Superintelligence for Fashion Retail: Catalog, Sizing

Superintelligence for fashion retail, today, means more capable AI agents that talk to customers, suggest pieces and check stock. It does not mean an intelligence above human level. What decides whether these agents work is the foundation the store prepares: a catalog with clear attributes, measurements for each garment, written returns rules and stock that can be queried. Without that, a better agent simply makes mistakes with more confidence.

Key takeaways

• A more capable agent does not make up for bad data: if the size chart is wrong or missing, it will confidently answer with something nobody verified.
• The catalog needs structured attributes (fabric, drape, fit, true color, care), not only a marketing description.
• Size is the most sensitive point: the agent should check the garment's measurements and recommend with a caveat, never promise how something will fit.
• Returns rules must be written with clear exceptions, and cases outside the rule go to a person.
• Stock and collection status are checked at the moment of the answer, never remembered by the agent. The pilot starts with a single flow, with decisions logged.

What does superintelligence mean for a fashion store?

The term tends to create the wrong expectation. For a store, what exists and can be operated now are language models that converse well, read a catalog, call tools and follow long instructions. That is already enough to change customer service. But it is different from the idea of an intelligence that decides on its own what the company should do. If you want to separate these concepts clearly, read superintelligence myths that get in the way of business decisions.

For this article, one reading rule helps: what comes from vendors' official documentation is technical fact, what comes from implementation experience is practice observed in projects of this kind, and what is projection is a hypothesis. The advice that agents work better with well-described tools and simple flows appears in public material, such as Anthropic's guide on building effective agents. Applying it to fashion, with its particular size and returns issues, is an implementation reading, not a measured result.

What catalog data does the agent need so it does not invent answers?

In fashion, the photo sells and the description decorates. The agent, however, can only answer with what is written or structured. If the question is "does this fabric wrinkle?" and the catalog only says "versatile, elegant piece", the agent has two options: admit it does not know or improvise. The store's job is to make the first option possible and the second unnecessary.

A minimum checklist of attributes per product:

• A unique code for the piece and for each variation (color and size), identical to the one in the stock system.
• Fabric composition, and whether there is elastane or lining.
• Fit: slim, regular or loose, and whether the manufacturer considers it to "run small" or "run large".
• Length, waist, bust, hip and other measurements of the finished garment, in centimeters.
• Height and size worn by the model in the photo.
• True color in words, because the photo changes with light and screen.
• Washing care, and whether the piece can be washed at home.
• Occasion of use only when the brand has actually defined it (for example, an evening line).
• Collection status: permanent line, current collection, last unit or discontinued.

You do not need to fill in everything before starting. Pick the best-selling and most-asked-about categories, complete these attributes and leave the rest out of the agent's scope. One well-described category is worth more than a catalog that is half done.

The most common technique for the agent to find this information is retrieving data before answering, known as RAG (retrieval-augmented generation). In practice, it means the agent checks the catalog for each question instead of relying on what it "knows". When the data does not exist, the instruction should tell the agent to say so and offer a human attendant.

How do you handle sizes and measurements without promising fit?

Size is where fashion customer service generates the most exchanges, and where an overconfident agent does the most harm. Sizing varies between brands, between collections of the same brand and even between fabrics. An agent that says "the M will fit you" without a basis is gambling with the store's money.

The safest approach has three parts:

• Garment measurements, not just the label. The agent compares the measurements the customer gives with the actual measurements of the piece and explains the expected ease.
• Recommendation language, not guarantee language. "Based on the measurements you shared, the M should be comfortable and the L would be looser" is honest. "It will fit perfectly" is not.
• A context question. If the customer prefers a close fit or a relaxed one, the agent should ask before recommending.

There is also a personal data question. Body measurements, photos and health information tied to the body deserve care. Ask only for what is needed to recommend a size, say what the information will be used for and do not keep more than the flow requires. For official guidance on data protection in Brazil, the starting point is the national data protection authority (ANPD). If the store has doubts about the legal basis, review it with whoever handles LGPD at the company before the pilot.

What should the returns and exchange rules say?

A returns rule is the kind of content where imprecision is expensive. If the agent states a deadline or a condition the store does not honor, the customer has the conversation record in hand. That is why the policy needs to exist as short, unambiguous text, kept in a single place.

For each situation, write three things: the rule, what the agent can do alone and what requires a person. A simple decision guide:

• Size exchange for an unworn piece with tags: the agent explains the procedure and opens the request, if the store allows it.
• Piece with a visible defect: the agent collects photos and the order number and hands off to the team, without promising an outcome.
• Piece on sale, personalized or from a special line: the agent cites the store's specific rule and, if in doubt, hands off.
• Request outside the deadline or standard: the agent does not negotiate, it only logs and forwards.
• Angry customer or one threatening a public complaint: immediate handoff with a summary of the conversation.

The line between automating and keeping people involved is also covered in AI customer support: what to automate and what to keep human. The same logic applies here: the agent resolves what has a clear rule and passes on what depends on judgment or a commercial exception.

How should the agent check stock and collection status?

An agent should not "remember" that a model exists in a certain color. It should check the system at the moment of the question, through a tool with read-only permission. This avoids two common errors: offering what sold out and hiding what just arrived. Anthropic's guide on writing effective tools for agents reinforces that tools with clear names and clear return values help the model use them well.

Some practical decisions:

• The stock tool returns quantity per variation and the date of the last update, so the agent knows whether the data is fresh.
• Holding a piece for a customer is a separate action, with a time rule set by the store and confirmation from the customer.
• Pieces from a previous collection or a discontinued line carry a marker in the catalog, so the agent does not sell them as new.
• If the physical store's stock and the e-commerce stock are different, the agent says which one it is checking.

When service happens on WhatsApp, the channel's rules also come into play. The conversation has time limits for free-form replies, and outbound messages depend on an approved template and on the customer's permission. Read the WhatsApp 24 hour customer service window and how an AI agent should act before designing alerts such as "the piece you wanted is back in stock".

Illustrative scenario: a dress for a wedding

This is an illustrative scenario, unrelated to any real customer. A customer writes on WhatsApp asking for a midi dress for a daytime wedding and asks whether size M will fit, sharing her bust, waist and hip measurements.

The agent searches the catalog, filtering by the occasion defined by the brand and by length. It finds two models. For the first, it compares the measurements she gave with the garment's measurements and answers that the M should be snug at the hip, and the L more comfortable but loose at the waist. For the second, which has elastane, it says there is more room for adjustment. Before suggesting a color, it checks stock and reports that the L of the second model has only a few units left, with the time of the last update.

The customer asks whether she can exchange it if it does not fit. The agent cites the store's written rule, without adding a deadline of its own. She then asks for a discount for buying two pieces. That is outside what the agent can decide: it logs the request, summarizes the conversation and hands off to a sales associate, who picks up where it left off.

In this scenario, the agent did not invent a measurement, did not promise a fit, did not improvise a policy and did not negotiate a price. The quality of the answer came from the prepared foundation, not from a more "intelligent" model.

How to apply it with a method: map, limit, pilot, hand over

The path we usually recommend for this kind of project follows four steps. For a broader view of the architecture behind it, see the company superintelligence layer in five building blocks.

• Map: collect the real questions that arrive through WhatsApp, Instagram and the physical store, the most frequent reasons for exchanges and where the team loses time. Record how service works today and define what will be measured, such as response time or handoff rate.
• Limit: choose a few categories, give the agent read access to the catalog and stock and keep sensitive actions, such as refunds or discounts, out of its reach. Define what it never answers and when it passes to a person.
• Pilot: run one real flow, for example size questions and exchange status, logging every answer and every lookup. Review the conversations with the team and adjust the catalog and instructions, not just the prompt.
• Hand over: deliver the operation with documentation, owners and criteria. Then decide, based on the pilot data, whether to expand, adjust or stop.

One detail that is often underestimated: the pilot reveals gaps in the catalog more than flaws in the model. Every time the agent says "I don't have that information", there is a spreadsheet row to complete.

Frequently asked questions

Can an AI agent replace the sales associate in a fashion store? That is not the best design. It handles repetitive questions well, checks stock and organizes exchanges, but style advice, relationships and negotiation remain with people. The gain is in freeing the team for the work that needs a human eye.

Do we need our own model trained on our catalog? Generally not as a starting point. Keeping the catalog organized and searchable through retrieval is usually simpler to maintain, because a collection update goes in without retraining anything. The technical decision depends on the case and should be tested in the pilot.

What should we do when the agent gets a measurement or a returns rule wrong? Treat it as a data or instruction failure until proven otherwise. Find the conversation in the log, see what the agent consulted, fix the source and add the case to the review tests. If the error affected a customer, the human team should contact them directly.

Where do we start if the catalog is disorganized? Start with one category that gets a high volume of questions and complete the garment measurements, composition and fit. Run the pilot only on that category. It quickly shows the effort needed for the rest, without betting the whole operation.

Sources

• Anthropic: Building effective agents
• Anthropic: Writing effective tools for agents
• AWS: What is retrieval-augmented generation
• ANPD: Autoridade Nacional de Proteção de Dados