Leaders Insights
Leaders Insights

Rester au meilleur niveau, un peu chaque jour.

DomainesMarketingDataFinanceIA
RessourcesApprendreTestOutilsBlogGlossaire
© 2026 Leaders Insights — Tous droits réservés.
Formations/AI in fashion/AI in fashion/Optimizing size, fit, and inventory to cut returns and waste
3/4+150 XP

AI in fashion

1Forecasting trends with AI: from runway signals to demand curves+1502Generative design and AI-assisted collection development+1503Optimizing size, fit, and inventory to cut returns and waste+1504Hyper-personalization and AI-powered supply chain traceability+150

Optimizing size, fit, and inventory to cut returns and waste

# Optimizing size, fit, and inventory to cut returns and waste

A customer orders the same dress in three sizes, keeps one, and ships two back. Multiply that by millions of orders and you have the single most expensive problem in online apparel: returns. In some fashion categories, return rates climb toward 40%, and much of that is driven by one thing: shoppers guessing at fit.

Zalando, Europe's largest online fashion platform, tackled this head on. By building fit prediction into the buying journey (recommending a size based on what actually fit similar bodies), the company reduced size-related returns on participating products. Fewer returns means less shipping, less repackaging, less landfill, and higher margin. This lesson shows you how that machinery works and how to deploy it.

Why returns are a fashion-specific crisis

Returns are not free. Each returned garment carries reverse logistics cost (shipping back), inspection labor, repackaging, and markdown risk because a returned item may miss its selling window. Industry estimates put the cost of processing a single return at a meaningful fraction of the item's price, sometimes enough to erase the profit on the sale entirely.

There is also a sustainability cost. A share of returned apparel never gets resold. It is discounted, liquidated, or in some cases destroyed. Cutting returns is therefore both a margin play and a waste-reduction play, which is why it now sits on the CFO's and the Chief Sustainability Officer's desks alike.

The root causes split roughly into two buckets:

  • Fit uncertainty: the shopper does not know their size in this brand or cut.
  • Bracketing: the shopper deliberately buys multiple sizes intending to return some.

AI attacks both.

The three levers

Lever 1: Body-scan and measurement data

The most direct way to recommend size is to know the body. Several approaches exist, ranging from low friction to high accuracy:

  • Self-reported measurements: height, weight, age, bra size, usual fit preference. Cheap, but noisy.
  • Photo-based scanning: apps that estimate body measurements from two smartphone photos using computer vision. Companies like 3DLOOK and Bold Metrics work in this space.
  • In-store 3D body scanners: booths that capture precise measurements, used more in made-to-measure and uniform businesses.

The catch: body data is sensitive personal data. Under the EU's General Data Protection Regulation (GDPR), and depending on interpretation, body scans can qualify as biometric data requiring explicit consent and careful handling. Build privacy in from day one: minimize what you store, get clear consent, and let users delete their profile.

Lever 2: Size recommendation engines

Most large platforms do not rely on body scans alone. The workhorse is a recommendation engine trained on purchase and return history. The core insight: you learn a garment's true fit from how it behaved on thousands of past customers.

Zalando's approach, built partly on its acquisition of the fit specialist Fits.me and later research teams, models both the customer and the article in a shared "size space." If customers who bought Article A and kept it also tend to keep Article B in the same size, the system infers those articles fit similarly.

A simplified version of the logic:

python
# Pseudo-logic for a size recommendation signal
# For a given customer + candidate article:

def recommend_size(customer, article, purchase_history):
    # 1. Estimate customer's latent "true size" from kept items
    true_size = infer_true_size(purchase_history)  # items kept, not returned

    # 2. Estimate this article's fit bias (runs small / large / true)
    fit_bias = article.fit_offset  # learned from aggregate return reasons

    # 3. Adjust and score each available size
    scores = {}
    for size in article.available_sizes:
        expected_fit = true_size + fit_bias - size_to_scale(size)
        scores[size] = probability_of_keep(expected_fit)

    return max(scores, key=scores.get)

The critical training signal is the return reason code. When a customer returns an item and selects "too small" or "too large," that label is gold. It tells the model which direction a garment's fit is biased. Retailers that capture granular return reasons train far better fit models than those who only see "returned: yes/no."

Lever 3: Demand-driven size and store allocation

Fit prediction reduces returns per order. Allocation reduces a different waste: stock in the wrong place.

Classic buying orders each style in a fixed size curve (say 10% XS, 20% S, 30% M, 25% L, 15% XL) applied uniformly. The problem: a beach store in one region sells a different size mix than a city store in another. Uniform curves guarantee you sell out of the popular middle sizes while XS and XXL sit unsold, get marked down, and eventually become waste.

Demand-driven allocation uses local sales history to forecast size demand per store or per fulfillment node, then allocates stock to match. The payoff:

  • Fewer broken size runs (a style is "broken" when the popular sizes sell out and only odd sizes remain, killing full-price sell-through).
  • Fewer markdowns.
  • Less inter-store transfer cost.

For deeper reading on how forecasting and allocation connect, McKinsey's State of Fashion reports track how leading players are operationalizing this.

🎬 [VIDEO: "How Zalando Uses Machine Learning" — youtube.com — Zalando engineers explain size recommendation and personalization at scale]

Making it work operationally

Technology is the easy half. Here is where deployments succeed or fail.

Coverage matters more than accuracy. A brilliant recommendation shown on 5% of products barely moves the return rate. Prioritize getting a decent recommendation on the widest possible catalog. A "runs small, size up" nudge on every product often beats a precise scan on a handful.

Placement in the journey. The recommendation must appear at the moment of size selection, not buried in a tab. A single visible line ("We recommend size M based on your profile") drives adoption.

Handle the cold start. New customers and new products have no history. Use self-reported measurements and category-level fit priors until enough data accumulates. New articles can inherit fit bias from similar past styles (same brand, same fabric stretch, same block).

Close the loop with return reasons. Make the return flow capture *why*. Structured reasons ("too tight in shoulders," "length too long") feed the model and, crucially, feed back to design and buying so recurring fit flaws get fixed at the source.

Watch bracketing behavior. Some retailers charge for returns or flag serial bracketers. This is commercially and reputationally delicate. Better fit prediction is the positive-sum move: help people order right the first time rather than penalizing them.

Vérification des acquis

1. Why are returns considered a fashion-specific crisis rather than just a routine cost of doing business online?

2. A retailer notices that a customer routinely orders the same item in multiple sizes and returns all but one. Which root cause does this behavior illustrate?

3. Why does the lesson frame cutting returns as both a margin play and a waste-reduction play?

CHOIX MULTIPLES

4. Select ALL correct answers. Which costs are described as making a single return expensive?

Sélectionnez toutes les réponses correctes.

CHOIX MULTIPLES

5. Select ALL correct answers. Which statements accurately reflect how fit prediction reduces returns?

Sélectionnez toutes les réponses correctes.

Measuring impact honestly

Do not claim a win from a vanity metric. Track:

  • Return rate by reason: isolate size and fit returns, since those are what fit tech influences. A drop in "changed my mind" returns is not your model's doing.
  • Keep rate on recommended size: did customers who followed the recommendation actually keep the item more often?
  • Full-price sell-through by size: proof that allocation is reducing broken runs and markdowns.
  • Net margin per order, not just returns, because a scheme that suppresses returns but also suppresses sales is a loss.

Run it as a proper experiment. Show the recommendation to a random test group, withhold it from a control group, and compare. Correlation between "users of the fit tool" and "lower returns" is misleading, because engaged shoppers behave differently anyway. Only a controlled test tells you the true lift.

The waste dividend

Every avoided return is roughly one avoided round trip of shipping plus the emissions and packaging that go with it. Every better-allocated size curve means fewer garments marked down to clearance or liquidated. For a mid-size online retailer, shaving even a few points off the size-related return rate can translate into meaningful cost savings and a genuine sustainability story you can report, provided you measure it rigorously rather than estimate it generously.

Key Takeaways

  • Returns are the margin and waste crisis of online apparel, and fit uncertainty is the biggest fixable driver. Zalando cut size-related returns by building fit prediction directly into the buying journey.
  • Return reason codes are your most valuable training data. Capture granular reasons ("too small," "shoulders tight") to teach both the model and your design and buying teams.
  • Coverage beats precision. A simple, visible size nudge across the whole catalog moves returns more than a perfect body scan on a few products.
  • Demand-driven size allocation matches stock to local demand, cutting broken size runs, markdowns, and waste that uniform size curves create.
  • Prove impact with a controlled test, measuring return rate by reason and net margin per order, not vanity metrics or correlations.

Précédent

Generative design and AI-assisted collection development

Suivant

Hyper-personalization and AI-powered supply chain traceability