AI for Home Textile Stores: Exchanges and Sizes With Clear Rules

AI for Home Textile Stores: Exchanges and Sizes With Clear Rules

How a home textile store uses an AI agent for mattress sizes, bedding sets and exchanges, with written rules, logs and handoff to a person.

•

Want to apply this in your company?

Book a call with Juan →

An AI agent for a home textile store answers questions about mattress measurements, sheet fit, bedding combinations and exchange status, as long as it works from a size table and return rules written by the store itself. The model helps with the conversation, but the store's process decides what is true and what can be promised. This guide shows how to organize those rules, where the conversation should move to a person, and how to test everything before real customers write in.

Key takeaways

  • Size and set-combination questions are the most frequent topics and the easiest to get wrong without a table.
  • The exchange policy must be written down before the agent starts, separating regret, size or color exchanges, and defects.
  • The agent looks up and explains. Exceptions, refunds and defect analysis stay with a person.
  • Every conversation leaves a record of the order, the reason, the answer given and who made the decision.
  • Start with a small flow, test it with real questions, and only then widen the scope.

Which questions the agent can answer on its own

In a store like this, messages tend to repeat. It helps to split the topics into four groups, from the safest to the most sensitive:

  • Measurements and fit: which fitted sheet suits a mattress of a given height, whether a duvet covers a double bed, which mattress protector matches which size.
  • Composition and care: fiber type, washing temperature, drying advice, the difference between percale and jersey.
  • Building a set: how many sets to buy, color and pattern combinations, pieces that match.
  • Order status and exchange window: where the delivery is, whether the order is still within the store's rule, how to open a request.

The first two groups depend on content the store controls. The last one depends on order data, so the agent needs to query a reliable source instead of answering from memory.

The size table: the foundation of every answer

The most costly mistake in this kind of service is the guess. When a customer asks whether a sheet fits a particular mattress, the agent should not take a chance. It should answer from a table maintained by the store, which lists the actual measurement of each item, the accepted mattress height range and the relevant notes.

A sequence that usually works well to build that table:

  • Measure each product in a standard way and record the real measurement, not only the nominal size on the label.
  • State the mattress height or accepted range for each fitted sheet, mattress protector and bed cover.
  • Record the manufacturing variation the store recognizes and explain that margin to the customer in plain words.
  • Flag cases where the answer depends on something the agent cannot see, such as a mattress topper, a high headboard or a box spring with a skirt.
  • Define a standard reply for those cases. The agent says it will check with the team and does not estimate.

The table needs an owner. Someone in the store reviews it when new products arrive, when a supplier changes a measurement, and when a return reveals a recurring error.

Return rules written before the agent

Returns are not one single topic. For the agent to respond well, the policy must separate at least three situations, each with its own conditions.

  • Regret on a purchase made at a distance, such as through a website or app. The Consumer Defense Code provides a right of withdrawal under those conditions. The store should check the legal text and its own policy with whoever handles legal matters before fixing what the agent tells customers. The agent explains the procedure the store follows and opens the request.
  • Exchange for size or color. This is a commercial policy of the store. Define the conditions, such as unused product, tags and packaging, who pays shipping, and what happens to bath or bed items that show signs of use. Do not let the agent decide alone whether a used item can go back on the shelf.
  • Defects. This involves legal warranty and technical analysis. The agent does not conclude there is a defect just because the customer says so. It collects the necessary information, such as order number, photos and a description of the problem, and passes it to the team.

Write each rule so the agent can apply it: when it applies, who decides, what the customer needs to send, and what the agent cannot promise.

Checklist before putting the agent on WhatsApp

Before opening the channel to customers, confirm the following:

  • The size table was reviewed by someone who knows the stock, and a named owner is responsible for it.
  • The exchange policy sits in a single document, with each case separated.
  • There is a list of what the agent cannot promise: delivery dates without confirmation, refunds, rule exceptions, defect diagnosis.
  • Access is limited. The agent reads order status but cannot change address, price, customer record or commercial terms.
  • There is a handoff path to a person, with business hours and a named responsible contact.
  • Every conversation generates a record with the order, the answer given, the referral and the human decision when there is one.
  • A set of real questions, taken from past service history, exists to test before opening to the public.

If the agent will also send active messages, such as delivery confirmations or exchange reminders, the WhatsApp conversation and template rules must be followed. The subject is covered in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act.

Illustrative scenario: bedding exchange with a claimed defect

Illustrative scenario. A customer writes on WhatsApp saying the sheet set showed a split seam after the first wash and asks for a replacement. The agent finds the order using the number provided and confirms the product and delivery date in the system. It does not try to judge whether the split is a defect. It explains that the team performs the analysis, asks for photos of the product laid flat and a description of how the item was used, and opens a request with that data. At the end, it states when the team will reply and who will contact her. If the customer asks for a refund at that moment, the agent logs the request and passes it on, without promising an amount or a date.

The value of this flow is not deciding faster. It is making sure the customer gets a consistent answer, the team receives the complete case, and no promise is made outside the rule.

When to hand off to a person

Not every conversation should stay with the agent. Use this as the starting criteria for handoff:

  • There is a claim of defect, transport damage or any safety risk with the product.
  • The customer asks for a refund, discount, deadline exception or an exchange outside the written policy.
  • The tone is that of a formal complaint, or the customer already tried to resolve it and did not succeed.
  • The product is not in the size table, or the question depends on something the agent cannot verify.
  • There is a registration error, such as a swapped address, or the message contains personal data beyond what the service needs.

The practical rule is simple: when the answer requires deciding on behalf of the store, the conversation goes to a person. At that point, the agent should summarize the case so the customer does not have to repeat everything.

Customer data and the LGPD

Store orders carry name, address, phone and, in some cases, tax ID (CPF). Brazil's General Data Protection Law (LGPD) requires that data processing be limited to the purpose of the service. In practice, the agent should receive only what it needs to answer each question, not the entire customer base. Conversation logs also need a defined retention period and restricted access.

If the AI tool processes data outside Brazil, the store should assess the international transfer before connecting the system to real customers. The topic is covered in International Data Transfer Under LGPD When Using AI Providers. The national data protection authority (ANPD) publishes official guidance on the subject, and that is the place to check current requirements.

How to pilot and measure without promising results

The approach we use follows four steps: map the current process and the baseline, limit data and actions, pilot a real flow with logging and an agreed metric, then hand over and decide whether to expand, adjust or stop. Applied to bedding, bath and table linen service, this starts with collecting a few weeks of real conversations, sorted by topic, with the answer the team considered correct in each case.

That set becomes the reference test. Before each change to the agent, it runs again to check whether answers still match the table and the policy. For guidance on how agent evaluations are designed, look at published engineering material from AI model providers. To test order integration without risk, use simulated environments, as explained in Mock Tools for Testing AI Agents: Safe Sandbox Testing Guide.

Pilot metrics should be chosen before starting. Good candidates are the share of answers that follow the table, the share of cases that reach the team with complete information, and the number of conversations reopened after handoff. It is reasonable to expect the agent to reduce the queue of repeated questions, but that stays a hypothesis until the pilot measures what actually happens in your operation.

On what is implementation and what is law: consumer and data protection rules come from legislation and should be confirmed with whoever handles those topics in the company. Building the table, the flow and the log is implementation work. The expected gain is a hypothesis that needs data.

Frequently asked questions

Can the agent answer size questions without a ready table?

Not safely. Without a table it will tend to estimate, and a wrong size estimate usually becomes an exchange and frustration. The first step is building the table with the people who know the stock, and only then connecting the agent to that source.

Does a small store need order system integration to start?

Not necessarily. It is possible to begin with answers based on documents: the size table, the exchange policy and a handoff script for a person. Order integration comes in when the pilot shows that the gain justifies the complexity and the data exposure.

Can the agent approve an exchange on its own?

It can open the request and explain the rule, but approving exceptions, used items and refunds should stay with a person. For routine exchanges, the criteria must be written so the agent applies the same rule in every conversation. When a situation does not fit the criteria, the conversation goes to the team.

Does the agent replace the customer service team?

That is not the proposal. The agent handles repeated questions with a clear rule and leaves the team with cases that require judgment, such as defects, exceptions and complaints. That split is what the pilot should test, and the decision to expand or stop depends on what the records show.

Sources

AI agentsHome textileCustomer supportReturns policyWhatsApp