The CAD File Is Not a Static Artifact Anymore

🏭 Breaking the Bottleneck is a weekly newsletter and interview series covering manufacturing technology and physical AI. Want to chat? Reach out at aditya@machinafactory.org or connect with me on LinkedIn.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.

“A CAD file is no longer an artifact, it’s part of a connected data system”

You started your career as a Navigator and Gunnery & Missiles Officer, then spent two decades running digital transformations at Chase, OpenTable, and Yahoo before taking over Autodesk’s manufacturing business in 2018. Consumer internet and banking are slightly different than manufacturing, with software-native data and operations, and far lower integration and deployment overhead. What from those playbooks actually translates into how Autodesk approaches go-to-market, and what lessons could manufacturing software take from those markets?

I come from a family of factory workers and civil servants, so I had early exposure to how things get made and a real sense of wonder about making them. In its own way, a navy ship is a factory: steam propulsion, nuclear reactors, mechanical guns, radar, all of it running as interconnected systems.

Each industry I worked in taught me something different. Banking is about regulation and risk management; consumer businesses run on short cycles and the ability to move fast. But it is the commonalities that transfer. The first is ease of use. Nobody likes things that are hard to use, and running the mobile banking app at Chase taught me the relentless focus on simplification, making things easy for the customer. The second is efficiency: booking a restaurant table, ordering ahead, and the same theme kept recurring across every industry.

The third is a growing recognition of the importance of data. In banking, better data meant more accurate risk prediction and fraud detection. Manufacturing can benefit from this as well. A CAD file is no longer a static artifact; it is a living, breathing data system. Conceptual design, mechanical design, and manufacturing data are the foundation.

And the lesson beneath it all is that inertia can kill a company. The real question is how our software raises productivity relative to the old ways of working and the point solutions people have relied on. Consider what those cost today: data corruption, lost data, endless migrations between systems, or the sheer time spent producing 2D drawings, which can account for 40% of an engineer’s job.

You’ve framed AI as the third great shift in design and make, after parametric modeling in the ‘80s and Windows-based CAD in the ‘90s. For the median shop, the gap between paper-and-process and agentic orchestration with MCP integration is significant. How can machine shops and manufacturers begin that journey, and how do they balance this journey across different technical maturities, for example, design maturity with Fusion 360, with a legacy ERP for operations?

There is still some paper-and-pencil out there, but manufacturers have generally embraced technology, even before migrating to a modern stack. The gap you describe does sound intimidating; in my experience, the path is more practical than it looks.

It starts with data. Smooth, connected, flowing data is the currency that ties your entire process together, and that includes the processes themselves, the knowledge of how things actually get done. You cannot layer agentic AI on top of tribal knowledge.

That is worth stressing, because design and manufacturing are fundamentally different from text-first domains. They are more complex, and they demand precision. We spend a lot of time on how frontier models can augment design and make. Still, we also see their limits firsthand: to be useful on a factory floor, a model has to be trained specifically on real physics and real geometry. Engineers define intent that must hold up across geometry, tolerances, materials, toolpaths, and quality systems. Close enough is not good enough; it will break a machine and cost you a fortune.

So we usually begin with a more basic question: how do you connect your data? Most manufacturers are wrestling with fragmented legacy data, and much of our industry has built proprietary formats that do not work with anyone else’s systems. We don’t believe in that approach. What we have found is that AI becomes far more effective once engineering data has context and is connected across the product lifecycle, and, interestingly, AI can help get it there by interpreting and structuring legacy data and helping one system talk to another. I remember hard-coding systems to work together, and it was not that long ago; then came API connections, and now MCPs make it easier still.

That is really the answer to the maturity question. Interoperability is far easier in a more open ecosystem: you may have a legacy ERP and still connect to it. But I would add a caveat: if you are only using an MCP to push data into a legacy ERP, you may not be capturing the full advantage AI can offer.

Then it comes back to ease of use. We show customers how simple this is to adopt, and we hear it constantly with Fusion. Unlike some legacy CAD tools, you don’t need two decades of experience with Windchill, Teamcenter, or Enovia to make it work. From there, we talk to manufacturers about modernizing their end-to-end system so they can capture the full competitive benefit. Fundamentally, manufacturers that standardize and operationalize their engineering and design data will be best positioned to leverage AI as a competitive advantage.

One thing did surprise me. I think constantly about the mid-market, and I came in wondering how sophisticated those manufacturers would be. I have been struck by how much they have embraced technology; this is not my parents’ generation. But they do need help, and they need easy tools that connect everything for them.

In your February 2025 testimony to the House Energy & Commerce Committee, you argued that domestic reindustrialization fails unless manufacturers of all sizes become digitally enabled. The political reality is that reshoring dollars and ribbon-cuttings gravitate toward buildings, tooling, and steel. If the next wave of incentives once again funds machines but skips the “digital plumbing,” what more is needed to drive cost-competitive reindustrialization? And what role have you seen SBA loans play across the SMB ecosystem in reindustrialization so far?

First and foremost, we have to invest in modernizing our design, data, and manufacturing systems, both the software and the hardware. I have said the same to a room full of chief information officers in India: you cannot simply throw bodies at this problem. If you are not modernizing your hardware, your software, and your processes, you will fall behind.

You are pointing to two problems we have been trying to reconcile. The first is that if reindustrialization reaches only the top of the pyramid or only certain sectors, we do not actually rebuild the industrial base, which is a long-term danger. The second, and more common, is that everyone thinks about hardware: let us offer incentives to go and build. But without rethinking your processes and your software, you will not be competitive.

Small and medium-sized businesses are not a nice-to-have here. They make up around 90% of manufacturing firms. They are the backbone of any manufacturing economy, not only in America, and a supply chain is only as strong as its weakest link. So, to your framing: if the next wave funds machines and skips the plumbing, we will have a problem.

That is why we are constantly reminding lawmakers that our global competitors are not investing solely in physical factories. We need incentives aimed specifically at digital transformation and AI adoption among smaller manufacturers. Loans from the Small Business Administration are one instrument; there are also tax credits, matching funds, and other targeted measures. Larger manufacturers can benefit too, though they already enjoy some advantages of scale.

In some ways, AI can be a great equalizer, giving a small shop access to capabilities that were once the preserve of large companies with deep pockets. As I mentioned, connecting old legacy data to a modern system, something that used to take a lot of people, is now far easier with MCPs, which can reduce the integration required.

SBA loans are an important source of capital for small manufacturers, but far from the only one. The very fact that we were invited to Congress to discuss competitiveness was encouraging, but we still need programs at both the national and state levels to expand access to low-cost capital. I said something in Congress that I firmly believe: manufacturers, certainly in America, and I suspect more widely, do not want handouts. They want to compete. Give them access to capital so they can invest and stay competitive. And these programs pay for themselves: a loan is not a grant, so it is repaid, there is a return on the investment, and firms then pay more in tax as they succeed. A broader, more resilient manufacturing base also brings a halo effect of jobs and domestic sales.

If I were king of the forest, I would set up directed programs, “directed” in the sense that firms choose from a menu. Perhaps a company has already modernized its software; perhaps it has just built a new factory and realized it must upgrade that software; perhaps what it really needs is skilled people. Let manufacturers decide what they need, but give them a menu, because success requires investment across all of these. Imagine a state or federal program that said: to compete in manufacturing, you will need to modernize your digital plumbing; you may need a new or retrofitted brownfield factory; and you may need to train your workforce. Here are three programs that support exactly that, choose what you need and apply.

Unifying design and manufacturing is a promise PLM vendors have been making for 25 years. How are you embedding AI? What changes are you making across the stack? Furthermore, how is Autodesk positioning itself with the MaintainX acquisition to further integrate manufacturing data upstream into the design process?

We have not closed the MaintainX deal yet, so I will keep that part hypothetical. But let me start with the premise, because I do not think PLM systems converge design and manufacturing on their own. They are one part of the process. We think in terms of CAD, CAM, PLM, and supply chain, and how all of those systems fit together. PLM performs a critical function, data and change management across the product lifecycle, but on its own, it stitches data together rather than truly converging it. What we are striving for with Fusion and our industry clouds is for CAD and CAM to communicate natively.

We believe in a far more integrated industry cloud that goes beyond point solutions. Those point solutions have served our industries extremely well for decades. American manufacturing was built on them, and they still work for many people. But manufacturing productivity gains have been plateauing, and to reach the next level of breakthrough, we believe you need AI within a converged industry cloud. It is not PLM alone.

So we embed AI at every step, CAD, CAM, and PLM, and we think about it on three levels. The first is task-based automation: we look for individual tasks that could be made more efficient. Automated drawings are a good example. We rolled that out a couple of years ago, using AI to spare designers and engineers the time required to produce them; we have done the same with sketch constraints, and we recommend machining strategies from a trained CAM model. These are the tasks customers tell us are low-value and inefficient.

We are now moving into workflow-level automation, using AI to power an entire workflow. One of the most exciting is how a design begins: going from a text description to 3D geometry, sketch to mesh to B-rep, invisibly to the user. I think AI will be game-changing in PLM in particular, because PLM systems have been so rigid and so dependent on tribal knowledge and bespoke rules. AI should be able to simplify the interface, capture the intent a user would otherwise enter by hand, understand it, and effectively build the PLM system for them, without anyone ever having to log in.

Last year, we ran an analysis with a large consulting firm and surveyed hundreds of PLM customers of all sizes. To a one, they hate their PLM system; switching it on is the worst part of their day. It should not be that way. They understand how vital lifecycle management is; the interface is cumbersome. You will hear more from us on this in the autumn.

The third level is system automation. We envisage a world in which entire systems are connected through AI, so that you can take a design change you are contemplating and predict its ramifications across the factory, across operational processes, even down to the logistics of the supply chain and what is actually available. AI connects the whole system and gives designers, manufacturers, and engineers visibility into the implications of their decisions at every stage.

On operations, as Andrew put it on our earnings call, we have long thought about the “design and make” process, but “operate” was always the logical next step, and it cuts across our architecture and construction business as much as manufacturing. People operate large buildings; people operate factories. When we can capture data coming off the factory floor and feed it back into manufacturing to make that system smarter, then back into design to make that smarter, with information flowing both ways, the value is immense. Some of our manufacturing tools have already moved into the operations group: we have done factory design work with Inventor for years, and we acquired FlexSim, which builds digital twins of a machine setup and makes a factory much easier to design.

Where this really compounds is context. We have been building a context graph: picture a single customer, every product they design, all the accumulated knowledge in their processes, and what does and does not work for a given product. We can capture that as a time series of context around design, manufacturing, processes, data, and what the customer cares about. With design and manufacturing in place, and now operations too, you build a far richer body of context, and it becomes much more feasible to interpret and consume it through a model tuned to a specific task. Large language models have none of that. Like everyone, we are impressed by what they can do; we also see firsthand where they struggle, and we think hard about how to use their strengths while giving customers what they actually need. When we apply that across multiple industries, we will unlock the power of AI throughout the economy.

As AI takes on tasks like drafting and G-code generation, you’ve argued that the role shifts from “button pusher” to problem-solver and systems orchestrator. Yet the upskilling and change management these roles demand is often very difficult, given entrenched ways of working, time constraints, and a changing workforce. How will jobs change with AI, and what steps should manufacturers take to overcome some of these operational hurdles?

There are two challenges here, perhaps three. One is the aging workforce; the second is upskilling. The third, which is really the backdrop, is something Andrew has said many times: there is a large backlog of work across many of our industries. Infrastructure projects are delayed for want of capacity; manufacturers would love to produce more but cannot, and construction is riddled with the same constraint.

Our view is that AI can augment these processes and unlock more capacity to get things done. In our industries especially, there is pent-up demand, and we see AI as a way to meet it despite the aging workforce, but it will require upskilling people. On our side, that begins with embedding AI in our products, so that adopting our tools makes people faster at the work they are already doing and helps clear the backlog.

For the next generation, we give our software away free to universities, vocational schools, and secondary schools. We believe part of our role in society is to train young people on modern tools so that the economy stays competitive, and we invest heavily there. We have now committed $350m over three years to provide access to technology, training, and certification for the AI-powered jobs that design and build the physical world, a continuation of what we have done for a long time. Part of that, $10m, is pledged to the RAISE program, a bipartisan effort to improve the training of future workers. It is a small fraction of the total, but it all adds up.

Specifically on the aging workforce, the single most important thing I tell manufacturers is to start documenting their processes. A great deal of knowledge is locked up in retiring senior people, and we can now capture it. One MIT professor I shared a panel with described using visual-learning systems to record what people do on an assembly line and learn from it, so that a new hire who lacks thirty years of experience still gets the benefit of that experience, because it has been captured and trained into a model. It starts with manufacturers deciding to document the existing processes they want to keep and then using AI to train new workers on them.

On change management, we try to make our AI tools intuitive enough that people can pick them up and take advantage of the task, workflow, and system automations, and we have people on our team who guide customers through deploying the software. But the single biggest factor is simply getting people to experiment. Even our own software engineers, some of whom have built the same products for ten or twenty years, are now having to learn entirely new ways of working, and what I have found is that once people start trying things, the capabilities tend to blow them away. One colleague who has been here for twenty years and in the industry far longer told me this is the most exciting period he has known in technology. Still, it began with him embracing the change and then getting out to experiment. He now does his own G-code generation on a little machine shop at the back of his house, testing and running it himself. I tease him that I worry about his marriage. That kind of enthusiasm is contagious. As with any change management, even before AI, you first have to get people to embrace the idea of change and the better outcome it can bring. Once they do and start experimenting, it becomes exciting and contagious.

________________________________________

Note: Edited and Paraphrased for Length

To contact Jeff, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Ironloop, Upkeep, and Industry 4.0 Club.