Learning from 3D: Part 2 - Where We're Going in 2026 and Beyond

3D is dead! When H&M's Head of Digital Product Creation takes the stage at The Fashion Tech Show Europe, he knows that's the number one thing many in the audience will walk out thinking. Ram Kummamuru addresses it head-on. But the real question isn't whether AI replaces 3D. It's whether the industry is ready to stop adding AI features to their current tools and start asking: what if we threw it all out and rebuilt for a completely different world? Here's where H&M is headed in 2026, and what it means for everyone else.

This is Part 2 of a two-part conversation. Read Part 1 here.

Why 2026 Is Different

The technology finally caught up. But that's only half the story. The other half is what happened inside H&M.

Ram, why is this transformation possible now in a way it wasn't in 2024? What changed?

Ram: One of the greatest learning curves for me was how fast AI evolved in 2025, especially generative AI for image generation. At the beginning of the year, I didn't believe it was there. Three months in, I still felt maybe not. Then all of a sudden, new models started coming up that got our attention. The technological development completely baffled me.

The second key learning: I didn't believe we'd do this transformation in 2025 until the C-suite entered the conversation. When our top executives said, "Let's make this happen," that was the driving factor. If you want to bring this kind of transformational change and challenge your norms, you cannot push from bottom up. You need major traction within the company to ride the wave.

Third, I was surprised by the response. I was worried designers would say, "No, I don't want it." We didn't get that feedback. We got constructive conversational dialogues. Where it worked, where it didn't, what they thought about it. These conversations helped us and continue to help us.

And here's what's interesting: use cases and features we are expanding in 2026 - they are not our ideas anymore. They're ideas from users calling us up saying, "We have an idea to explore AI for components ideation. Can we explore it?" Due to positive adoption, we moved away from selling the tool, but rather focusing on "guardrailing" to ensure we continue to have an established process before tools.

Holger, from a technology maturity perspective, what needed to happen for AI-first design workflows to become production-ready?

Holger: I think 2025 was the year of initial adoption, and now you see questions coming up: Does it really create all the value we expected? If you listen to all the buzz, 2025 was throwing AI at every problem. But it doesn't solve every problem. It solves specific problems extremely well.

In 2026, we'll see AI being placed where it adds value, like Ram does at H&M. We see our clients adopting similar mechanisms, but more conservatively. Some clients tell me the generation quality isn't good enough yet. One out of ten, one out of five images is usable. So you still need to improve the technology.

Data security is also very important. Generating images is one thing, but learning all about your products and giving that to an external company? Maybe not the best idea. Your visual IP is in those images. You don't want to train every common AI on all your patterns.

Ram: I believe 2026 will also be a year where we continue failing in AI. There will be good stories of success, but we'll also see a lot of failures and loss of trust. People will get excited, pick up AI use cases that aren't mature enough, invest a lot of money, not get the right outcome, and get frustrated.

I think there'll be equal balance. Maybe even higher on failure. Are we going to see major things happening across the industry in 2026? No. Number one reason: bureaucracy. Enterprise companies have bureaucracy. It takes time to get approvals. But I'll be very surprised if you don't hear more stories of direct impact from generative AI and agentic AI within the year. I'd say 2027-2028 is when this becomes normalized.

Governance and Focus

Everyone wants AI. The key is saying no to the wrong use cases.

You created a "Creative Board" to govern feature priorities. Why was that important?

Ram: The critical aspect of the creative board is having forward-leaning experts who have a say in our tool. What became critical is we weren't doing AI for the sake of doing AI.

Here's an example: Wouldn't it be cool if you pressed a button and AI generated a video of a product? A really exciting feature. But then I thought: wouldn't it be more interesting if, rather than generating random hex code colors, we could generate proper Pantone or H&M color codes? That has a higher value. The first one excites me as a tech geek. The second one solves a business problem.

Rather than me thinking about it, I put it in front of the creative board. They're brutal. They don't care about AI or what cool stuff it does. They want to solve business problems. So when we get excited about new features, they say, "Ram, great feature, but this other thing isn't even working. Can you fix that first?"

We get humbled in that room — challenged to get back to what critical business problems actually need to be solved. We're not thinking from a tools perspective. We're looking at the process and what technology best fits.

How is the board composed?

Ram: There are two groups. One is overarching DPC priority, beyond AI, just everything we need to do. That board has representatives from each customer group, heads of prints and products. Not the managing directors, but the people responsible for assortment planning and development. They set the tone on what needs to be delivered.

When it comes to AI specifically, we deliberately included both AI enthusiasts and AI skeptics. We have people who are natively excited about AI, but we also have people who are not. We get both perspectives. It keeps us honest.

Holger, what advice would you give brands about governance as they adopt AI tools?

Holger: Ram pretty much covered everything. You need your users on board. You need a pull, not a push. But you also need management support. Ideally, you see this as a transformation of your process.

You need to figure out where you can generate the right value, then put your stakes there. To do that, you need to talk with people on the ground and talk with management to look at the figures. I've been thinking about this question for some time because the question in my head always was, "What does it (AI) do to create the key value and how, and where?" In the end, I settled with it being a revolution, but at the core it's a transformation. Very impactful, game changing, but another transformation.

You should use the same tools that worked in your company before. Probably at H&M, you had specific boards for specific challenges before, right? The same way digital creation was a board-level priority at Adidas that needed buy-in from all the business units. Treat it with respect, bring in the users, and figure out where you can generate value.

Technology Stack Decisions

Fashion brands aren't software companies. Ram says stop pretending to be one.

Ram, you're using open models rather than training custom models on H&M's data. Why that choice?

Ram: Enterprise contracts and legality are important. When we use AI models at H&M, we use enterprise agreements. There's an agreement between companies that they cannot train on our data.

But it's also important to realize: custom models might work, but you have to justify the cost-to-value ratio. Out-of-the-box models have a very high cost-to-value ratio. The moment you build custom models, maybe the initial value is there, but someone has to maintain that custom model, update it, refine it.

You have to weigh the pros and cons. Some brands need higher data security. Maybe they'll say they need custom models and accept the additional cost. Look at the full picture. There's no one rulebook. Look at what kind of brand you are, what brand DNA you have, what security guidelines you need, and make the right decision.

And I'll always say this: compared to Meta and Google, we are a fashion brand, not a software company. Don't try to be one.

Holger, when should brands use open models versus custom-trained models?

Holger: We've seen both. Usage of existing models because they're extremely versatile and well-trained, and custom or adjusted models because there's specific IP you don't want to give out.

It's always a question of what kind of data you have (accessible) and where (can you get it from). You need to train your users. You can't just dump your vision for the next season into an unlocked tool or go to ChatGPT with no security means and drop stuff in. It will learn it.

It's a function of the company to ensure you have the right means to protect your data. There's a cost if your data goes public, and you need to balance benefit versus risk.

Building your own model from scratch? Probably impossible these days. No brand besides the big ones like Meta, Microsoft, OpenAI, etc. can do that. It's specialized knowledge, at a high cost. You can adopt models, enhance them, train them. But be aware of costs and your business case. In the end, it's always the business case. Don't introduce something if it doesn't generate value. And data is value. If your data goes out, you lose a lot of value as a brand.

Looking Ahead to 2026

The goal for 2026? Keep the momentum, stay small, and keep failing on purpose.

Ram, what are you most excited about for the rest of 2026? What's on the roadmap?

Ram: I made a promise to my team entering 2026: let's just try to do what we're doing. We had a really good ending to 2025. Let's keep the same momentum. That's my ambition.

From a features point of view, we're experimenting more with agentic AI, vector databases, and RAG models. In practice, that means building AI systems that can autonomously retrieve relevant information from our internal databases (like our pattern library or material specs) and use that context to generate smarter, more accurate outputs. Obviously, I can't expose all the use cases. Then I'm giving too much away. But we're experimenting with things that require that kind of architectural significance.

We also plan to add more users with different use cases: component ideation, print development, and everything in between. Expansion with more advanced and complex features, and failing at those features while experimenting. That excites me for 2026.

I think we have a lower percentage of failures and higher percentage of successes compared to last year. But I do believe we should continue to fail. Otherwise you're doing something really wrong.

What about team size? Are you scaling up?

Ram: I plan to keep the team small. Two reasons: with the development of AI for coding and programming, we can do a lot more with fewer developers. And organizationally, getting too excited, expanding the team, then another financial crisis and reorganization? Not my favorite topic.

I'll be pragmatic about long-term versus short-term needs. Long-term scale-up requires permanent headcount. Short-term feature development for three to six months? Maybe we bring in more people. Maybe a partner like 4D Pipeline that has deep experience in these problems across multiple brands. We'll keep the same mindset: small, intact team delivering great value.

Holger, what do you think the fashion industry needs to understand about AI in 2026?

Holger: Embrace it. It's there. It's not going away.

Make use of AI in the best possible manner. Find out where you see value. Try and fail. Try and fail fast. Don't go crazy and build the next spaceship. Do the small incremental things where you really generate value. Do so by doing small proof of concepts. We're more than happy to help our clients with that.

I'm really curious how the speed of AI will continue. 2024-2025 was crazy. Every week, every month, new models, new improvements. It feels like we're on a plateau in some places, specifically on large language models. But you will have to embrace it. It's not going away.

And don't be afraid. AI will not change that you're a fashion brand. It will not change that you have to create products. It will just help you make better products, get more inspiration, get faster, get better quality. Whatever you pick. It's a great tool.

Final Thoughts

Ram knows what headline the industry wants to write. He's here to give them a better one.

Ram, you're presenting at The Fashion Tech Show Europe in March. What do you want people to take away?

Ram: I'll put it two ways: what I don't want them to take, and what I do want them to take.

I don't want them to assume things. I don't want them to walk out assuming 3D is dead. I want them to have clarity on what I'm trying to say. What it is and what it is not.

What I want them to take away: hopefully they understand that what we're doing at H&M is not something you can copy everywhere. Maybe it's meant for our use cases. But I want them to get a thought process where AI is not this hypothetical thing. There are production use cases. It is possible to develop them.

I want them to walk out dreaming and visioning what AI could do for them. I want them to leave saying, "What if I use this for my processes? Can I solve these problems in my company?"

If I can make them believe that the "what if" is not so hypothetical, not a dream state, that it's actually doable in some scenarios, I'm happy. That's the outcome I'm hoping to walk out of that room with.

Ram Kummamuru will be presenting at The Fashion Tech Show Europe in London, March 30-31, 2026. Holger Schmidt will also be at the show and available for consultation. For more information about H&M's AI transformation or to schedule a meeting at the event, visit 4dpipeline.com or connect with Ram on LinkedIn.