A chat box is not a conversational-commerce strategy. It is one possible interface to a much larger operating system: product discovery, inventory, policies, customer context, cart state, checkout, and post-purchase service.
Fashion retailers that begin with the interface often create a polished assistant that answers easy questions and fails at the moment that matters. It recommends a sold-out size, cannot explain a return exception, loses the selected color during checkout, or invents an answer when the catalog is incomplete.
The practical way to build is from the transaction backward.
The six layers
1. Intent capture
Fashion requests are frequently contextual: an event, climate, dress code, body preference, existing wardrobe, delivery deadline, or care constraint. The assistant should preserve those constraints across turns and distinguish requirements from preferences.
Do not convert every phrase into a permanent customer profile. Ask for consent before storing sensitive or durable preferences, and make it possible to shop without personalization.
2. Catalog retrieval
The retrieval layer should query current product and variant data, not a frozen copy of marketing pages. It needs structured filters for price, availability, size, color, material, fit, category, and fulfillment conditions, followed by a relevance step that can explain why a result matches.
OpenAI says its 2026 shopping discovery uses merchant product information supplied through the Agentic Commerce Protocol and other providers. Shopify says its own UCP tools can search either a global catalog or a single merchant’s storefront catalog. The implementation details differ, but both approaches depend on a usable catalog.
3. Policy service
Shipping, returns, exchanges, pickup, discounts, and final-sale rules must be retrievable as current policy objects. A general language model should not improvise an exception. When the system cannot determine the answer, it should link to the controlling policy or hand the question to staff.
4. Cart state
A conversation can last several turns. The cart therefore needs a durable state containing exact variant IDs, quantities, localization, promotion eligibility, and estimated totals. Shopify’s UCP documentation describes carts that can be updated as buyers iterate and then converted into a merchant checkout.
5. Checkout and consent
The customer should see the selected product, variant, quantity, price, shipping, taxes or fees where known, and merchant identity before committing. Any agent action that changes money or inventory needs an explicit confirmation boundary.
Some channels complete payment inside their experience; others send the shopper to the merchant. OpenAI’s March 2026 product-discovery update said it was emphasizing merchant-controlled checkout experiences while continuing to support deeper integrations and apps. Retailers should design for both native and handoff paths rather than assume one universal checkout.
6. Order and service state
The same conversational layer should be able to retrieve a fresh order status, explain the next legitimate action, and escalate when a return, address change, or exception requires staff. Shopify’s developer documentation places order monitoring after discovery, cart, and checkout in the UCP buyer journey.
What a small retailer can build first
An independent boutique does not need to create a general-purpose shopping agent. A useful first release can answer five high-value question types:
- Which available items match a stated occasion, budget, size, and material preference?
- What are the verified measurements and fit notes for this exact variant?
- Can this item arrive at a given destination by a requested date?
- What return or final-sale rule applies to this item?
- Can the assistant prepare the correct variant in a cart and hand it to checkout?
Everything else can fall back to conventional navigation or a human.
A minimum architecture
Source systems: commerce platform, inventory, product information, order management, policy documents, and approved editorial content.
Commerce services: product search, variant lookup, cart mutation, checkout creation, order lookup, and escalation.
Conversation layer: intent and constraint tracking, tool selection, answer composition, uncertainty handling, and confirmation.
Control layer: authentication, permissions, rate limits, event logging, data retention, fraud controls, and human review.
Measurement layer: query class, products considered, reasons surfaced, handoff success, checkout completion, escalation rate, correction rate, and customer feedback.
The assistant should never receive more authority than the action requires. Reading public product data is different from accessing an order. Adding an item to a cart is different from placing an order.
Test cases before launch
- A requested size sells out during the conversation.
- The product page and feed disagree on price.
- Two colors share an image incorrectly.
- A final-sale item is added after the return policy was explained.
- The delivery deadline cannot be guaranteed.
- A shopper changes currency or destination mid-conversation.
- A promotion applies to only one variant.
- The shopper asks the assistant to complete a purchase without reviewing the total.
- The assistant cannot find a policy answer.
- A returning customer asks to delete stored preferences.
Pass conditions should be explicit. “The response sounded helpful” is not enough.
Build the boring services first
Conversational commerce works when the invisible systems are dependable. The catalog must be current. Policies must be queryable. Cart and checkout state must be deterministic. The assistant must know when not to answer.
The most credible fashion shopping experience will not be the one with the most human-sounding chat. It will be the one that preserves intent, shows its evidence, and hands the customer to a correct transaction.
Sources and verification
- Shopify Developers: Build commerce agents with UCP, accessed August 31, 2026.
- Shopify Developers: Carts and checkout for agents, accessed August 31, 2026.
- Universal Commerce Protocol, specification site accessed August 31, 2026.
- OpenAI: Powering Product Discovery in ChatGPT, March 24, 2026.
- Shopify Help: Shopify agentic storefronts, accessed August 31, 2026.