This article was originally released in The Interline’s AI Report 2026.
Including profiles and exclusive interviews with 17 AI companies, stories and opinions from 12 different industry perspectives, and real survey feedback from around 100 voices from every level of fashion, bottled and analysed, The AI Report 2026 is essential reading for anyone interested in AI for fashion.

I spent a lot of 2025 researching for my MA thesis at the Munich University of Applied Sciences. As part of that research, I interviewed nine practitioners, whose work straddled AI, entrepreneurship, fashion forecasting, and supply chain disciplines. I asked each of them what they saw happening today, and what they predicted for the future, when it came to AI’s influence on not just understanding the market, but on optimising sourcing and manufacturing, and managing overproduction.

Some interesting strands emerged that united them all, and The Interline asked me to capture those themes, and to present my own analysis of them for this report.

The quotes that appear in this article are real, sourced from industry experts who spoke to me on condition of anonymity. The argument that connects them all is mine, but it’s one I suspect the wider industry will recognise as a reflection of reality – whether we choose to confront it or not.


The story fashion gets told about AI

The accepted wisdom, today, is that machine learning and computer vision are making demand forecasting sharper. Transformer-based models, the architecture behind large language model families like the GPT, Gemini, and Claude series, synthesise sales data, social signals, and visual trend information faster than any human team could. Generative AI reads consumer behavior almost in real time, across a signal array that far outmatches our own capabilities to parse. 

None of this is wrong. The gains in predictive accuracy over the last three years are real. The world’s biggest forecasters know this, and their businesses – and yours – are changing as a result.

But I want to put forward an uncomfortable proposition: forecasting, no matter how good AI makes it, has only ever been half the answer to fashion’s overproduction problem. And it may not even be the most important half.

This was the most striking finding from my interviews, and it’s borne out by market data: BoF and McKinsey’s State of Fashion 2025 estimates that the industry turned out between 2.5 and 5 billion items of excess stock in 2023 alone, representing somewhere between $70 billion and $140 billion in unrealised sales. 

The regulatory net is also tightening as I write: from 19 July 2026, the EU’s Ecodesign for Sustainable Products Regulation will prohibit the largest brands from destroying any unsold clothing, accessories or footwear, in response to the EU Commission’s finding that 4–9% of Europe’s unsold textiles are scrapped unworn, generating more than 5 million tonnes of CO₂ annually.

Knowing what to make, and having confidence that those predictions are based on the most accurate and complete picture possible, is a problem that’s approaching a new solution. But I see AI’s real contribution being judged in the harder, more consequential questions of how much of that new plan to commit to, and which upstream model to use. 

What predictive AI can, and can’t do

Before making a case for where they fall short, I want to be honest about where AI’s predictive capabilities have impressed me and the wider industry. 

Convolutional neural networks (CNNs), deep learning models well-suited to image data, can extract attributes from millions of social media images: silhouette, fabric, Pantone color code, and demographic signals. Transformer architectures, the family behind the most talked-about generative AI models, capture non-linear relationships between variables that traditional regression-based models miss. Generative AI itself can synthesise conversational and qualitative data from online sources and combine it with historical transaction data, producing richer forecasts than fashion has had before.

These are real gains. But they also have a clear ceiling the industry needs to be honest with itself about.

A consistent finding from my interviews: forecast accuracy is tightly correlated with forecast horizon. The shorter the window, the better AI performs. “The closer to the present, the better,” as the founder of a Portuguese fashion tech platform specializing in on-demand production put it. 

For one-month windows, particularly for core or carryover products with stable demand patterns, AI performs well with measurable gains. Stretch the window out to six, twelve, or eighteen months, and the picture changes fundamentally. At those horizons, demand prediction starts to look mathematically like a random walk: a process where each step is essentially independent of the last, meaning the future cannot be reliably extrapolated from the past. The CEO and founder of a UK-based AI supply chain platform for fashion put it more bluntly:

That, I think, is one of the most useful sentences I came across. AI is not giving us predictions. It is giving us faster, more granular signal detection. And signal detection, on its own, is not what reduces overproduction.

This is also why forecasters are not rushing to fire their human analysts and replace them with machine models. What comes out of AI is just part of a tapestry of contextual and cultural analysis that remains, at least for now, the preserve of people.

Data-backed bets: how much, not just what

The most consequential finding of my research is the distinction between predictive AI and optimisation AI.

Predictive AI tells you what is likely to sell. Optimisation AI helps you decide how much of that thing to make, when, through which factory, from what materials, and along what logistics path to ship it, and what channels to allocate it to. It also needs to take account of the non-trivial constraints you are operating under:minimum order quantities locked in with suppliers who don’t operate on a flexible model. Fabric availability and lead times. Factory capacity. Shipping windows. The same UK CEO put it more directly:

Fashion’s combinatorial complexity is also part of why this distinction matters, and why it’s been a folly – at least using traditional means – for fashion to try and hit targets that fundamentally different sectors can hit. 

The contrast with other industries, like consumer electronics, is stark: Apple ships six SKUs in a category, HP perhaps thirty, while a fashion brand ships thousands per season across colorways, sizes, and materials. Planning and producing at that scale without significant waste is mathematically a different kind of problem, one that spreadsheets and human intuition cannot solve at the necessary speed or precision. 

Machine-learning-driven optimisation, neural network-based dynamic production scheduling, and computer-vision-driven fabric cutting, on the other hand, are well-suited to it.

The implication runs further than the supply chain. As the same Portuguese founder put it, “The main value of a designer isn’t to design but to design in the most optimal way,” meaning creating in a way that accounts for costs, materials, and environmental impact up front. Optimisation, in this sense, is not a back-office concern to be handled by grizzled ops teams once the creative work is done. It is, increasingly, part of the essential creative work.

My argument here is straightforward. Predictive AI without optimisation AI leads to more accurate overproduction, in a climate where the commercial and legislative mandate for curbing excess are stronger than ever. 

The supply chain we built can’t keep up

If optimisation is one half of the under-told story, the other half is a tale of setup and structure. Even a perfectly optimised production decision is wasted if the supply chain cannot translate it into products on shelves in a timeframe before the value of the forecast has leaked out.This, more than any limitation of AI, is where I think fashion’s overproduction problem actually lives: we can get much better at deciding what to make, and much smarter about how to make it, but the limiting factor is the appetite and the ability of upstream actors to execute.

The co-founder of a German fashion trend forecasting and color intelligence platform articulated the disconnect for me:

Accurate forecasting, then, is necessary (it’s probably even essential, taking account of what fashion brands see when they look out over the parapets) but it’s not sufficient. The supply side of the equation must be incentives and able to respond at the right speed and scale – and most of it currently isn’t. 

The nine-to-ten-month industry calendar, described to me by an advisor at a Singapore-based AI think tank for fashion as the industry’s “bible”, opens up a wide window in which consumer preferences move around, material availability changes, and the original forecast diverges more from the changing conditions of reality.

Compounding this, brands typically place orders without real-time visibility into factory capacity or material availability: buying blindly, in effect, at the same time as placing big bets far in advance.

Three AI-enabled architectures emerged from my research, sharing a premise: shorter forecast horizons and faster feedback loops between consumer signals and production decisions.

The first is test-and-scale production. Under this model, rather than committing to large speculative runs, brands produce small initial batches, monitor consumer response through AI-driven signal analysis, and scale only the items that show genuine traction. The co-founder of a Turkish AI trend forecasting platform put this in concrete terms:

Shein is the most discussed embodiment of this model and also the most controversial. Litigating that company’s legacy is beyond the scope of this article, but I will say this: the majority of the waste Shein generates appears to be post-consumer, which suggests that, for better or worse, it’s making what the market is willing to actually buy. By the strict definition, Shein clearly has a waste problem, but it doesn’t have an overproduction one.

The second is nearshoring and reshoring. Machine learning models can help match demand signals to available local capacity, but a change in operations and orchestration has to go alongside them: producing closer to the consumer collapses the window during which the forecast is under external stress, and  reduces speculative volume, but AI is not going to address the work involved in brands lifting and shifting their sourcing models.

The third, and possibly the most underrated, is collaborative upstream planning. Instead of designing in isolation and placing large orders into a supply chain they cannot see into, manufacturers and suppliers share capacity and material data upfront, and brands design around it. Generative AI already enables early product previews and pre-order mechanisms, letting consumers and retail partners weigh in before any commitment. The shift here would be philosophical as much as technological: from design-then-source to design-around-resources.

Commercial and climate: AI cuts both ways

According to the UN Environment Programme, fashion accounts for up to 10 percent of global carbon emissions, more than international flights and maritime shipping combined. Of every ten designs a brand introduces, only one typically sells fully. 

These are not commercial inefficiencies that happen to have environmental consequences. They are environmental crises that happen to also be commercial inefficiencies. So while the initial catalyst for AI adoption, upskilling, and uptake will come from the bottom line, climate and humanitarian alarms are also sprinting for that same bottom line, and AI has the potential to at least slow them down.

The shifts I’ve described could, according to industry estimates, reduce overproduction by five to fifteen percent, with proportional reductions in water, chemical, and carbon footprints. That is not transformational on its own, but in an industry running up against Paris Agreement targets, it is meaningful. Combined with regulatory momentum behind initiatives like the EU’s Digital Product Passport, this kind of impact mitigation is rapidly becoming non-optional.

Here I want to be honest about something the industry tends to talk around, but that readers of this report have already seen brand founders tackle head-on. AI is not directionally neutral, and the same techniques that can reduce waste can also accelerate it. 

The danger is that AI deployed for trend acceleration, rather than waste reduction, generates more waste, more quickly, than the system it replaced. Automation is inherently appealing, because it allows us to take work-intensive processes and reduce the effort involved in them, but making fashion more “set and forget” without challenging the damage its dominant business model can cause is exacerbating a problem rather than addressing it.

The same machine learning techniques that let a brand test two hundred items and scale only the winners can also be used to test thousands of designs per day, accelerating the trend cycle and producing volumes of waste no optimisation model can recover from, and excess inventory on a scale that dwarfs even the most cynical projections from today.

There is also the environmental cost of the AI itself. Training and operating large generative models requires significant energy and computing resources, and multiple interviewees (as well as other writers for this report) raised this directly. It does not negate the case for AI in waste reduction, but the net environmental calculation depends entirely on what AI is deployed for, not just whether it is deployed.

What I take from this is simple. AI is a tool with a cost. Its impact on fashion’s sustainability footprint depends on the intent and discipline of its deployment, not on the technology itself. A brand using AI to produce less, more accurately, will reduce waste and will likely trim their overall net footprint, even when the impact of AI inference is taken into account. 

By contrast, a brand using it to produce more, more often, will increase it. The technology does not choose between these outcomes. We do.

Data is the difference

None of this is possible without the unglamorous foundation every AI capability depends on: data. This was the most consistent finding of my interviews, and the VP of a French creative AI studio captured it best:

The implications here for executives are concrete. Most fashion brands currently lack the data infrastructure to support the AI capabilities they are being sold, and the task of plugging that gap is unglamorous, expensive, but mandatory. 

Sales data, right now, is fragmented across channels; upstream supply chain data sits in different systems from downstream retail data, often in different organisations. Internal AI literacy is uneven, and the gap between what vendors promise and what organisations can operationalise remains the most under-discussed barrier in this industry. Successful AI adoption is, in my experience, as much organisational and cultural as technical.

Industry analysis projects a widening performance gap between AI leaders and laggards, with that data groundwork being one of the variables that makes the biggest difference. The BoF-McKinsey State of Fashion Technology Report estimates that brands embedding AI into their business models could see a 118 percent cumulative cash flow uplift by 2030, while those slower to invest face a 23 percent relative decline.

I’ll close, though, where I started: with the problem. Fashion simply produces too much, both commercially and environmentally, for that issue to be solved by sharper forecasting alone. 

Predictive AI will only ever do half the work: the visible, well-discussed half. The more meaningful other half is optimisation AI, paired with the supply chain flexibility that lets a brand actually act on what AI tells it, and supported by transformation projects that reflect the cultures they are intended to be embedded in. 

That is where the real waste reduction will come from. It is also where both the commercial and sustainability returns will compound for brands that take it seriously.

You May Also Like
Is There Such A Thing As An Authentic Synthetic Image?
Read More: Is There Such A Thing As An Authentic Synthetic Image?

Is There Such A Thing As An Authentic Synthetic Image?

Individual brands have their own idiosyncratic relationships with generated images, but fashion, on aggregate, has embraced the value of synthetic photography for eCommerce in particular - even in the face of cultural polarisation. What’s next?