What product data does fashion AI need?

One row per variant with size, size system, colour, material, fit, occasion and care for apparel, and metal, purity, weights, stones, hallmark and price build-up for jewellery, plus live stock and price per store and each garment's own size chart.

Is virtual try-on biometric processing under the GDPR?

Not if it only locates a body or face to place a garment and never identifies the person. Under GDPR Article 9 and EDPB guidance, processing to uniquely identify someone is special category data. Track without templates, keep processing on the device where possible, and complete a DPIA. This is not legal advice.

Which AI outputs need human approval?

Anything the public sees or that changes your systems: message templates, campaigns, social posts, design concepts, product images, price or rate changes and listing edits. Shopper answers should be grounded in live catalogue data rather than individually approved.

How should brands handle AI-generated models?

Get a release that names AI use, pay as for a shoot, let the model control the replica, label the result, and never turn a shopper's photo into a brand asset. H&M's digital twins set this pattern in 2025.

How should a brand choose an AI model?

Per job: test on your hardest products with your own reviewers, buy rather than build unless your data is the advantage, check whether the provider trains on your data and where it is processed, and count image-generation cost and energy.

Data foundations and governance for fashion AI

Before a fashion or jewellery brand chooses an AI model, it needs five foundations in place: product data clean enough for a machine to recommend from, shopper data collected only with clear consent, an approval step between anything AI drafts and anything a shopper sees, written rights for every image and every person in it, and a model choice made per job rather than once for the company. Every voice in our research points here. Investors call product data the bottleneck, platforms publish feed specifications, houses announce consent rules alongside their AI imagery, and researchers document the bias and accuracy limits that make review necessary.

This is part eight of our fashion AI research series and the second of three implementation guides. It follows the 90-day pilot plan; most of what is below should be settled in that plan's first month. It is a practical guide, not legal advice: take advice from your own counsel in each market.

What data does fashion AI actually run on?

Data	What it feeds	Who owns the problem
Product attributes, one row per variant	Stylists, search, external AI agents, tagging	E-commerce and merchandising
Live stock and price, per store	Anything that recommends or quotes	Operations and IT
Size charts per garment, and metal rates for jewellery	Size advice, pricing	Product and finance
What shoppers try, ask for and buy	Insight, range planning, design briefs	Retail and planning
Consented shopper profiles	Follow-ups, clienteling	CRM and privacy
Images, with rights and releases	Try-on, content, campaigns	Brand and legal

What does good product data look like?

The platforms have written the spec for you. OpenAI's product feed specification asks for one row per variant, nine required fields (including title, description, brand, image, availability and price), each size and colour as its own row under a group, a size system with no conversion implied, colour-matched images, material and gender, and reviews of the product rather than the store. Shopify's Agentic Storefronts and Google's Shopping Graph reward the same discipline. a16z put the principle in four words: AI shopping is "bottlenecked by content, not compute".

For a fashion brand, the fields that matter most to a stylist or an agent are the ones a good sales associate would mention:

Apparel: fabric and composition, weave or knit, fit (slim, regular, relaxed), length, neckline and sleeve, occasion, season, care, and the garment's own size chart in its own size system.
Jewellery: metal, purity and karat, net and gross weight, stones with their weights and grades, hallmark or certificate, ring and bangle size ranges, and how the price is built.
Both: accurate colour names, images that match the variant, and the plain-language reason someone buys it. Shopify tells merchants to make the brand "legible" to agents, because otherwise agents compare on price.

Two traps catch brands. First, size. Research from the University of Washington and Google shows image try-on does not show fit, and Zalando's controlled test found the measurable returns reduction came from size advice built on garment data and returns. Keep each garment's chart; one brand's M is another's L. Second, jewellery prices. A price built from a metal rate must have a rate that is current and confirmed. Ateliq, for example, prices from net weight, the day's rate and purity plus making charges and stones, has head office confirm the rate, and takes any piece off the screen, or shows "Price at the counter", if the rate is more than 48 hours old.

What consent and privacy rules apply to fashion AI?

The rules differ by market, but four questions decide most cases.

Does the camera identify anyone?

In Europe and the UK, the line is unique identification. Under GDPR Article 9, biometric data processed to uniquely identify a person is special category data, normally needing explicit consent. The EDPB's video guidelines distinguish a shop that detects a body for a purpose from one that recognises the same person again; the ICO takes the same view and expects a data protection impact assessment first. Try-on needs to find a body or a face; it does not need to know who it is. Track, don't template, and keep the processing on the device where you can. We cover the US state rules (Illinois, California, New York City) in privacy by design for in-store AI.

Does the shopper know they are talking to AI?

The EU AI Act's Article 50 sets transparency duties, including telling people when they are interacting with an AI system, and rules on marking AI-generated content. Dates have moved under the EU's AI Omnibus, so confirm timing with counsel; the safe design is to say so plainly at the start of every conversation.

Is the consent specific, and can it be withdrawn?

India's Digital Personal Data Protection Act, 2023 is now operational: the government notified the DPDP Rules in November 2025, with most obligations phased in over 18 months. Whatever the market, the pattern that works is the same: consent for one purpose at a time, confirmed by the shopper, logged without more data than needed, and easy to withdraw. Ateliq's follow-ups, for instance, go only to shoppers who opt in and confirm with a code, use approved templates, run at most twice a week and never between 9pm and 9am, and carry an unsubscribe link in every message.

What are you inferring?

The riskiest AI features infer things the shopper never said: body shape, weight, ethnicity, mood. Research shows image models default to a narrow look, with over 96% white-presenting people from neutral prompts in one 2026 audit and documented weight bias in another. Don't build what you can't defend. A useful rule: ask, don't guess. A stylist can ask the sizes someone wears; it should not estimate them from a camera.

Where do approvals belong?

Every voice in this series puts a person between AI drafts and the public. LVMH describes AI that supports client advisors; Hermès formed an AI governance committee; Anthropic's commerce agents blueprint stages merchant changes until a person approves; Shopify's campaign tools require merchant approval. Its guidance on building agents adds a reason: agents trade cost and compounding errors for autonomy, so start with the simplest workflow that works.

Set approval by what the output touches:

Output	Who approves	Rule
Answers to shoppers about stock, sizes and prices	Grounding, not approval	Only from live catalogue data; no invented prices or promises
Message templates, campaigns, social posts	Marketing lead	Approved before use; approving a post does not publish it
Design concepts	Design director	Drafts only; humans decide what is sampled
Product images and shoots	Brand and product teams	The piece itself is never changed
Price or rate changes	Head office or finance	Confirmed; large jumps get a second look
Anything written to the POS, ERP or listings	Named owner	Staged until approved, or not allowed at all

Ateliq is built this way: its twelve agents save drafts, reports and suggestions, none can show anything to a shopper or post in public, and it reads from a brand's systems without writing back to its POS or ERP. Whatever you use, write the approval map before the pilot starts, and log every approval.

How do you handle IP, likeness and imagery?

People in images. H&M's AI "digital twins" came with written consent, watermarks, the models' control of their replicas and standard pay; Zara asked models for consent and paid a standard fee. Brands that skipped this, from Levi's in 2023 to the Guess AI model in 2025, met a backlash, and New York's Fashion Workers Act took effect in June 2025. Get a release that names AI use, pay as for a shoot, label the result, and never turn a shopper's photo into a brand asset.

Training data and tools. Adobe says Firefly is trained on licensed and public-domain content and offers IP indemnity to enterprise customers. Whatever image tool you choose, ask what it was trained on, whether it trains on your uploads, and what indemnity you get. Attach Content Credentials where you can; Google has joined the C2PA and shows them in Search.

Your own designs. The US Copyright Office's January 2025 report concluded that prompts alone do not make a user the author, while human selection, arrangement and modification can be protected. Keep the designer's contribution substantial and documented. In the EU, showing concepts to the public can start the clock on design novelty, as the EUIPO explains. We go further in from demand signals to new designs.

Rights on every upload. Require a statement with each image: you own the photo, or you hold a release. It is a small step that stops most problems.

How should a brand choose AI models?

Choose per job, not once. The research gives four tests:

Fit for the task. a16z's Raspberry AI memo argues general image models "lack the domain-specific detail" apparel needs; academic work shows fine detail such as logos, buttons and engraving is the weak spot. Test on your hardest products, such as jewellery close-ups and print-heavy garments, with your own reviewers rather than benchmark scores.
Buy or build. Menlo found 76% of enterprise AI use cases were bought rather than built in 2025. Build only where your data is the advantage.
Data terms. Does the provider train on your data? Where is it processed? Can you delete it? For shopper data, these answers decide whether you can use the tool at all.
Cost and footprint. Image generation used over 60 times the energy of text generation in a 2024 study. Generate deliberately, and don't claim sustainability you have not measured.

Keep the option to switch. Open protocols such as Anthropic's Model Context Protocol and the commerce protocols built on it make it easier to change models without rebuilding your connections to stock, catalogue and CRM.

What should the governance checklist include?

A product data standard per category, with an owner and a weekly quality score.
A consent design per market: purpose, wording, confirmation, logging, withdrawal and deletion.
A data protection impact assessment for anything with a camera, before launch.
Set retention periods for events, profiles and consent records, tested.
A list of what the system must never infer, and a test that enforces it.
An approval map, as above, with named approvers and logs.
Image rights: releases naming AI use, rights statements on uploads, labels on AI imagery.
A model register: which model does which job, its data terms and its review results.
An incident route: who stops the system, who tells shoppers, who tells the regulator.

None of this is glamorous, and all of it decides whether the pilot in part seven can scale. It also decides what you can measure, which is the subject of the final part: how to measure the ROI of in-store and design AI.

Data foundations for fashion AI: consent, privacy, approvals, IP and model choice