AI for a Mattress Store: Deliveries, Returns and Refund Approvals

An AI agent for a mattress store can answer delivery window questions, confirm existing appointments, and guide customers through the steps of a return, all from the order record and a written policy. Refund decisions, exceptions to the return period, and any claim that a mattress is defective or soiled should stay with a manager until the agent has been reviewed on real cases. The useful question is not whether the agent can read the order, but which actions it may take alone and which it may only prepare.

Key takeaways

• Delivery and return questions suit an agent first because the answers live in the order record, the carrier schedule and a written policy.
• An agent can check return eligibility against the policy, but a person should decide the refund and any exception.
• Every money-moving action needs one named owner, a record of what was read and approved, and a way to stop it.
• Set approval rules by action type and amount, not by who happens to be asking.
• Measure the agent on whether customers get a correct answer the first time, not on how many messages it closes.

Why deliveries and returns suit an agent first

Mattress retail has a specific shape. The product is heavy and bulky, a delivery usually needs someone at the door and a clear view of the stairs, and a return means a pickup, a condition check and a restock decision. Most messages from customers repeat the same few questions: when will it arrive, will it fit up the stairs, can I send it back, and what happens to my money. Those answers already exist in the order record, the carrier schedule and the return policy. An agent that reads those sources and states them accurately does useful work, and it does not need to make judgment calls to do it.

Delivery windows: what the agent needs to know

A delivery answer is only as good as the data behind it. Before the agent answers a window question, it should read the confirmed window from the order system, not infer one from the order date. It also needs the store's most common access rules, written in plain language that staff already use. Confirming an existing window is a read and a reply. Moving a delivery to another day changes a carrier booking, so it is an action with consequences and should follow the approval rules you set.

The delivery memory should include:

• The confirmed window and status, read from the order record at the time of the question
• Service areas and postcodes the store does not deliver to
• Stair, elevator and access rules the store has agreed to, with the wording staff use
• Carrier cut-off times for changes requested the same day or the day before
• The rule for who may move a window, and what the customer is told when a move is not possible

Tell the agent to say when it does not know. A reply such as 'I can see the order is scheduled for the stated day, and I have passed your access question to the team' is far better than a confident guess about when a truck will arrive.

Return questions the agent can answer alone

Customers usually want four things: whether the return is allowed, how to start it, what they must provide, and when they will hear a result. The agent can answer all four if the return policy is written in full. That means the return period, the excluded product categories, what counts as an opened item, and which photos help a review. It can open a return request, collect the order number, the reason and the photos, and then hand the case to staff. It should not describe a sag, a stain or a squeak as a defect, because that is a finding a person must make, not a fact the agent can read from the record.

Illustrative scenario: A customer writes three days after delivery to say the mattress feels too firm and asks to send it back. The agent reads the delivery date, confirms the item is inside the written return period, explains the pickup steps, and opens a return request with the reason and the order number. A staff member confirms the pickup slot. The agent does not promise a refund amount or a deduction until a manager has reviewed the case against the policy.

For a closer look at how support, catalog and returns work fit together in an online store, see superintelligent agents for ecommerce catalog, support and returns.

Which refund decisions stay with a manager

A refund is a payment, and payments need an owner. The agent can prepare the case, but a person should decide every refund until the agent has been evaluated on the cases it will handle. The guide below is a starting point. Adjust it to your written policy and to what your payment system allows.

• Refund inside the written return period, with no product question: the agent prepares it and a staff member approves it
• Any refund outside the written return period: a manager decides
• Any claim of damage, sag, stain or defect: a manager decides after reviewing the photos
• Partial refund, restocking deduction or goodwill credit: a manager decides
• Refund to a different person or payment method than the original purchase: a manager decides and confirms identity
• Card disputes and chargebacks: a manager handles them directly, outside the agent's queue

Two safeguards matter more than the list itself. First, the agent should never be able to issue the same refund twice. Payment providers document idempotency keys for exactly this reason, so that a retried request does not create a second payment. Second, every refund decision should record who approved it, what the agent read and what it recommended. Rental operations face the same logic around returns and charges, which is covered in AI for a car rental company: bookings, returns and fleet flags.

Narrowing approval later is possible. If evaluations show the agent applies a written rule correctly within one category, such as a full refund inside the return period with no product question, you can let it complete that category alone. Make the change once, record it, and review it whenever the policy changes.

A staged plan for handing work to the agent

Most teams should widen the agent's authority in steps, with a review before each step.

• Step 1: Write the policy the agent will read, including delivery rules, the return period, eligibility and what counts as a defect.
• Step 2: Give the agent read-only access to orders, the carrier schedule and the policy text.
• Step 3: Let it draft replies and return requests that staff approve before anything is sent.
• Step 4: Allow it to confirm existing delivery windows and send the confirmation to the customer.
• Step 5: Add refund preparation, with every payment still approved by a person.
• Step 6: Review a sample of conversations and decisions each week, and change one permission at a time.

Check authorization on every action, not once when the conversation starts. The OWASP authorization guidance is a useful reference: the agent's access should be limited to the specific customer and order in the conversation, and a request to act on someone else's order should be refused. Treat text inside customer messages as data, not as instructions. A message that tells the agent to ignore the policy and refund the full amount is an attempt at prompt injection, and the OWASP prompt injection prevention cheat sheet describes the defenses.

How to tell whether the agent is helping

Pick measures a manager can check without reading a dashboard of technical traces. Useful ones include the share of sampled delivery and return answers that a reviewer marks correct, the number of prepared refund decisions a manager had to change, the time between a return request and a confirmed pickup slot, and the number of messages that reached a person because the agent did not know the answer. Track complaints as well. A fast wrong answer about a delivery costs more than a slower right one.

Review the agent's record of what it read and did, the same way an operations team reviews any tool that touches orders. The guide to ecommerce operations beyond customer chat explains why the work reaches well past the chat window. The NIST AI Risk Management Framework is a practical reference for defining these measures and the human oversight around them.

How Be Human approaches this in practice

Be Human starts with one role, here delivery and return questions, and defines the memory the agent reads: the policy text, the order fields it may see, the carrier windows and the store's access rules. It sets which tools the agent can call, which actions need a named approver, and an observable success criterion before anything goes live. Under managed operations, the work continues after launch, with monitoring, recurring evaluations of sampled conversations, permission reviews when the policy or the carrier schedule changes, integration maintenance, and incident response if the agent acts outside its boundary. This is an operating arrangement with clear limits, not a promise of particular results.

If you are not sure which part of your delivery or return work should go to an agent first, the free Be Human AI diagnostic is designed to find that first task.

Frequently asked questions

Can an AI agent issue refunds for a mattress store? An agent can prepare a refund when the case fits a rule you have written, such as a full refund inside the return period with no product question. In most setups a staff member still approves the payment, and that approval is recorded alongside the agent's summary of the order and the policy it applied. Refunds involving a defect claim, a partial amount or a card dispute should stay with a manager.

Who decides whether a mattress is defective? A manager should decide. The agent can read the customer's description, collect photos and check the warranty terms, but judging whether a sag or stain is a manufacturing defect or normal wear is a finding for a person. The agent's role is to make that review quick, complete and well documented.

What happens if the carrier changes a delivery window after the agent has confirmed it? The memory the agent reads must be refreshed when the carrier schedule changes, so it does not keep repeating an old window. A customer whose window moved should receive a clear message from the store, not a silent change. Include this case in your permission and policy reviews so the rule stays current.

Do customers need to know they are talking to an AI? Yes, the conversation should make clear that the assistant is automated and explain how to reach a person. Customers should not have to guess whether a human has read a refund request. Clear labeling also helps when a case is escalated, because the staff member can see what was already said.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.