The Murky Middle: Representation, Translation, and the Real AI Opportunity in Product Creation
Everyone can see AI at the edges of product creation.
On the front end, generative tools are producing concept imagery at a speed and volume that would have been unthinkable three years ago. On the back end, virtual photoshoots and AI-generated marketing imagery are compressing timelines and budgets for e-commerce at scale. Those applications are visible, well understood, and — for the brands investing in them — already delivering measurable returns.
It’s the middle that’s murky.
The middle is where a sketch becomes a pattern. Where a pattern becomes a simulated garment. Where a 3D model needs to behave like the physical product it represents — materials that drape correctly, structures that hold up, constructions that can actually be manufactured. This is where the most consequential decisions in the product creation pipeline are made, where the technical demands are highest, and where AI’s potential is least understood. It’s also where I spent the latter part of my career at Nike, building the digital product creation infrastructure that sits between concept and commerce. The middle is where the real leverage is. And almost nobody is talking about it.
The Bookends Are Easy to See
The visible applications are attracting the visible investment. According to McKinsey and BoF’s State of Fashion 2026, over 48 percent of global fashion brands have integrated machine learning into collection planning and 3D sample generation. The AI fashion market was valued at $2.78 billion in 2025 and is growing at nearly 40 percent annually. Nike’s A.I.R. project used generative AI and 3D printing to co-create footwear designs with athletes, compressing concept-to-prototype timelines from weeks to hours. Adidas trained a diffusion model on its entire sneaker archive dating back to the 1950s to generate new concepts. On the other end, platforms like Vue.ai are producing on-model fashion imagery at a quarter of the cost and five times the speed of traditional photoshoots.
None of this is the hard part. Concept generation and marketing imagery are the lowest-risk applications because getting them wrong costs relatively little. A concept that doesn’t work gets discarded. A marketing image that misses the mark gets reshot. But in the middle of the pipeline, decisions carry downstream consequences that compound — a pattern that doesn’t work delays manufacturing, a simulation that doesn’t reflect material behavior leads to costly physical corrections, a 3D model that can’t translate to production engineering wastes weeks of work across multiple teams. The stakes are structurally different.
Representation vs. Simulation
For two years there was a race in fashion tech to go from prompt to 3D model. The results looked impressive on screen — visually compelling shapes you could turn in space, light from any angle, drop into a review meeting. But what those tools produced were solid mesh forms with no manufacturing utility. You couldn’t take them into pattern making. You couldn’t simulate fabric behavior on them. You couldn’t grade them, spec them, or hand them off to a factory. They existed in the space between concept and product, which is precisely where representation alone isn’t enough.
The race I’m describing isn’t over — it’s intensified. Open-source models like Microsoft’s TRELLIS.2 now produce watertight meshes with PBR materials in minutes, dramatically better than the demos of two years ago. For 3D printing and additive manufacturing, the outputs are usable. For pattern making, sample engineering, and production specs in soft goods, they still aren’t. Better meshes, same gap.
The story I’ve been telling for the last two years — AI is good at representation, AI can’t do simulation — is starting to get out of date. Not because the AI itself learned to simulate fabric physics. It didn’t. The simulation engines didn’t suddenly need it to. CLO3D, Style3D, Browzwear, Marvelous, Substance — those tools already handle the physics, and they’ve been handling it for years. The question was never whether the math was good enough. It was who or what was driving the app.
That’s the shift actually happening, and anyone working seriously inside a product creation pipeline right now can feel it. The agentic layer is moving in. Not as a replacement for the simulation engine, but as the operator of it. Sketch in, agent runs the pattern step, agent drives the simulation, agent hands the result to the next stage. None of the major platforms is going to publish a press release titled “agents are now operating our software end-to-end” — but if you’ve been in the rooms where the work is happening, you know roughly where every one of them is. Some of it’s public. Most of it isn’t. Style3D has been the most direct about positioning around an agentic layer. Adobe rebuilt its Substance 3D pipeline on OpenUSD and automation and pushed 55,000 renders in three months. The line between automation and agent is getting thinner by the week.
So the rep/sim binary is collapsing into something more useful. AI does representation natively. The simulation engines do simulation. The middle of the pipeline isn’t waiting on AI to learn fabric physics — it’s waiting on the orchestration layer that lets an agent drive the engines that already exist. That’s a translation problem.
The brands already operating this way are quietly pulling away. Tapestry trained Adobe Firefly on proprietary design elements they call “Coach codes” — brand-specific materials, hardware, construction details — so a digital twin stays on-brand from concept through marketing. That’s not a demo. That’s a production workflow where AI operates at the level of fidelity each stage demands. And what’s public is a small fraction of what’s actually happening inside these organizations. The real experimentation is below the waterline.
The Translation Layer
The most immediate AI opportunity in the middle isn’t full simulation. It’s translation.
Once you stop expecting AI to do the physics itself and start asking how an agent operates the engines that already do, the whole problem reframes. The question isn’t “can AI simulate a draping garment.” The question is “can AI take a sketch, prep the pattern, set up the simulation in the right tool, evaluate the result, and hand it off to the next stage.” That’s translation, end to end. And it’s exactly the kind of structured problem AI is already good at.
Every product creation pipeline has handoff points where work crosses from one system, one team, or one format to the next. Creative output needs to become engineering input. A sketch needs to become a pattern a designer can adjust in the software of their choice. A pattern needs to become a simulated garment without someone manually constructing every seam. In footwear, a sketch needs to become a procedural model — something editable and adjustable — and then transform into a real-time rendering that teams can evaluate together. These translation steps are where enormous amounts of time and fidelity get lost.
The numbers support this. Brands using digital twins and AI-assisted workflows report physical samples dropping from 15–25 per style to as few as 1–2, with pre-production costs falling 60 to 80 percent and development timelines compressing from 12–16 weeks to 2–4. Rendering time for a fully layered outfit dropped from five minutes in 2025 to under 90 seconds in 2026 using GPU-optimized AI pipelines. CLO3D reports 28–40 percent time savings in pattern-to-simulation workflows with AI-assisted features.
These gains come not from AI replacing the people in the middle, but from AI handling the translation between stages — converting formats, bridging systems, maintaining fidelity across handoffs — so that designers and technical developers can spend their time on the decisions that actually require human judgment. Translation is a tractable problem. Full simulation isn’t — not yet. The distinction matters because it changes what you build toward today.
The Human in the Loop
The industry has an underdeveloped understanding of how to use AI in stages alongside the people who provide oversight, quality control, and domain expertise. Most conversations about AI in product creation jump straight to capability — what AI can generate, how fast, at what resolution. Very few start with workflow design — at which step does AI contribute, at what level of fidelity, with what kind of human review at each gate, and who owns the decision when the AI output needs correction.
This is a design ops problem, and it’s the same one I argued in my last piece. The brands that built the operational foundation around 3D pulled ahead while others invested in tools they couldn’t actually deploy. AI in the middle of the pipeline is heading toward the same pattern. The talent bridge — the people fluent in both the creative process and the digital tools — is the connector between where AI is now and where it needs to go inside a working product creation pipeline. Eliminate that capability in a downturn and you don’t just lose what you built. You lose your best shot at what comes next.
AI doesn’t need to replace the people in the middle. It needs to work alongside them — in stages, with oversight, at increasing levels of fidelity. Build that workflow and the middle of the pipeline stops being murky. Leave it unbuilt and AI just becomes another tool that promised more than it delivered — not because it couldn’t, but because nobody designed the conditions for it to work.