AI for Sporting Goods Stores: Sizes, Stock and Exchanges

AI for Sporting Goods Stores: Sizes, Stock and Exchanges

How an AI agent checks size and stock by store and stops where an exchange falls outside policy. Limits, logging and a step by step pilot plan.

•

Want to apply this in your company?

Book a call with Juan →

AI for sporting goods stores, in practice, means an AI agent that answers with data from the store's own system: whether a size exists, at which unit, and with what balance. When an exchange falls outside policy, the agent hands the case to a person. The value comes from two parts: a stock lookup read in real time, by store and by size, and a written limit on what the agent may decide on its own. Below, we show how to build that lookup and where the agent should stop.

Key takeaways

  • The agent should check balances in the stock system by product code (SKU) and unit, never answer from memory or from sales estimates.
  • Every availability answer should say which store the balance belongs to and when it was read.
  • Exchanges that fall within explicit rules can be opened by the agent with a log entry; the rest goes to a person with a full summary.
  • The limits are written, tested rules, not a general request for the model to be careful.
  • The pilot starts with one category and one store, with the metric agreed in advance, and the decision to expand, adjust or stop comes afterward.

Why size and stock are the first test

In sporting goods retail, the most common question before a purchase is whether the right size exists. When the agent says yes and the item is not on the shelf, the customer loses time, the store loses the sale, and trust in the channel drops. That is why size lookup is usually the starting point: it is simple to measure, it has an objective answer, and it exposes the quality of the system integration quickly. A fragile integration shows up in the first week of use.

There is also an operational detail. A physical store balance is not the network balance, and stock reserved for the online store may not be available for sale at the counter. Without that label, the agent mixes information that the customer treats as fact. Before the pilot, define which system field is the source of truth for each channel.

The term superintelligence, as used here, describes an agent with controlled access and clear rules, not a system that decides what the store will do. The useful idea for the leadership team is simple: the agent gains scope to answer and record, and loses scope to promise, grant and decide.

How the store lookup should work

A minimal flow, from the message to the log, has five steps:

  • Identify the product by sales code or catalog link, not by free text name, which causes confusion between similar models.
  • Confirm the size requested and, if the answer depends on region, ask for the customer's city or nearest mall.
  • Check the balance by SKU and unit in the stock system, recording the date and time of the reading.
  • Reply with what is available at the indicated unit, what is available at another unit and, if the store offers it, the option to reserve for a set period.
  • Log the query with product, size, unit and result, without storing more personal data than the interaction requires.

If the system does not respond or the balance looks outdated, the agent should say so and offer the store's contact. It should not estimate. The rule 'when I don't know, I say so and transfer' is worth more than any adjustment to tone.

Where the agent stops: exchange policy as a rule

Exchange policy is the part that gets lost most often in conversation. For the agent to apply it, the policy has to become rules a system can check. Phrases like 'we exchange when reasonable' are useless. Write the deadline, the condition of the product, the proof of purchase requirement, the item types excluded, and who decides each exception. Then sort situations into a simple decision table:

  • Within the deadline, no sign of use, with proof of purchase, and exchange for another size or color with balance at the unit: the agent may open the request and confirm the next step with the customer.
  • Past the deadline, no proof of purchase, or signs of use: the agent explains the policy and transfers to a staff member with a summary.
  • Promotional items with their own rules, kits, personalized products, or footwear with signs of prolonged use: transfer to a person.
  • Refund requests, credit outside the rule, or any exception: always a person.
  • Defect reports: the agent triages with fixed questions and follows the warranty procedure set by the store and the manufacturer.

The agent does not grant discounts, promise exceptions or decide warranty on its own. These actions have financial and legal effects and belong to whoever is accountable for the store. For tone and structure of the support flow, the article on AI customer support helps separate what to automate from what stays with the team.

Illustrative scenario: running shoes at two units

Illustrative scenario: a customer asks for size 39 in a running shoe. The agent checks the mall unit, which shows zero balance, and the street store in the same city, which has one pair. It reports both situations with the time of the reading and asks whether the customer wants a reservation for the period the store offers. Later, another customer says they bought a pair of shoes a few weeks ago, used them for training, and want to exchange them without the receipt. The agent explains the policy, does not open the request, and transfers to a staff member with a summary: product, purchase date as stated by the customer, reason given, whether there is proof of purchase, and what has already been explained. The staff member does not need to ask the questions again.

The same scenario carries a common risk. If the customer resends the message, the system must not open two requests for the same order. Payment platforms such as Stripe document this problem and use idempotency keys to handle it. The same logic applies to exchange requests: each order needs an identifier that prevents duplicates.

How to log the exchange and transfer without losing context

A transfer only works if the record is useful. A good summary contains what the customer asked for, what the agent checked, what was told, what is pending, and the agreed response time. Do not copy the whole conversation if the summary is enough, and define how long conversations are kept, according to the company's retention policy.

Under Brazil's LGPD, the necessity principle calls for minimal collection and a clear purpose. For an exchange, the order or receipt number and the product are usually enough. A tax ID (CPF) should only be requested when a tax obligation or the store's procedure requires it, and that needs to be written into the flow. The Brazilian National Data Protection Authority (ANPD) publishes guidance on the subject, and it is worth checking the official page before finalizing the design.

If support happens on WhatsApp, there is an additional channel restriction. The 24 hour window defines when the agent can reply freely and when it needs an approved template, which affects what can be sent after a transfer. Details are in the 24 hour WhatsApp API window guide.

How to pilot with one store before expanding

The method we apply has four moves: map, limit, pilot and transfer. Mapping means describing the current lookup and exchange flow, with response time and the share of conversations that currently go to a person, as a baseline. Limiting means giving the agent read-only access to stock, with no permission for discounts, refunds or order changes. Piloting means choosing one category and one store, running for an agreed period, and keeping a log of every lookup and every transfer. Transferring means handing the flow to the store team with documentation, then deciding, based on the metric agreed in advance, whether the flow expands, adjusts or stops.

The metric must be agreed before starting. One possible example is the share of availability answers confirmed correct in a weekly sample checked by the store itself. Before exposing real customers, exchange rules can also be tested with simulated cases in a staging environment, as described in how to test AI agents with mocks and a sandbox.

Mistakes that tend to show up

Some problems repeat in projects like this, and they are worth checking before expanding:

  • Letting the agent infer stock from past sales or demand forecasts.
  • Searching products by free text and confusing similar models with different colors or versions.
  • Showing a single balance for the whole network, without the unit and the time of the reading.
  • Allowing the conversation to reopen an exchange that is already open, with no order identifier.
  • Measuring only message volume, without checking the quality of answers or the reasons for transfers.

For return and exchange rules in another kind of retail, the article on superintelligence for paint stores, with formulas, colors and returns shows the same separation between what the agent looks up and what the team decides.

Frequently asked questions

Can the agent say a size is available without checking stock? No. An availability answer must always come from the system, with the unit and the time of the reading. If the system is unavailable, the agent says it could not confirm the balance and offers the store's contact.

Who decides an exchange outside policy? A person at the store. The agent can explain the rule, record the request and transfer it with a full summary, but it does not grant exceptions, discounts or refunds. That separation should be written into the flow and tested before going live.

Does this kind of agent need customers' personal data? Only the minimum needed. Checking size and balance requires no personal data. For an exchange, the order or receipt number and the product are usually enough, and a tax ID is only requested when a tax requirement or the store's procedure justifies it.

Where should a store that has never used an AI agent start? Start by mapping the current flow, choosing one category and one store as the pilot, writing the exchange rules, and agreeing on a metric with the team. Expansion only makes sense after the pilot shows where the flow works and where it needs adjustment.

Sources

sporting goodsretail AIinventory lookupexchange policyAI agents