Launching an AI shopping assistant: a 20-point checklist for fashion e-commerce
From product data and guardrails to EU AI Act transparency and measurement, the 20 checks fashion e-commerce teams should complete before an AI shopping assistant goes live.
KEY TAKEAWAYS Summary by the editors
- An AI shopping assistant is only as reliable as the product, stock and policy data it can retrieve, so data readiness should be checked before model choice.
- Under Article 50 of the EU AI Act, applicable from 2 August 2026, AI systems that interact with people must be designed so users know they are dealing with an AI unless this is obvious.
- A Canadian tribunal held Air Canada responsible in 2024 for incorrect information given by its website chatbot, rejecting the argument that the chatbot was a separate entity.
- Assistants should be launched with defined scope, escalation to human service and logging, and tested against realistic fashion questions on size, fit, materials and returns.
- Success should be measured with control groups on conversion, order value, returns and contact rate, not only on chat volume.
Before launching an AI shopping assistant, a fashion e-commerce team should confirm five things: the assistant reads accurate product, stock and policy data; its scope and guardrails are defined; legal transparency duties are met; it has been tested against real customer questions; and its business impact can be measured against a control group. The 20 points below cover those areas.
Why does a fashion shopping assistant need a checklist?
Shoppers are already using AI for shopping. Adobe reported that generative AI traffic to US retail sites grew 693% year on year in the 2025 holiday season. Fashion retailers are responding with their own assistants: Zalando said in March 2026 that its assistant had six million users. But an assistant that invents a fabric composition or a returns rule creates commercial and legal risk. In the 2024 Moffatt v Air Canada case, a British Columbia tribunal held the airline responsible for incorrect information from its website chatbot.
The checklist is grouped into five areas, each with an owner. Not every point needs a large project; many are decisions to document. But skipping an area tends to show up quickly after launch, usually as wrong answers about sizes or returns, unclear accountability when something goes wrong, or results that cannot be proven to management. Teams that complete the list before go-live also find it easier to expand the assistant to new markets later, because the same structure can be reused.
Is your data ready? (points 1 to 5)
Data gaps are the most common reason assistants fail in fashion. Before choosing a model or vendor, run a simple audit: take the top categories by revenue and check what share of styles has complete composition, fit, care and size guide data in every launch language. If the share is low, fixing it will improve the assistant, product pages and external feeds at the same time.
- Product attributes are complete for the categories in scope: size, size system, colour, material, fit, care and gender.
- Stock is available to the assistant at size level, with a defined refresh frequency.
- Prices and promotions come from the same source as the checkout, so the assistant cannot quote an expired discount.
- Delivery, returns and exchange policies are available as structured, current text the assistant retrieves rather than paraphrases from memory.
- Size guides and fit notes are digitised per style or block, not only as images.
Are scope and guardrails defined? (points 6 to 10)
- Written scope: which questions the assistant answers and which it declines, such as medical, legal or payment disputes.
- Answers on product facts are grounded in retrieved catalogue data, with an instruction to say 'I don't know' when data is missing.
- Prohibited outputs are listed: invented discounts, unverified sustainability claims, comments on body shape.
- A clear hand-off to human customer service, with conversation context passed on.
- Rules for competitor products, out-of-stock substitutes and price comparisons are agreed with commercial teams.
Which legal and compliance checks apply? (points 11 to 14)
Requirements differ by market, so these points should be reviewed with legal counsel.
- Transparency: Article 50 of the EU AI Act, applicable from 2 August 2026, requires systems interacting with people to make clear that users are dealing with an AI unless obvious. Label the assistant at the first interaction.
- Consumer law: the EU Unfair Commercial Practices Directive covers untruthful information before, during and after a transaction, and its amendment by Directive (EU) 2024/825 applies from 27 September 2026. Product statements by an assistant should be treated like any other marketing claim.
- Data protection: document what conversation data is stored, for how long, whether it trains models, and how personal data such as body measurements is handled.
- Accessibility: the assistant should work with keyboards and screen readers, consistent with Article 50's requirement that information meets accessibility requirements.
How should the assistant be tested before launch? (points 15 to 17)
- Build a test set of several hundred real customer questions from service logs, covering size, fit, materials, care, delivery and returns, in every launch language.
- Run adversarial tests: attempts to obtain discounts, inappropriate content, prompt injection through product reviews or user input.
- Have merchandisers and customer service review answers for factual accuracy and brand tone, and set an accuracy threshold for go-live.
How will you measure success? (points 18 to 20)
- Launch as an A/B test with a holdout group so that conversion, order value and returns can be compared fairly.
- Track operational metrics: share of conversations resolved, escalation rate, answer accuracy in sampled reviews, and latency.
- Set a review cadence and an owner: who reads conversation samples weekly, who updates content and who can switch the assistant off.
| Area | Points | Typical owner | Go-live blocker if missing |
|---|---|---|---|
| Data | 1 to 5 | E-commerce, PIM, merchandising | Yes |
| Scope and guardrails | 6 to 10 | E-commerce, customer service | Yes |
| Legal and compliance | 11 to 14 | Legal, data protection officer | Yes |
| Testing | 15 to 17 | Product, QA, customer service | Yes |
| Measurement | 18 to 20 | Analytics, e-commerce | No, but needed within weeks |
What happens after launch?
Assistants degrade when catalogues, prices and policies change faster than the content they retrieve. Treat the assistant as a product with a backlog: review failed conversations, add missing attributes to the product information system, and re-run the test set after every model or prompt change. As assistants from search engines and AI platforms also start answering questions about your products, consistent product data across your own assistant and external feeds becomes even more important.
Finally, plan communication. Customer service teams need to know what the assistant can and cannot do, so they can handle escalations smoothly, and store staff should be briefed if the assistant also answers questions about in-store availability. Internally, share early results honestly, including failures, so that expectations stay realistic and the assistant is improved rather than quietly abandoned when the novelty fades.
Frequently asked questions
Do I have to tell customers they are talking to an AI?
In the EU, Article 50 of the AI Act, applicable from 2 August 2026, requires AI systems that interact with people to be designed so users know they are dealing with an AI, unless this is obvious. Clear labelling at the first interaction is the safest approach.
Is a retailer liable for what its chatbot says?
Courts have treated chatbot statements as the company's own. In Moffatt v Air Canada (2024), a British Columbia tribunal held the airline responsible for incorrect fare information from its chatbot. Local law varies, so seek legal advice.
What data does a fashion shopping assistant need?
At minimum: complete product attributes, size-level stock, live prices and promotions, structured delivery and returns policies, and digitised size guides. Without these, the assistant will either refuse to answer or risk making things up.
How do I measure if an AI shopping assistant works?
Use an A/B test with a holdout group and compare conversion, order value, return rate and customer service contacts. Add sampled accuracy reviews and escalation rates to monitor answer quality.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- EU Artificial Intelligence Act: Article 50, Transparency obligations
- American Bar Association: BC tribunal confirms companies remain liable for information provided by AI chatbot
- European Commission: Unfair Commercial Practices Directive
- Marketing Dive: AI has changed holiday shopping, here's what the numbers say
- Retail Gazette: Zalando reports strong 2025 growth as AI strategy drives performance
