Superintelligence for Clothing Stores: Size Grids and Exchanges

In practice, superintelligence for a clothing store means an AI agent that serves the customer with real stock by size and color, suggests available alternatives, guides the exchange according to the published policy, and routes any deadline exception, defect or case outside the rules to store management. The agent does not decide on its own anything the policy does not cover. The gain comes from correct answers about availability and a clear record of every decision.

Key takeaways

• The agent should only promise sizes and colors that actually exist in the stock it checks, and the store should know how often that stock data is updated.
• Exchanges follow the store's written policy, and the agent applies only what is published there.
• Deadline exceptions, defects and requests outside the rules go to a person in management, with a full summary of the conversation.
• Every interaction leaves a record of what was checked, what was offered and who made the decision.
• The first pilot covers one type of request, one store and one agreed metric before any expansion.

Why the size and color grid is the starting point

Most problems in fashion customer service begin with a wrong promise: the customer asks whether the dress is available in size 42 in sand, the assistant says yes, and the piece had already sold. To avoid this, the agent needs to see the grid at the variant level, meaning each combination of product, size and color with its own balance. A parent product with a total balance is not enough.

Before connecting any system, map three points. First, where the real stock lives: in the ERP, in the e-commerce platform, in a spreadsheet, or in each physical store separately. Second, how often that balance is updated. If synchronization happens in batches, the agent should say that availability was confirmed at the moment of the check, and it should not guarantee a reservation. Third, which fields are reliable, such as size, color, reference code and list price.

In designing the agent's tools, each function should do one clearly defined thing: list the variants of a product, check the balance for a variant, check the balance in other units, and open an exchange request. Tools with clear names and lean responses reduce misinterpretation. Anthropic has published guidance on writing tools for agents, which is worth reading for anyone building the integration.

How the agent guides an exchange without inventing rules

The exchange flow has a simple sequence, and the agent should follow the same order in every conversation:

• Identify the purchase by order number or invoice number, asking only for the data needed to find the sale.
• Check whether the purchase is within the deadline set in the store's published policy.
• Check the condition of the item as described in the policy, such as tags, hygiene and use, and record the customer's answer without judgment.
• Offer alternatives of the same product in other sizes or colors, with balance confirmed during the check.
• If the customer wants a different product, explain the price difference rule written in the policy, without calculating discounts on its own.
• Generate the exchange request with the case summary, or route it to management if any step does not have a clear answer.

The central point is that the agent does not create new conditions. If the policy says exchanges happen in a physical store, the agent should not offer shipping by mail. If the policy says nothing about an item the customer has already washed, the case falls outside the rules and goes to a human queue. This discipline is what separates reliable service from a friendly conversation that later causes conflict at the counter.

Define the tone as well. The agent should say what it can do now, what depends on confirmation, and roughly how long management usually takes to respond, without promising timelines the store does not control.

Step by step for an exchange pilot

An exchange pilot can follow a lean sequence:

• Map the current flow: who handles exchanges today, which channels the requests arrive through, and which reasons come up most often.
• Write the policy as rules the agent can follow, with examples of allowed and forbidden cases.
• Limit the actions: at the start, the agent checks stock and opens requests, but does not confirm refunds or change orders.
• Connect the stock data in read-only mode and test size and color checks using test data.
• Simulate real conversations before exposing customers, including edge cases such as a missing invoice or a defective item.
• Define the pilot metric together with management, for example the share of conversations resolved without an improper handoff and the time until the first useful reply.
• Review the records weekly and decide whether the pilot continues, whether the rules need adjusting, or whether it stops.

Testing with simulated conversations is especially useful. A script of test cases with expected answers shows early where the agent improvises. This approach is close to what teams do with mocks and sandboxes before connecting an agent to real data.

Deadline exceptions and handoff to management

Exceptions are where the agent can do the most harm, because it is tempting to be flexible with a customer. The practical rule is that the agent does not promise an exception. It records the case, organizes the information and hands it off.

Illustrative scenario: a customer writes on WhatsApp two days after the exchange deadline, saying she was traveling and could not get back to the store. She has the tag and the order, and she wants the same dress in another size. The agent confirms that the purchase is past the policy deadline, checks whether the size exists in another unit, explains that the exception decision belongs to management, and opens the handoff with the order, the reason given, the size requested and the balance found. The customer is told that a reply will come through the same channel within a timeframe agreed with the store. Management responds with a decision that is recorded.

A simple guide helps decide what stays with the agent and what goes to the responsible person:

• If the policy answers clearly, the agent responds and records the request.
• If the policy is ambiguous or silent, the case goes to management with a summary.
• If there is a defect, a formal complaint, a refund request or an exception request, the case goes to management without any attempt at resolution by the agent.
• If the customer is upset or asks to speak with a person, the transfer happens without pushing back.

If the service runs on WhatsApp, also consider the message window rules, because management's reply may arrive after the period in which a normal conversation stays open. This point is covered in the 24-hour WhatsApp window with an AI agent.

Customer data, LGPD and what the agent should not keep

An exchange agent needs little data: order number, a name to locate the purchase, a contact for follow-up and a description of the reason. Collecting more than that adds risk without improving the service. In practice, the necessity principle of Brazil's LGPD guides the store not to ask for tax ID numbers, full addresses or payment data when the exchange does not require them.

Conversation records also need rules. Who can access them, how long they are kept and how they are deleted should be defined before the pilot starts. If the agent refuses or changes an exchange in a way that affects the customer, treat that decision as automated and provide a path for human review, the subject of LGPD Article 20 and the right to review. We recommend that the store's legal team review the exchange policy and the text the agent shows customers.

For AI tools that process customer data, you also need to check where the processing happens and which providers are involved. This is covered in detail in international data transfer under LGPD when using AI providers.

How we apply the method in fashion stores

The method we use in AI customer service projects follows four stages: map, limit, pilot and hand over. For exchanges, mapping means describing the current process and the baseline, such as response time and handoff volume. Limiting means defining which data and which actions the agent can use. Piloting means running a real flow with records and a metric agreed in advance. Handing over means transferring the operation to the store's team, with a plan to expand, adjust or stop. These are work stages, not promises of results: each store decides what to measure and what counts as a good outcome.

For readers evaluating the topic more broadly, it helps to separate the marketing term from what is actually operated. An agent with restricted access, written rules and a decision log is a customer service system, and that is what a store should compare against its alternatives. The gap between the pitch and the operation is easier to see in superintelligence vs ChatGPT: the real difference for business.

Frequently asked questions

Can the agent see stock by size and color in real time?

It depends on the integration. If the store has a system that updates the grid by variant, the agent checks that balance at the moment of the conversation. If updates happen periodically, the agent should say that availability was confirmed at that moment and should not guarantee a reservation.

Can the agent approve an exchange after the deadline?

It should not. A deadline exception is a commercial decision for the store. The agent records the case with the necessary information, tells the customer that management will decide, and forwards the request with the full history.

How much customer data should the agent keep?

Only what is needed to locate the purchase, contact the customer and record the reason for the exchange. Payment data and documents the exchange does not require stay out of the flow. Retention periods for conversations should be defined in the store's data policy.

Where should you start an exchange pilot with an AI agent?

Start with one type of request, one store and one metric agreed with management. Map the current flow, write the policy as rules the agent can follow, connect the stock data in read-only mode and test with simulated conversations. Only after reviewing the records does the store decide whether to expand, adjust or stop.

Sources

• Anthropic: Writing effective tools for agents
• Anthropic: Building effective agents
• OpenTelemetry: observability concepts