The clay model

Nearly a century after automakers first started sculpting car designs in industrial clay, they still haven't stopped. Every major manufacturer does it by hand, at full scale. A single model takes three or four months of refinement by specialized sculptors, burns through hundreds of pounds of heated plasticine, and is so central to the process that even Tesla, a company that rethought almost every other part of car manufacturing, hasn't been able to skip it.

I've been thinking about this because software is in the middle of a similar transition right now. I've built a professional career around designing and developing software, but I've also spent years restoring cars and bikes in my garage, most recently a Mustang down to bare metal. Having worked in both mediums, the thing that strikes me about how cars are actually conceptualized and designed is how precisely it maps to what's happening in software. Not as a loose metaphor. As the same structural pattern, playing out in a different material.

The fidelity ladder

Picture every design discipline as a ladder. On the bottom rung are things like sketches or back-of-napkin drawings, the types of artifacts you make and discard just as quickly. And then on the top rung, you have the actual thing you set out to build. Every rung in between is an intermediate representation that exists for the same reason: the real thing is too expensive to work in directly, so designers build proxies for it at increasing levels of fidelity. Architects draw plans, build scale models, run simulations. Car designers sketch, surface in CAD, sculpt in clay. Software designers drew wireframes, built mockups, assembled click-through prototypes.

The history of every one of these disciplines is the history of the intermediate rungs changing. New tools compress some of them, replace others, and occasionally make one obsolete. But nobody in any of these disciplines set out to make a better intermediate artifact. They were always trying to get to the top of the ladder faster. The blueprint was never the point. The building was.

The pipeline

The modern automotive design pipeline runs three to five years from concept to production, moving up a fidelity ladder where each rung costs more and reveals more truth.

It starts with sketches, hundreds of them, loose and expressive, drawn by hand or on tablets. The best ones get rebuilt as digital 3D models, where engineering realities like aerodynamics and safety begin to constrain the artist's vision.

Then comes clay. CNC machines rough out the basic shape from the digital model in about two days. Then the sculptors take over, and the process slows down deliberately. They spend three to four months refining the surfaces by hand, shaping curves measured in fractions of a millimeter. The model is full scale. You can walk around it. You can see how light falls across the hood at different times of day. This is the step in the pipeline that has resisted every attempt at automating it away.

After clay, the model gets laser-scanned back into digital, refined into production-grade surfaces, and eventually moves to functional prototypes: drivable vehicles built from production-intent materials. But the clay model is the last major checkpoint before the manufacturer commits to tooling that costs tens of millions of dollars. It's the final proxy. After that, you're building the car.

Why clay survived

The obvious question is why. CAD, VR, and 3D printing have all been around for decades. Clay has outlasted them all.

It comes down to light. No screen can show you how it behaves on a curved surface at full scale. Digital modeling creates two-dimensional images of three-dimensional space. Even VR, which puts you inside the model, mediates the experience through a headset and rendered light. Clay is actual three-dimensional form under actual light. Designers and executives walk around the model, see it from every angle, crouch to check the character line at eye level, step back to evaluate the proportions from across the studio. Decisions get made in those moments that are impossible to make from a rendering.

Apple ran into a smaller version of this when they rebranded Apple TV. They ended up cutting the new logo from actual glass and filming real light moving across it, because the refraction was impossible to fake. The same reflex shows up in how Apple evaluates its own typography: SF and SF Symbols get printed at full scale, pinned to studio walls, and spread across tables, because no amount of zooming on a screen tells you how the whole system reads from across the room.

The representations that got squeezed out were the ones further from reality. Scale models shrank in importance as CNC milling made full-size clay faster to produce. Some wind tunnel work moved to CFD simulation. Tape drawings, full-scale line studies made with thin black photographic tape, are rarer than they used to be. The rungs closest to the real car held their ground. The rungs further down were the first to go.

Even the companies with the strongest incentive to break this pattern haven't been able to. Tesla, which approached almost every other part of the car-building process as something to be rethought from scratch, still employs master sculptors for their production vehicles. Jeff Martin, Tesla's clay modeling lead, worked on both the Model 3 and the Cybertruck. Rivian has gone further than anyone in trying to compress the clay phase, moving design reviews into VR. A review that normally requires a week with a physical model can now happen in about an hour with a headset, saving roughly a million dollars per vehicle program. But Rivian hasn't eliminated clay. They've compressed its role. Fewer reviews in clay, more in VR, but clay remains the final checkpoint before production tooling. When a company with every reason to skip the step actively tries to and still can't, that tells you the step is structural.

The software parallel

For decades, the software pipeline ran the same way: idea, wireframe, mockup, prototype, code. Each step increased fidelity, but each transition cost time and introduced translation loss. The wireframe couldn't capture the visual design, and the prototype couldn't represent real-world performance. By the time someone could actually hold the app in their hand and feel whether it worked, months had passed and the cost of changing direction was high.

The tools got better over the years. Figma made mockups collaborative and fast. Prototyping tools added animation and interaction. Design systems encoded decisions into reusable components. Every one of these improvements still kept you working in a representation of the software, not the software itself. You could make the mockup more interactive, the prototype more realistic, the handoff more precise. You could not make any of them into the actual product.

AI collapsed the cost of that last step. Tools like Claude Code and the wave of AI-assisted development environments that followed made it possible to go from idea to working software in hours instead of months. Not a prototype. Not a simulation. Working code, running on a device, responding to real input. The fidelity gap between "representation of the product" and "the product" went from months of engineering effort to an afternoon.

The same pattern played out. Visual design judgment still matters, though not as a moat, because you're now exercising it on the real thing rather than a static approximation of it. But the artifacts that existed to bridge the gap between imagination and reality? Wireframes, annotated mockups, redlined specs, click-through prototypes? Those were workarounds for a cost barrier. When the barrier dropped, the workarounds became overhead. Jenny Wen made this argument from inside the design field: that designers had become more fluent in process artifacts than in the products those artifacts were meant to serve.

The value of an intermediate representation is inversely proportional to how close you can get to the real thing. When working in the real medium is prohibitively expensive, the intermediate steps are essential. When the cost drops, the steps furthest from reality are the first to lose their justification. Clay survives because it's one rung from the real car. Canvas-based design tools are several rungs from real software. The pattern predicts exactly which one is in trouble.

The sculptor's job changed

So what happened to the sculptors? GM alone employs roughly 400 clay sculptors across its studios, and there are only about 25 schools globally that teach automotive clay modeling. These are trained artists whose core skill is feeling a surface with their hands and knowing how light will behave on it at thirty feet. The clay sculptor didn't disappear. But the job changed substantially.

Before CNC milling, sculptors built up the entire form by hand from raw clay, weeks of rough shaping before the refinement could even begin. CNC machines now do that rough shaping in two days from the digital model. The sculptor's work shifted from building the form to refining it, from generalist shaping to the specific, high-judgment work that machines can't do. This kind of work takes the shape of all the invisible elements that make a vehicle feel organic and expressive, like the subtle transitions between surfaces or the way a crease softens as it runs along a door panel. These are design cues you would never notice unless you're standing next to a full-scale object under real light, and perhaps not even then.

The same pattern is playing out in software design. The work that required a canvas is exactly the kind of intermediate-fidelity work that AI handles well. The work that requires a real product in your hand is different. Things like the rhythm of a transition, or how a screen handles network strain, don't translate to a prototype or a canvas. That still requires a human with judgment, exercised in code on a running application, not on an artboard.

The sculptor's job still matters. They mold and shape the clay itself, not a sketch of the clay. The designer who thrives in this shift is the one who can evaluate the real thing, not the one who was most fluent in the tools that approximated it.

Closing the gap

Nobody in the history of automotive design set out to make a better clay model. They were trying to make a better car. The clay was a craft-intensive, hundred-year concession to the fact that stamping metal is expensive and slow. Every tool that came after it, from CAD to VR to generative design, was another attempt to close the gap between the designer's intent and the physical vehicle. The tools that succeeded didn't replace clay by being better representations. They replaced the steps that were worse representations, and they made clay faster to arrive at.

The same logic applies to software. Nobody chose to design on a canvas because a canvas was the ideal medium for design decisions. They chose it because code was too expensive, too slow, and required skills that most designers didn't have. The canvas was the clay. A useful intermediate, but one that existed because of a cost barrier, not because of an inherent advantage.

The gap between intent and reality in software is closing faster than it ever has. And history says only the rungs closest to reality survive that change. The fluency in tools that exist to bridge the gap becomes optional; the judgment to evaluate the real thing doesn't.