The most effective sustainability technology in fashion is not a recycled fibre, a compostable polybag, or a certification badge on a product page. It is a demand forecast that is slightly less wrong than the one before it.
This sounds anticlimactic. It is also where the largest environmental gain in the industry sits, because fashion's defining waste problem is not disposal. It is production — specifically, the production of clothing that nobody ever buys.
A garment that is manufactured, shipped, warehoused, marked down twice, and finally liquidated has consumed every litre of water, every kilowatt of energy, and every kilogram of cotton that a sold garment consumes. It simply generated no revenue to justify any of it. This is the quiet centre of AI in sustainable fashion, and it is far less discussed than recycled polyester.
Overproduction Is the Emission
Every stage of a garment's environmental cost is incurred before a customer decides whether to buy it. Fibre cultivation, spinning, weaving, dyeing, cutting, sewing, and freight all happen upstream of the sale.
Which means the sustainability question is not really "how do we make this garment greener." It is "how do we avoid making the ones that will not sell."
Forecasting models attack this directly:
- Smaller initial production runs, with reorder capacity for whatever performs.
- Size curve optimisation by location, so a store is not sitting on extra-smalls it will never move.
- Allocation intelligence that shifts stock between locations before a markdown becomes necessary.
- Earlier exit signals, so a declining style is stopped rather than reordered out of habit.
The ultra-fast retailers demonstrated the commercial version of this — tiny test batches, aggressive reordering — and then applied it to accelerate consumption rather than reduce it. That is an uncomfortable truth worth stating plainly: the same technology that could shrink overproduction is currently being used to increase the number of styles offered. The tool is neutral. The business model is not.
The Waste on the Cutting Table
Before a garment exists, fabric is laid in long stacks and pattern pieces are arranged on it. Whatever falls between those pieces becomes offcut waste, and across an industry producing billions of garments, the fraction that ends up on the cutting-room floor is substantial.
Marker making — the arrangement of pattern pieces to maximise fabric yield — is a classic optimisation problem, and it is exactly the kind of task algorithms handle better than people. Modern nesting software evaluates enormous numbers of arrangements against constraints like grain direction, print matching, and fabric flaws.
The gains are measured in a few percentage points of yield per lay. That sounds trivial until you multiply it by the annual fabric consumption of a mid-sized brand, at which point it becomes one of the highest-return sustainability investments available — and it pays for itself in material cost alone, without any environmental argument needed.
Other cutting-room applications:
- Defect detection on incoming fabric rolls using computer vision, so flaws are mapped before cutting rather than discovered after.
- Offcut matching, identifying which remnants can be used for smaller components, trims, or accessory lines.
- Roll allocation, assigning specific fabric rolls to specific orders to minimise leftover partial rolls.
Sorting: The Bottleneck Nobody Sees
Textile recycling has a public image problem and a private technical one. The public believes clothing donated or dropped in a collection bin gets recycled into new clothing. In reality, the overwhelming majority does not, and the reason is sorting.
To recycle a textile into new fibre, you need to know its composition. A garment that is sixty per cent cotton and forty per cent polyester behaves entirely differently from pure cotton, and a single polyester sewing thread can contaminate a mechanical recycling stream. Care labels are unreliable, frequently missing, and often wrong.
This is where machine vision has produced a genuinely new capability. Systems using near-infrared and hyperspectral imaging can identify fibre composition from the material itself, at speed, on a moving conveyor. Companies including Refiberd, Matoha, and the Fibersort project have built on this approach.
The sequence that makes recycling economically viable looks like this:
- Identify fibre composition per item, without relying on labels.
- Detect and flag contaminants — zips, buttons, elastane content, coatings.
- Sort by stream into recyclable, reusable, and unusable categories.
- Route each stream to the recycler capable of processing it.
- Record the composition data so the recycler receives a characterised feedstock rather than a mystery bale.
Without step one working at industrial speed, textile-to-textile recycling remains a pilot project. With it, the economics start to make sense.
It is worth remembering how fragile these economics still are. The bankruptcy of the textile recycler Renewcell in 2024 was a hard reminder that technical capability and commercial viability are different things, and that demand from brands has to be real rather than announced.
The Dye House Is the Real Water Problem
Ask where fashion's water and chemical footprint concentrates, and the answer is dyeing and finishing. It is energy-intensive, chemically heavy, and notoriously imprecise.
The imprecision is the opening. Colour matching frequently requires multiple attempts, and every failed batch is a full cycle of water, heat, and chemistry discarded. Process optimisation models that predict the correct recipe for a given fabric, dye lot, and machine — getting the colour right on the first attempt more often — deliver savings that show up simultaneously on the utility bill and the environmental report.
Related applications include predictive maintenance on machinery, energy scheduling that shifts intensive processes to cleaner grid hours, and wastewater monitoring that catches a chemical excursion in real time rather than in a quarterly audit.
Data Is About to Become Mandatory
European regulation is moving toward digital product passports for textiles — a requirement that individual garments carry structured, verifiable data about materials, origin, and end-of-life handling.
Most brands cannot currently produce this data. They do not know, with certainty, which factory made which batch, which mill supplied the fabric, or what the exact composition is beyond what a supplier declared. Supply chains run three or four tiers deep and go dark somewhere around tier two.
Supply chain traceability is therefore becoming an AI problem by necessity:
- Extracting structured data from unstructured supplier documents, invoices, and certificates.
- Cross-checking declared compositions against test results and flagging inconsistencies.
- Mapping subcontracting relationships that were never formally disclosed.
- Scoring supplier risk on environmental and labour indicators rather than on price alone.
The brands treating this as a compliance chore will spend heavily and gain nothing. The ones treating it as an opportunity to finally understand their own supply chain will get better forecasting, better quality control, and lower risk as a side effect.
Where AI Makes Sustainability Worse
Any honest treatment of this subject has to include the counter-argument.
- Model training and inference consume energy. Not enormous relative to textile manufacturing, but not zero, and rarely counted in the brand's own reporting.
- It accelerates consumption. Better recommendation engines and faster trend detection sell more clothing. More clothing sold is more clothing made.
- It enables sophisticated greenwashing. A well-designed impact dashboard built on unverified supplier declarations produces confident, precise, and completely unreliable numbers.
- It concentrates advantage. Small producers without data infrastructure get scored badly by systems that mistake poor documentation for poor practice.
The fourth point deserves attention. A supplier scoring model trained on documentation quality will systematically penalise smaller factories in regions with less administrative capacity, regardless of what actually happens on their floor.
The Second Life Problem
Resale, repair, and rental are the parts of the industry that extend a garment's useful life, and they share a shared operational headache: every item is unique.
A traditional retailer sells the same style ten thousand times and describes it once. A resale platform receives ten thousand different garments, each requiring individual photography, categorisation, condition assessment, and pricing. That labour cost is the reason secondhand platforms struggle to be profitable, and it is a well-suited problem for machine vision:
- Automatic categorisation and attribute tagging from a single photograph.
- Brand and authenticity verification from stitching, labels, and hardware detail.
- Condition grading — pilling, fading, stains, seam wear — scored consistently rather than by whoever is on shift.
- Dynamic pricing based on what comparable items actually sold for, not on the original retail tag.
Repair services benefit from the same capability: diagnosing a fault from a photograph and quoting a price without a physical inspection removes most of the friction that stops people repairing clothes at all.
A Practical Checklist for Brands
If you want the environmental gain rather than the marketing line, the order of operations matters:
- Measure your sell-through honestly, including markdown and liquidation. This is your real overproduction number.
- Fix forecasting before buying anything sustainability-branded. It has the largest impact and the clearest payback.
- Invest in cutting yield. Material cost savings fund everything else.
- Get composition data right at the design stage, because you cannot retrofit it later.
- Reduce returns. A returned garment often carries more emissions than the original delivery.
- Verify supplier data rather than collecting it. Unverified data is worse than none, because it creates false confidence.
- Publish the awkward numbers, not only the flattering ones. Credibility compounds.
The Uncomfortable Conclusion
Technology can make each garment cheaper to produce, easier to trace, and simpler to recycle. What it cannot do is answer the question underneath all of it, which is how many garments the world should make.
AI in sustainable fashion shortens the distance between demand and supply. Whether a brand uses that to produce less waste, or to produce twice as many styles with the same waste, is a decision made in a boardroom rather than by a model.
The technology is ready. The business model is the part still under construction.
0 Comments