How much of the new AEC computing stack does Autodesk intend to own, and how many meters does it intend to attach to it?
Autodesk’s vision of agentic AI isn’t difficult to understand. Forma, Assistant, AI Orchestrator, Model Context Protocol (MCP), Context and Autodesk’s AI models point towards a future where users specify what they want and agents work out how to do it.
This arrives just as AI threatens the economics that have served Autodesk so well. For decades, more people doing more design and documentation generally meant more application seats. If one architect can supervise agents performing work previously undertaken by several Revit users, that relationship starts to break down.
Autodesk University 2026 offered some clues as to what comes next. Autodesk isn’t giving up on the seat. It is building orchestration and consumption above it – the implications of which extend beyond which applications Autodesk’s AI agents will use.
Autodesk is already introducing charges for machine-driven compute and data access, while its position on licensing existing Revit automation remains unclear. The question is whether AI will reduce the cost of producing buildings or simply introduce more ways for Autodesk to charge for producing them.
Autodesk reveals the destination
For years, Autodesk’s messaging around Forma has zigzagged between replacement, complement and connected future. At AU 26 it removed much of the ambiguity: over time, design, analysis, coordination and documentation will move into Forma, with Revit becoming a Forma Connected Client.
At AU, Autodesk announced the destination while Forma remains some distance from being capable of replacing Revit. In an ‘Ask me anything’ session with customers, Autodesk CEO Andrew Anagnost was asked how long the Revit to Forma transition would take. According to our notes from the session, he initially said that depended on customers and that Revit would be around for as long as they wanted it, then suggested Forma could reach comparable architectural capability in about two years.
Autodesk has good reasons for announcing its plans early. Large AEC firms make technology decisions years in advance, and giving customers visibility of Autodesk’s direction is preferable to surprising them later. It’s true that the BIM 2.0 challengers aren’t sitting there with complete Revit replacements either.
But announcing the destination changes the competitive calculation. One of Revit’s strongest defences has always been the disruption involved in replacing it. Firms have enormous investments in content, standards, families, training, APIs and workflows. Staying put was easier than moving. Now Autodesk has told customers that staying put isn’t the long-term plan. At some point, the way they work is going to change to a different system anyway.
That gives Motif, Snaptrude, Arcol, Hypar, Qonic and, whoever comes next, a different sales pitch. They no longer must persuade firms to abandon an eternal Revit workflow. If substantial migration is coming anyway, Autodesk’s destination has to compete with everybody else’s, and the challengers arrive without Autodesk’s baggage: no Enterprise Business Agreement (EBA) history, no perpetual-licence grievance, no record of annual price rises and, for now, no meter above the seat. Their distance from producing a usable drawing set may also prove shorter than Forma’s distance from Revit parity.
None of this means they win. No BIM 2.0 tool has yet been proven on a large project, none have anything like Revit’s content ecosystem, and launch pricing from venture-funded companies says little about what they will charge at scale. But Autodesk has relied on customer inertia for a very long time, and inertia only works when staying put is the path of least resistance. After AU 26 it isn’t. Firms that have spent a decade absorbing price rises for software that hasn’t visibly kept pace now have both a reason and a window to look elsewhere. Autodesk may have weakened one of Revit’s strongest competitive defences by announcing its replacement before the replacement is ready. It is betting that customers will continue investing in Revit while Autodesk spends the next few years building Forma into a credible alternative.
Choice, Autodesk style
Autodesk stresses that customers aren’t being forced to move from Revit to Forma. That may be technically correct, but its licensing history gives customers good reason to distinguish between technical and economic choice.
During the move from perpetual licences to subscription, Autodesk didn’t switch perpetual software off. Its investor filings describe the maintenance-to-subscription programme, which started in 2017, as designed to “incentivise” maintenance customers to move to subscription while Autodesk increased maintenance pricing for those who stayed. In short, customers were priced across.
Autodesk needn’t repeat that playbook. It doesn’t need an end date for Revit at all. Put the newest AI capabilities into Forma first, make collaboration easier there, integrate Assistant more deeply and develop new workflows there rather than in Revit. Remaining primarily in Revit can become progressively less attractive without anybody being formally forced to leave.
Amy Bunszel, Autodesk’s outgoing EVP of AEC Solutions, told AEC Magazine editor Greg Corke that the company intends to work out the new paradigm on the Forma side. That matters because the question isn’t whether Revit continues, but where Autodesk puts its future development effort and what staying with Revit might eventually cost customers.
Autodesk’s August 2026 financial results offer a more current example of how it manages pricing. The company said it was winding down multi-year maintenance-to-subscription renewals and reducing discounts on multi-year contracts to improve what it calls ‘price realisation’. That does not mean Revit customers will be priced into Forma, but it shows that changing the economics of staying on an older commercial arrangement remains part of Autodesk’s playbook.
From seats to orchestration
One of the most important announcements at AU wasn’t another generative AI feature. It was Autodesk Assistant becoming a layer above Autodesk’s applications. Instead of a user deciding that a task requires Revit, Civil 3D or Forma Building Design, they tell Assistant what they want to achieve, and Autodesk’s AI Orchestrator works out which models, services and capabilities are required.
The wider AI industry is moving in this direction, with applications becoming collections of capabilities that agents call on as required. Control starts to move from the application to the agent deciding which applications get used.
Today, a practice might use Revit, Rhino, Speckle, Excel and an internal Python tool to complete a job. If Assistant starts choosing the tools, will it select a non-Autodesk application when that is better suited to the task? A customer-controlled agent could make those choices independently, reducing the importance of any one application.
For Autodesk, control of the orchestration layer may prove more valuable than supplying every tool an agent uses. Autodesk has an interest in deciding how those tools are selected, whether they are its own or supplied by somebody else.
The data toll
Autodesk’s Project Intelligence proposition rests heavily on Context. Generic AI doesn’t know why a wall moved, why an engineer rejected an option, which company standard applies or what happened six months ago when somebody made a similar decision.
Autodesk wants Forma to become the environment in which that knowledge accumulates. Models, drawings, decisions, project history, standards and collaboration data can form a persistent project intelligence layer that makes agents far more useful.
It could also create a formidable data toll. For decades, customers have worried about getting their DWGs out of AutoCAD and their RVTs out of Revit. In an agentic world, geometry may be the easy bit.
If an AI system has spent ten years learning how a practice designs hospitals, it could know its standards, preferred assemblies, cost assumptions, regulatory interpretations and thousands of previous project decisions. Moving platform would mean transferring far more than files. Project history, decisions and relationships would need to be exported in a machine-readable form that another AI system could understand. Even if customers retain ownership of that information, its usefulness may depend on Autodesk’s interpretation layer. If that knowledge cannot easily be transferred, the cost of switching could dwarf anything created by a proprietary file format.
Autodesk should answer the portability question clearly. Context may turn out to be one of its most valuable AI assets, but customers need to know whether their accumulated knowledge can leave with them.
Seat plus
The obvious assumption is that Autodesk needs to move beyond seat-based pricing because agents will reduce the number of humans required to perform traditional application work. What Autodesk told us at AU suggests it may instead retain the seat while adding consumption above it.
Autodesk told us that every automation machine requires a licensed named-user seat. Taken at face value, that raises an awkward possibility: an architect supervising five agents running Revit on five separate machines might still need five machine-associated seats, as well as paying any consumption charges. Agents would not necessarily reduce the seat count; they could become part of it. The precise licensing boundary remains unclear, as the next section explains.
Autodesk could retain application licensing while charging separately for cloud data, headless compute and specialist AI operations. AI labs already combine subscriptions with usage charges, but the underlying economics differ. AI inference has a variable compute cost, model prices have generally been falling, and switching providers can be relatively straightforward. Autodesk could meter access to applications and data that customers already license, in a market where prices have historically risen and replacing the underlying software can take years.
Pricing remains another unknown. AI providers publish model and API prices, so a company can estimate the cost of an agent before deploying it. Autodesk has not yet provided equivalent visibility into its emerging agentic economics.
Licensed to script?
For Revit customers, Autodesk’s position on automation raises an immediate question: which existing workflows require additional licences? Large AEC practices have been automating Revit for years using Dynamo, pyRevit, Rhino.Inside.Revit and internally developed scripts, some of which run unattended, overnight or on dedicated machines.
Autodesk’s licensing refers to ‘Background Activity’, but when AEC Magazine asked what the term covered, the company declined to define it. Without that definition, IT directors cannot tell where ordinary scripting ends and separately licensable machine activity begins, particularly when processes run without a user present.
The implications extend beyond future licensing costs. Firms could discover during an audit that workflows they have operated for years, believing them covered by existing subscriptions, fall outside Autodesk’s interpretation of permitted use. Customers need a clear definition against which to assess their compliance, rather than discovering the rules retrospectively.
Meanwhile, Autodesk is developing the infrastructure for a broader consumption-based model. In December 2025, it introduced a headless token class to Autodesk Platform Services (APS), establishing a mechanism for charging machine-driven compute. The AEC Data Model API now provides a practical example of these charges, detailed in the box-out at the end of this article. The Q4 FY27 Enterprise Business Agreement (EBA) renewal cohort will offer further insight, particularly given the importance of negotiated enterprise licensing to Autodesk’s largest AEC customers.
Autodesk Assistant’s commercial terms in 2027 will be another significant milestone. Andrew Anagnost has discussed amortising expensive AI operations across cheaper ones, while Amy Bunszel has referred to rate tables for computationally intensive tasks. These comments point towards a model in which customers could pay not only for access to software, but also for the work machines perform using it.
The full economics of Autodesk’s agentic strategy remain unclear, but the question of existing automation licensing should not. Customers need to know what they are required to license today, before Autodesk introduces additional ways of charging for what those machines do tomorrow.
Open, but metered
Autodesk’s adoption of Model Context Protocol (MCP) is an important move towards openness. Agents should be able to discover and call Autodesk capabilities without every connection requiring a proprietary integration. But an open protocol can still lead into a metered platform.
Nor does moving the AI onto the customer’s own hardware escape the meter. An AEC practice could run its own model and orchestrator, keeping the reasoning layer and more corporate knowledge on hardware it controls, while using MCP to connect agents to Revit, Rhino, Speckle, internal databases and other applications.
Running the model on its own GPUs can remove the cloud AI charge for reasoning. It doesn’t remove Autodesk’s charges when the agent crosses into its software and services. Local AI gives the customer control of the intelligence layer, but not necessarily the execution layer.
This runs against another trend in AI. Models are becoming more interchangeable, MCP makes tools easier to access, and local inference is becoming more practical. Software is becoming more composable at the same time as Autodesk’s commercial advantage grows from bringing data, applications, context and workflows together inside Forma.
How many layers will Autodesk meter? A customer could conceivably pay for the human’s application access, machine access to an Autodesk application, cloud data calls, headless compute and specialist AI operations within the same workflow. Each may be defensible on its own; stacked together, the economics could look very different.
Conclusion
The easy criticism of Autodesk’s AI strategy is that the technology isn’t good enough yet. That’s probably the least interesting objection. Models will improve, agents will become more capable, and Autodesk has the engineering resources, domain knowledge and installed base to develop credible agentic AEC technology.
The more interesting problem is what happens if Autodesk succeeds. Bringing software, project history, AI models and cloud services together has technical advantages: agents work better with access to context and applications, and customers do not want to assemble dozens of incompatible services. But technical convenience and commercial dependence have always travelled rather well together in software.
In the desktop era, Autodesk wanted to own the tool used to create the building. In the cloud era, it has worked to own more of the platform containing the project. In the agentic era, it is moving higher again, towards the layer that interprets what the user wants and decides how the work gets done.
There is also the matter of Autodesk’s revenue expectations. The company is forecasting around $8.3 billion this fiscal year, having grown 16% in its most recent quarter. With $10 billion the next milestone investors will be watching. But if AI delivers the productivity gains Autodesk is promising, firms could produce the same amount of work with fewer people. Whether that also means fewer Revit licences is now an open question. That presents a problem for a company whose growth has depended heavily on selling software to individual users.
Autodesk has a strong commercial incentive to capture some of those productivity gains. Higher subscription prices, more expensive Enterprise Business Agreements and consumption charges for automated workflows could all offset a decline in seat numbers. Its dominance of the AEC design software market puts it in a strong position to make those charges stick. We expect design software to become more expensive, not cheaper, even as AI reduces the labour required to produce drawings and models. Much of the saving could end up with the software vendor rather than the practice doing the work.
Customers will have to decide how much of this new computing stack they are prepared to let Autodesk control. Whether accumulated Context can be exported, whether firms can run their own AI agents and how many charges those agents incur when accessing Autodesk software may prove more important than the capabilities of Assistant or the eventual replacement of Revit.
If Autodesk succeeds, it will have built something considerably more powerful than another generation of BIM software. It will have positioned itself as the tollbooth between what an AEC firm wants to do and the machines that actually do it.
The price of data out
Autodesk began charging for the AEC Data Model API on 6 October 2026. Its published rate is one Flex token for every 5 MB of data exported beyond the included allowance. At an indicative $3 per token, moving a gigabyte through the API would cost around $600 before allowances; actual costs depend on token purchasing terms. This is not a general charge for cloud storage or ordinary file downloads; it applies to data transferred through the AEC Data Model API.
The distinction between what is free and what is charged is revealing. Autodesk says data imports, administration and API transfers between its own products are not billed under this rate. The charge applies to ‘Data Out’: information moved from the AEC Data Model to an external data store. The meter is therefore positioned at the point where customers want to use their model data outside Autodesk’s platform.
Eligible subscriptions come with monthly allowances, pooled across the Autodesk team. Each AEC Collection seat contributes 50 MB; Navisworks Manage contributes 25 MB; Forma Build Essentials 25 MB; Forma Build 100 MB; and Forma for Construction Operations 350 MB. APS also provides limited access through a Free tier. These allowances increase with the number of qualifying subscriptions, but enterprise agreements do not automatically include them: EBA customers without eligible subscriptions are billed on consumption.
Autodesk illustrates the volumes involved with its 90 MB Snowdon Towers sample model. Extracting every door uses 1.55 MB, while the windows require another 1.12 MB. A practice pulling that sort of information every day to feed external project dashboards, quantity reports or compliance checks could generate recurring Data Out across dozens or hundreds of models. Even relatively small extractions can mount up as project portfolios and reporting frequency grow. The bill depends on how much data actually leaves Autodesk’s platform and on the firm’s pooled allowances.
APS argues that API access is now more closely aligned with product subscriptions, making integrations easier to build and scale, and that customers pay only when they need higher-volume access or export capabilities. But Autodesk has also committed to giving at least 90 days’ notice of API pricing changes, which offers limited certainty to firms building long-term data infrastructure. Once external dashboards, analytics systems and AI agents depend on regular access to Autodesk project data, their running costs are exposed both to usage volumes and to changes in the rate table. Firms could find themselves paying repeatedly to extract information from models they have already paid to create and maintain.
The scale of any bill will also depend on implementation: caching and incremental queries can substantially reduce the volume transferred.
