Separating real use cases from vendor hype

The rule: if a problem can be solved with a rule you could write on a whiteboard, ML is overkill. It adds cost, latency, and failure modes for no gain.

Stress test 1: catalog scale

Demos use a curated subset. Your catalog is the real test.

Ask the vendor: "Show me this running on 500 random SKUs I choose, including our worst product photos." A virtual try-on trained on clean studio images often breaks on:

• Flat-lay photos (garment laid on a table, no model).
• Patterned or sheer fabrics that confuse garment segmentationsegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → (the model separating the clothing from the background).
• Accessories, footwear, and non-standard shapes.

SKU (Stock Keeping Unit): a unique code for each product variant, for example one dress in size M, blue.

A mid-size apparel retailer can carry tens of thousands of active SKUs, refreshed seasonally. If a try-on tool needs manual 3D preparation per garment, the labor cost scales linearly and kills the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition →. Ask directly: cost per SKU to onboard, and how much is automated.

Stress test 2: data readiness

Data readiness (whether your data is clean, labeled, and accessible enough to train or run a model).

Most AI projects in fashion stall here, not on the algorithm. Run this checklist against your own systems before any pilot:

| Data asset | Ready? Ask yourself |

|---|---|

| Product images | Consistent angles and backgrounds? Or a mix from 50 suppliers? |

| Attribute tags | Is "color" filled for every SKU? Standardized values? |

| Return reasons | Do you capture "too small" vs "didn't like fabric"? |

| Purchase history | Linked to a stable customer ID across web and store? |

If half these cells are "no," a personalization or fit model has little to learn from. The vendor's demo works because their sample data was pristine. Yours may not be.

A quick way to check attribute completeness before a meeting:

If a third of your catalog has no color tag, an "AI stylist" that reasons about color coordination is starting blind.

Stress test 3: the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → math, honestly

Vendors quote upside ("try-on lifts conversion 30 percent"). Interrogate the baseline and the cost side.

Return rates in online apparel are high; commonly cited industry estimates put fashion e-commerce returns in the range of roughly 20 to 30 percent (estimate, varies by market and category, and higher for Europe in some segmentssegmentsDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition →). Reducing returns is often a bigger prize than lifting conversion, because returns carry shipping, handling, and restocking cost.

Worked example (illustrative, plug in your own numbers):

• Annual online apparel revenue: 50,000,000 USD
• Return rate: 25 percent, so 12,500,000 USD returned
• Vendor claims virtual try-on cuts returns by 2 percentage points (a realistic, modest claim; treat bigger claims with suspicion)
• 2 percent of 50,000,000 = 1,000,000 USD fewer returns
• If handling cost is 15 percent of returned value: 150,000 USD saved per year

Now the cost side:

• Onboarding 40,000 SKUs at, say, 3 USD each automated: 120,000 USD one-time
• Annual platform fee: 100,000 USD (illustrative)

Year-one net: 150,000 saved minus 220,000 cost = negative 70,000 USD. It only pays off if the return reduction holds and onboarding is automated. This is why the "cost per SKU" question decides the deal.

Always demand the claim be stated as incremental lift over your current baseline, measured by a proper test, not a vendor's other client.

Stress test 4: how will you prove it works?

Insist on a controlled test: randomly split traffic, show the AI feature to one group (treatment) and not the other (control), then compare. This is an A/B testA/B testA/B testing is a controlled experiment that compares two versions of something (A and B) by splitting traffic randomly to learn which performs better on a chosen metric.View full definition →.

Red flags in a vendor's evidence:

• "Conversion went up after launch." (No control group. Could be seasonality or a promo.)
• Case studies with no sample size or time window.
• Metrics that are easy to move but do not affect the business, for example "engagement with the try-on widget" instead of purchase or return rate.

A good vendor welcomes a holdout test and helps you design it. A hype-driven one resists measurement.

For a grounding in how these systems are actually evaluated, Google's People + AI Guidebook is free and non-technical.

Stress test 5: failure modes and trust

AI features fail in public. Plan for it before buying.

• Virtual try-on on a plus-size or non-standard body: does the garment warp unrealistically? A bad fit render can lose a sale and offend.
• AI stylist hallucinationhallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition →: a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → may confidently recommend a product you do not stock, or invent a fabric claim. HallucinationHallucinationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.View full definition → means the model generating fluent but false output.
• Bias: if training images skew toward one body type or skin tone, results degrade for everyone else. Ask the vendor how they tested across demographics.

Regulatory note for 2026: in the European Union, the EU AI Act (the EU's risk-based AI law, phasing in through 2025 to 2027) sets transparency duties. Systems that generate or manipulate images (a try-on that alters a user's photo) may trigger disclosure obligations, meaning you tell users the image is AI-generated. In the US there is no single federal AI law; the FTC (Federal Trade Commission) has warned it will act against deceptive AI marketing claims. Practical takeaway: a vendor claiming "no compliance concerns" without naming these is not paying attention.

Putting it together: a five-minute vendor script

Ask these in order. The pauses tell you as much as the answers.

1. "Could a rule or filter do this instead of ML?"

2. "Run it live on 500 SKUs I pick, including flat-lays."

3. "What is your fully automated cost per SKU to onboard?"

4. "State your lift as incremental over baseline, measured by A/B testA/B testA/B testing is a controlled experiment that compares two versions of something (A and B) by splitting traffic randomly to learn which performs better on a chosen metric.View full definition →."

5. "How did you test across body types, and what triggers under the EU AI Act?"

If a vendor handles all five with specifics, you likely have a real use case. If they redirect to the polished demo, you have hype.

Key Takeaways

• Rule-out ML first. If a whiteboard rule or a metadata filter solves it, AI is expensive overkill.
• Catalog scale and cost per SKU decide ROI, more than the model's cleverness. Demand a test on your messy catalog.
• Your data readiness, not the algorithm, is usually the bottleneck. Check attribute completeness before the pilot.
• Only a controlled A/B test proves lift. Reject "conversion went up after launch" as evidence.
• Name the regulation. The EU AI Act's transparency rules and FTC scrutiny of deceptive claims apply to try-on and stylist features in 2026.