone version is meant to sit alongside the assistants embedded in individual applications. Amy Bunszel, executive vice president of AEC solutions, said Autodesk will continue to expand those in product tools. The distinction is that a task spanning applications or projects could begin in one place, with the Assistant selecting relevant capabilities from Autodesk’s wider software environment.

That points to a bigger role than the word “Assistant” suggests. The envisioned system could start agents, reach across products, disciplines, datasets and organisations, and generate an interface for a particular task. Autodesk is positioning it as a route into its Forma, Fusion and Flow industry clouds, with the user describing an outcome instead of first learning where each command or workflow lives.

Bunszel said the intended users include people following what Autodesk internally calls “critical user journeys”: tasks customers need to complete across products and projects. That could span experienced specialists who know the tools well and people who need to complete a task without knowing which application or workflow is appropriate.

Under the interface, the system would draw on a mix of components. Bunszel described Model Context Protocol connections, Autodesk built agents, customer supplied agents and cloud services available through Autodesk Platform Services. The mix is intended to let a task call on the right capability across systems, although the demo and product vision do not yet establish how broadly that will work in day to day production.

Project context is the differentiator

Autodesk’s case for a domain specific approach is that design work cannot be represented by text alone. Geometry, engineering intent, project history and relationships between decisions all matter. Haley argued that large language models cannot solve every design problem by themselves; the system needs to understand the underlying project information as well as the request.

CTO Raji Arasu gave that idea a name: Autodesk Context. She described it as information built from the decisions organisations make, rather than information pulled only from a prompt or one data store. The goal is to capture why a project changed, not simply what changed.

Arasu used a familiar example: after a project is complete, a team may see that a wall moved but no longer know why. The reasoning may have depended on constraints, data, people and workflow conditions that were understood at the time but never preserved in a form others can retrieve. Autodesk Context is intended to connect those elements, including the state of the workflow when a decision was made.

That record of relationships could be more useful to an agent than a pile of files. If the system can find the reason behind a design choice, it may be able to judge whether that choice still applies when requirements change. Autodesk is betting that this project history is harder for a general purpose model to reproduce than a single geometry generation task.

Projects also cross systems beyond Autodesk’s own products. Assistant Builder, currently in preview, is designed to let organisations connect their own AI systems and those of their partners to Assistant. That gives firms a route to incorporate existing tools, though the article’s demonstrations did not establish the limits of those connections or the work needed to maintain them.

Autodesk AI Orchestrator is another part of the proposed stack. Arasu said it selects models based on factors including accuracy, speed, security and cost. The system may use one model or several, depending on the task. Its pool includes Autodesk models, Claude, OpenAI models and open weight models.

That choice matters because design tasks can involve geometry, schedules, materials and simulation, each with different requirements. Arasu said customers should not need to benchmark every model against every workflow themselves; Autodesk intends to handle that selection. In practice, that makes transparency important: users will need to understand what model handled a task and what its use costs, especially when work spans multiple agents.

Autodesk
Autodesk CTO Raji Arasu

Smaller models and Neural CAD

Anagnost argued that large general models can be an expensive fit for tightly defined engineering tasks. He said they consume many tokens when a problem has to be described in a constrained way. Autodesk’s own models, which he said once counted as large, are now medium or small by comparison. The orchestration system is designed to direct suitable work to these narrower models, which he described as costing pennies relative to dollars for some uses.

His argument is that a model trained for a narrower task can be faster and more accurate on that task. Autodesk is also bringing in an open weight model that, Anagnost said, performs as well as the best Claude model on certain problems. These are company claims about model performance and economics; the event did not provide benchmark results or a price schedule for the workflows described.

Neural CAD is Autodesk’s umbrella term for its generative models related to geometry, design, manufacturing and architecture, engineering and construction. Tonya Custis, senior director of AI research, said current applications such as Fusion, Revit and Forma generally use one Neural CAD model alongside other models.

The longer term research aim is to train one model across those tasks. Custis said a model learning multiple tasks together could improve across them. That remains a development goal, rather than a description of a single model already deployed throughout Autodesk’s products.

Sketch, speech and task specific interfaces

Autodesk’s research team also demonstrated Project Quill, which combines sketching, speech and AI. A rough drawing can express a shape that is awkward to describe, while speech can specify an attribute that would take time to draw. In the demo, a user could circle a feature and request a change such as making a door larger or adding solar panels, then iterate on the resulting 3D design repeatedly.

The interaction is significant because it treats sketch and speech as complementary inputs rather than expecting one prompt to define a finished design. Autodesk presented it as a way to reach an intended result more directly; the demonstration did not establish a release schedule.

The company is also exploring interfaces that change with the task. Michelangelo Capraro, a distinguished experience architect at Autodesk, described asking Assistant to generate a user interface for a specific operation in Revit. If the system can produce a focused set of controls on demand, users may spend less time navigating a fixed interface. The open question is how those generated controls will handle complex settings, validation and review.

What the Assistant demo showed

Autodesk’s stage demo began with a project manager behind on emails, requests for information and issues. Assistant gathered high priority items from multiple projects using agents. In a residential project, a client wanted more rentable area and requested that a wall move three feet. The project manager asked about schedule effects and downstream impacts; Assistant returned a Gantt chart identifying schedule hotspots.

The next task moved from analysis to geometry. An architect asked for a Level of Development 300 apartment with walls and doors. Assistant called an industry foundation model trained to understand and generate building geometry, then used Building Layout Explorer to stream a 3D model. The demo showed room layouts fitted into units of different shapes and sizes. Autodesk framed this as minutes of work instead of a day.

An agentic workflow in Assistant Builder then took dimensions from an attached image and ran a slatted facade workflow. Assistant interpreted a sketch and generated metal brackets for wooden slats. The workflow crossed into Fusion: a design engineer received the specification, found a similar bracket in an archive, updated it, generated more than a thousand fascia options from images and sent the results back to the architect.

The through line was that users specified the result and the system assembled the steps. That is the central ambition behind Assistant: reduce the need to search for applications and capabilities before starting a task. The demo presented a connected workflow, but it was a staged example; firms will need to assess how well the system handles their own data, design rules and review requirements.

Cost visibility and local processing

Agent workflows can use models rapidly, particularly when one agent starts several subagents. Nvidia’s Sean Young said this can drive cloud token use beyond what a firm’s IT budget supports. Anagnost described Autodesk’s plan to average costs across customers, with some simple tasks costing more than their use and complex tasks costing less, so Autodesk can offer the service at a sustainable margin.

Predictable average pricing does not necessarily show a firm who used a costly feature or how much a particular workflow consumed. Bunszel said a rate table for expensive features was “very realistic” and suggested customers may receive enough detail to understand what happens when they use them. Routine AI would generally be included in a subscription, while charges could apply to features that are highly valuable and expensive to run. Specific rates and reporting details were not provided.

Local processing is one way firms are weighing cost and data control. Kevin Philips of Symetri said hardware is becoming attractive to AEC firms. He estimated that the top 5% of developers could spend as much as $10k a month on tokens, while some practices do not permit use of frontier models. Bunszel said Autodesk is comfortable with firms running open weight models on their own systems and combining locally handled work with Autodesk services where useful.

Model Context Protocol, or MCP, provides a connection between language models and software. Autodesk has built MCP servers for Revit, Fusion and other products, while third parties have built their own. Young said developers can also create a server for an application that exposes an API. That gives firms more choice in how to assemble workflows and where the model processing happens.

Three workstation approaches

HP and Perplexity showed a local setup on the ZBook Ultra G3a mobile workstation. Its AMD Ryzen AI Max Pro processor combines the CPU and GPU in a system on a chip, and the machine has 192 GB of system memory, with up to 160 GB allocatable to the integrated GPU. In the demo, Perplexity’s own language model, a post trained version of Qwen, drove the Revit 2027 MCP server on the same machine. Perplexity’s Daniela Miranda-Smith said local model use makes electricity the main operating cost.

Greg Schleusner, formerly director of design technology at HOK, has been testing the same workstation for production workflows. His work goes beyond a single Revit connection: he described running large models or several smaller models together on the device. Examples included an LLM that sets up Revit projects, machine learning for real time clash detection, and an image model that checks room contents against requirements. Schleusner said local hardware matters when a firm wants to keep data from leaving its own environment.

Dell and Nvidia demonstrated a different arrangement for Revit and Fusion. A Dell Precision laptop ran the CAD applications and custom MCP servers, while a Dell Pro Max with GB10 ran Qwen3.6-27B. The compact desktop is built around Nvidia’s GB10 Grace Blackwell superchip, combining an Arm based Grace CPU and Blackwell GPU on one piece of silicon.

The Revit demonstration used a third party MCP to create a basic building from a prompt. It was a proof of concept, but it showed a possible advantage of open weight models: firms can fine tune them on their own Revit data. The Fusion demonstration was more developed, using multiple agents and feedback loops to create a model, check it and produce supporting outputs.

The Fusion workflow used the Hermes harness from Nous Research, which supports MCP and can connect to different models. In this case, it connected to the local model on the GB10. A project manager agent interpreted a design brief and two concept images, then assigned work to specialist agents. A CAD agent built and organised the model in Fusion’s Design workspace, added initial materials and properties, checked dimensions against requirements, and passed the result to a visualisation agent.

The visualisation agent set camera, lighting and render properties, compared the result with the concept images, and adjusted materials and scene settings in a loop. After review, a drawing agent created 2D engineering drawings and a parts list, checking views, dimensions, annotations and list completeness. The project manager agent then reviewed the outputs.

Young described a related approach in which an agent reads product requirements, creates geometry and runs two types of finite element analysis at once. If a simulation finds a problem, the agent can return to the geometry and try again. These examples point to a workflow where agents verify and revise one another’s work, though a demonstration does not establish how reliably such loops perform on varied production designs.

Lenovo showed a lower hardware barrier. Its custom agent interface for AutoCAD Civil 3D performed cut and fill volume calculations on a ThinkStation Ultra with a 16 GB Intel Arc Pro B50, a class of graphics card found in some existing workstations. Lenovo said agents could sit on different workstations across a network and be called by users as needed.

Nvidia’s Autodesk Fusion agentic demo

Licensing and the business model

Local agents raise questions for Autodesk’s cloud services and product licensing. Autodesk Platform Services includes a Revit Automation API that runs the Revit engine in the cloud to create, extract and modify Revit data without a local installation. If a firm runs comparable work locally, it may use less of that cloud service. If agents take on modelling, querying and documentation at scale, firms could also need fewer licensed seats.

Young described agents controlling multiple applications at once through subagents, with MCP connections and no need to interact with each application’s user interface. In some cases, the agent would make API calls without viewing a viewport. That could change how much work is performed through conventional desktop interactions.

Autodesk said customers can run multiple Revit instances on one workstation, provided each instance is properly licensed and users are signed in. The company also said its public MCP lets customers build agent workflows for internal business use; developers building commercial applications can use Autodesk Platform Services. These statements give firms a basis for planning internal automation, while workflows using third party MCP servers may need separate review against applicable terms.

The commercial details matter because the technology could alter both cloud usage and the number of seats required. Firms planning local workflows will need to account for hardware, model performance and licensing alongside the direct cost of cloud calls. The terms described at the event reflect Autodesk’s current position; firms should track how those terms apply as their workflows and products change.

Fusion’s expanding AI toolset

Autodesk also outlined several AI capabilities planned for Fusion, aimed at repeated design and manufacturing tasks. AutoTimeline is a feature recognition tool that turns imported geometry into editable parametric parts. Richard Stubley described the intent as rebuilding a model’s “parametric DNA”: inferring the construction steps so users can make structured changes instead of treating the import as fixed geometry. That could help when working with supplier models or migrating data from another CAD system.

AutoAssemble is intended to move imported components into workable assemblies. It recognises parts, predicts their positions and creates appropriate assembly relationships. Assistant is also planned to help generate hundreds of product configurations from conversational requirements by building the configuration logic needed to produce valid variants.

For reuse, Assistant will search company design libraries using geometry and metadata, find released alternatives, surface active change orders and help replace components where an existing design is suitable. Find Similar adds AI based part discovery beyond part number lookup or basic geometry comparison, helping teams locate comparable components already in their available design data.

AI generated CAM programming is also planned. Autodesk’s stated aim is to help automate machine setup, toolpath generation and production preparation, so teams can reuse manufacturing workflows rather than repeat setup work job by job. The scope and availability of these functions were not detailed.

Inventor is set to become the first Fusion Connected Client. The connection is intended to link desktop engineering work in Inventor with Fusion’s cloud data and processes across the product lifecycle. The announcement signals a closer relationship between the desktop application and the cloud environment, though the event did not specify how the connection will affect existing workflows or deployment choices.

What firms should watch

Autodesk’s strongest argument is project context: preserving the reasons behind design decisions and connecting them to geometry, people and workflow history. Generating a model from a prompt is becoming a visible capability across different systems. A project record that helps an agent understand why a design exists could be harder to replicate, and potentially more useful when a change affects several disciplines.

At the same time, local demonstrations showed that firms can assemble capable workflows from open weight models, MCP connections and workstation hardware. Some setups keep model processing near the design data and can be tuned to a company’s own work. Others, such as Lenovo’s Civil 3D example, run on hardware firms may already have. Each approach brings tradeoffs in model quality, maintenance, cost and integration.

The standalone Assistant is expected next year, leaving firms time to assess where cross product automation could help and how they want to manage model use. Cost reporting, data boundaries, review processes and licensing will matter as much as the generated geometry. Autodesk is making a bet that customers will value a system that assembles project context and tools around an outcome; local AI shows they also want choices over where that work runs and what it costs.

Source: Greg Corke · develop3d.com

Filed alongside