EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Retail & E-commerce

AI for Fashion and Apparel

Fashion has the highest return rate in retail, the shortest product life, and a compliance requirement arriving at the end of the decade that is really a product data project starting now.

Tier 1
Our depth here
Returns
The margin problem
~2028/29
Textile DPP compliance

Return rates of a third or more are normal in apparel, and each return costs shipping both ways, handling, and frequently the margin on the sale. Reducing them is the highest-value AI work in the sector and it is mostly a data problem about garments rather than a modelling problem about people.

In one paragraph

AI for fashion and apparel covers size and fit recommendation, return reason analysis and reduction, visual and attribute search, short-lifecycle demand forecasting and allocation, markdown and full-price sell-through optimisation, and the structured product data that Digital Product Passport compliance will require.

Returns, and why fit is a garment data problem

Most fit models try to learn the customer's body from their purchase and return history. That helps, and the larger and more tractable variance is on the other side: your own garments are inconsistently sized between suppliers, between production runs and between styles that carry the same label.

Return reasonWhat actually causes itWhat fixes it
Too small / too largeGarment measurement inconsistencyActual measured garment specs, not label size
Not as picturedColour rendering and stylingConsistent photography and colour calibration
Bought multiple sizesCustomer has no confidence in your sizingFit guidance at the point of selection
Quality not as expectedFabric and construction description gapsMaterial and construction attributes stated
Changed mindGenuinely unavoidableThe floor you are working toward, not zero
Worth knowing

Capture return reasons properly or nothing else here works

Free-text or single-category return reasons are close to useless. A structured reason taxonomy — too small at the waist, colour differed, fabric thinner than expected — collected at the point of return is the input that makes every downstream fix possible, and it costs a form change rather than a project. Retailers who tell us returns are their biggest problem and cannot break them down by reason are describing a data collection gap, not a modelling gap.

Forecasting products that live for twelve weeks

Fashion forecasting breaks the assumption underneath most demand models: that the thing you are forecasting has a history. A seasonal style with a twelve-week life and no predecessor has to be forecast from its attributes and its comparables.

Attribute-based analogues, not time series

Category, silhouette, fabric, colour family, price point, brand tier and channel. The forecast comes from how comparable items performed, weighted by attribute similarity — which means your attribute data quality directly determines your forecast quality.

Early sales are the strongest signal you get

The first two weeks of actual sales tell you more than any pre-season forecast. The valuable system is the one that reforecasts fast and feeds allocation and repeat-buy decisions while there is still time to act on them.

Size curve is a separate forecast

Total demand and its distribution across sizes are different problems, and getting the size curve wrong is how you end up sold out of mediums and marking down extra-smalls. It varies by style, by channel and by store.

Markdown timing is worth more than markdown depth

Marking down two weeks earlier at a shallower depth usually beats a deeper cut later. That is an optimisation over your own historical markdown response, and it is one of the clearer wins in the sector.

The Digital Product Passport, which is a data project

Under the Ecodesign for Sustainable Products Regulation, textiles are expected to receive their delegated act in late 2027 with compliance following roughly a year and a half later, around late 2028. Footwear is excluded from the first wave. That sounds distant and it is not, because the underlying requirement is supply chain data you do not currently collect.

  • The hard part is upstream. Fibre composition, origin, processing and treatment data sit with suppliers who have never been asked for them systematically.
  • Per-item rather than per-style. Passport data attaches to products in a way most PLM and PIM systems were not designed for.
  • It overlaps with what you already need. Better material and construction attributes reduce returns, improve search and support honest sustainability claims. The compliance deadline is a reason to do work that pays anyway.
  • Green claims are enforced separately and sooner. Unsubstantiated sustainability claims are already a consumer protection exposure, and generated marketing copy is a common route into one.
  • Start the supplier data collection now. The lead time on getting hundreds of suppliers to provide structured data is measured in years, not quarters.
Process

How an engagement runs

Return reasons and product attributes first, because everything else in fashion depends on them.

Weeks 1 to 3

Returns diagnosis

Return reasons broken down properly, garment measurement consistency assessed, and the addressable share quantified.

Weeks 4 to 8

Product data foundation

Measured garment specs, structured attributes and imagery normalised across suppliers.

Weeks 9 to 14

Fit and search build

Size and fit guidance at the point of selection, and visual or attribute search over the enriched catalogue.

Weeks 15 to 20

Forecasting and markdown

Attribute-based analogue forecasting with fast in-season reforecasting, and markdown timing optimisation.

Ongoing

Operation

Return rate by reason, full-price sell-through and forecast accuracy tracked against control.

Deliverables

What you receive

Fewer returns, better full-price sell-through, and product data that will still be useful in 2029.

01

Return reason taxonomy and analysis

Structured reasons captured at the point of return, with the addressable share quantified.

02

Garment measurement consistency review

Where your own sizing varies between suppliers, runs and styles.

03

Size and fit guidance

At the point of selection, measured on return rate rather than engagement.

04

Attribute and visual search

Over a normalised catalogue, which also serves marketplace and agent channels.

05

Short-lifecycle forecasting

Attribute-based analogues with fast in-season reforecast and a separate size curve.

06

Markdown timing optimisation

Optimised on timing as well as depth, against your own historical response.

Fit check

Is this the right starting point?

Worth being direct. There are situations in fashion and apparel where custom AI work is the wrong spend, and those are listed rather than buried.

Worth doing if

  • Return rates are a material margin problem and you cannot break them down by reason.
  • Your own sizing is inconsistent between suppliers or production runs.
  • Seasonal products are forecast from history that does not exist.
  • Markdown decisions are made on a calendar rather than on demand response.
  • You need textile product data for Digital Product Passport readiness and have not started.

Do something else if

  • Return reasons are not captured in a structured form. Fix the form first; it is cheap.
  • Product attributes are inconsistent and there is no appetite to normalise them.
  • The real problem is product design or quality, which no model addresses.
  • You want generated sustainability claims. That is a consumer protection exposure.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

What actually reduces returns in apparel?

Measured garment specifications and honest fit guidance at the point of selection, in that order. Most fit work focuses on learning the customer's body, and the larger tractable variance is usually your own sizing inconsistency between suppliers, runs and styles. Before any of that, capture return reasons in a structured taxonomy — a retailer who cannot break returns down by reason has a data collection gap rather than a modelling problem.

How do we forecast a product with no history?

From attributes and comparables — category, silhouette, fabric, colour family, price point and channel — weighted by similarity to items that have sold before. Which means your attribute data quality sets your forecast quality. Then reforecast hard on the first two weeks of real sales, because those tell you more than any pre-season number and there is still time to act.

When do we actually need Digital Product Passport data?

The textiles delegated act is expected in late 2027 with compliance roughly eighteen months later, around late 2028, and footwear is outside the first wave. The date is less relevant than the lead time: fibre composition, origin and processing data sits with suppliers who have never been asked for it systematically, and getting hundreds of them to supply structured data takes years. The same data also reduces returns and improves search, so it pays before the deadline arrives.

Is visual search worth building?

In fashion, more than in most categories, because customers frequently know what something looks like and not what you call it. It works best over a well-attributed catalogue, which means the attribute work comes first and delivers value on its own. Measure it on conversion from visual search sessions against a holdout, not on usage.

Should we optimise markdown depth or timing?

Timing, usually. Marking down two weeks earlier at a shallower depth commonly beats a deeper cut later, and the analysis runs on your own historical markdown response rather than requiring anything exotic. It is one of the clearer wins in fashion and it is frequently blocked by a markdown calendar that exists for planning convenience rather than commercial reasons.

Tell us what the problem looks like.

Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.