ecomcloser.com

Delivery van loaded with cardboard boxes

How to make a returns chatbot use the exact store policy

A returns chatbot for ecommerce uses the exact store policy when it checks the same facts your staff check before giving an answer: where the order was placed, which item the shopper wants to return, what condition it is in, when the return window closes and whether an authorized shipping label is available. Turn each policy clause into a condition the chatbot can verify, a question it can ask or a reason to hand the request to staff.

Start with the published policy, not a set of sample replies. Ask only for the fact that changes the next step, and use an authenticated order record when it already contains that fact. When another policy applies, send the shopper to the relevant page on your site instead of extending the default rule.

1. Turn policy paragraphs into decisions

Take a return question your team handles regularly, such as “Can I bring this back to a store?” Find the sentence that answers it, then underline every condition attached to that answer. “Bring it to a store” is the outcome. “Bought on our website,” “unopened” and “within 30 days of delivery” are separate checks. Keeping them separate lets the chatbot explain which fact is missing instead of asking the shopper to read the whole policy.

For each rule, write down the exact policy wording, the fact needed to apply it and where that fact comes from. An authenticated order might identify the product and sales channel. The shopper may have to say whether a seal is intact. A label record may show whether the store authorized a mail return. If you do not have access to that record, the bot cannot infer authorization from the shopper saying they printed a label.

Give each check three possible results: met, not met or unknown. A missing delivery date is not proof the return is late. A known marketplace order is not an unknown sales channel; it may belong to a different policy. Keep those distinctions in the decision list so the bot doesn’t turn missing evidence into a refusal or send a clearly identified shopper down the wrong path.

Finally, specify what each outcome allows the chatbot to say or do. A policy match might allow it to explain how to bring an item to a store. It does not automatically allow it to authorize a label or issue a refund. Give each rule a named policy section so a reviewer can trace an answer to the wording that supports it. If staff disagree on what a phrase such as “original condition” means, resolve that ambiguity with the policy owner or send the case for review; don’t let the bot define the phrase.

2. Write one complete example before mapping the rest

Illustrative policy only: Suppose a home goods store accepts unopened desk lamps bought on its own website for return at participating stores within 30 calendar days of delivery. The delivery date is day zero, and day 30 is included. The shopper needs the order confirmation. For this route, the shopper must bring the lamp to a participating store by the deadline, and no shipping label is required. These are example rules, not terms that apply to other stores.

In this example, mailing back the same lamp is a separate route. The shopper must request a label through the store’s online returns page by day 30. The store authorizes the label after reviewing the request; mailing instructions accompany the authorized label. Marketplace purchases follow the store site’s Marketplace Orders policy section instead. The example policy does not say how to handle a broken seal on a lamp that arrived damaged, so that request goes to staff.

  1. Identify the order and item. Check that the selected line is a desk lamp and that the order came from the store’s website before applying this rule.
  2. Check timing and condition. Use the lamp’s delivery date to calculate the final day. Ask whether the package is unopened if the shopper hasn’t said.
  3. Match the route to the request. For a store visit, give the deadline and what to bring. For mail, point to the label request process without implying that a label has been approved.

An illustrative reply might read: “You can bring the unopened lamp and your order confirmation to a participating store through June 3. You don’t need a shipping label for an in-store return.” That answer requires a verified May 4 delivery and the other conditions above. If the shopper instead says they bought the lamp through a marketplace, do not show June 3. Give them the site link labeled Marketplace Orders in the live chat, where they can read the rules for their order.

3. Find where the order was placed

When a returns policy refers to “store orders,” work out whether it means orders placed at a physical shop, orders from the retailer’s website or both. The place the shopper wants to return the item is a separate fact. So is the place they collected it. A website order picked up at a branch can still be a website order; use the channel recorded for that purchase rather than classifying it by the last place the shopper saw the parcel.

If the shopper is authenticated and the order record contains the channel and relevant item, use those details instead of asking for them again. If several orders could match, ask a question that distinguishes them: “Do you mean the white desk lamp or the black one?” Don’t treat the most recent order as the one they mean. A shopper could be asking about an older item, or about one of two identical products delivered on different dates.

If the channel is unavailable and it changes the answer, ask for it directly: “Did you buy it on our website, at one of our shops or from a marketplace?” Use the categories in the actual policy rather than adding options that have no bearing on the result. A reply such as “I bought it from you” may still leave the question open. Asking where the confirmation came from can clarify the channel without asking the shopper to paste the whole receipt into chat.

Check location separately if it affects the available route. In the lamp example, the order source determines whether the website rule applies, while the branch determines whether the shopper can use the in-store option there. If participation at a particular branch cannot be confirmed, point to the site’s participating-store information or the listed contact. Avoid “any branch” when the policy says “participating stores.” That wording difference determines whether the shopper can act on the answer or may make a wasted trip.

4. Ask about the condition that matters for that item

Start with the exact product line, then look up the condition that applies to it. If the unopened requirement applies to desk lamps only, keep it attached to that category. An order containing a lamp and a cushion does not inherit one condition rule for both items. When someone asks to return “everything in the box,” identify the relevant lines and explain their outcomes separately if the rules differ.

Turn condition terms into plain questions without changing their meaning. For the illustrative lamp policy, “Has the lamp package been opened?” is more useful than “Is it in good condition?” A lamp can look new after its package has been opened. Conversely, opening an outer delivery carton does not tell you whether the product package is open. If the policy distinguishes between them, make that distinction explicit in the question. If it doesn’t and the distinction decides the outcome, send the question for review.

Don’t ask the shopper to repeat a fact already given. “I haven’t opened the lamp box” may answer the condition check, while “I never used it” does not. If the answer is ambiguous, ask one narrow follow-up. If staff inspect goods on receipt, distinguish what the shopper has reported from what staff have actually verified. The chatbot can explain a route based on the reported condition without claiming the item has passed inspection.

A request can also describe something the standard rule doesn’t address. For example, “The lamp arrived damaged, so I opened it to check” is not just another way to say “opened.” Check whether the store publishes a process for damaged deliveries. If it does, direct the shopper there; if it doesn’t, explain that the standard return wording doesn’t settle the request and offer the store’s review contact. Keep the reason in the handoff so staff see why the chatbot didn’t simply apply the unopened rule. That gives staff the facts without making the shopper start the conversation over.

5. Calculate the window from the event the policy names

A return window needs a starting event and a deadline event. The starting event might be purchase or delivery, depending on the store’s policy. The deadline might apply to bringing an item to a store, requesting a label or another action the policy specifies. “You have 30 days” leaves both ends unclear. For each route, write down what starts the clock, how the days are counted and what the shopper has to do before time runs out.

Use the date that belongs to the item, not the date that happens to be easiest to find. If two items from one order arrive in separate shipments and the rule starts at delivery, check which shipment contained the return item. An order confirmation date is not a substitute. If the record only shows dispatch, ask for the documented delivery date where appropriate or refer the case for review. Don’t invent a date from a usual shipping estimate.

Test the calculation using actual dates. In the illustrative policy, a lamp delivered May 4 is on day zero that day. June 3 is day 30, and June 4 is outside the example window. A shopper who wants to visit a store has through June 3 to bring the item in. Someone choosing mail has through June 3 to request authorization. That does not make June 3 the date they have to post the package; the example says mailing instructions come with the label.

State the calculated date and the action together: “Request your label by June 3.” That is easier to act on than “You’re still within the window.” If the policy documents a cutoff time, use it; if it doesn’t, avoid inventing one. Before launch, check the day before the deadline, the deadline itself and the next day for each route. When a case falls outside a documented window, state that the standard route is unavailable rather than guessing whether staff can make an exception.

6. Separate a label request from a label authorization

Map the label process only after identifying the correct policy and return route. An eligible in-store return may not need a label at all. An eligible mail return may require the shopper to request one. The statement “You can request a label” describes a next step; “Your label is ready” describes a status and needs evidence. Even if both replies contain the word “label,” they answer two different questions.

Write a response for each status your process can verify. If no request exists and the item qualifies for the mail route, direct the shopper to the store’s label request page. If the request is pending, explain what the store says happens next and avoid giving an approval timeline the policy doesn’t state. If an authorized label exists, give the documented way to retrieve it and any instructions the store provides. If chat cannot check the status, say where the shopper can verify it or who can help.

Check that the label belongs to the return the shopper is asking about. For example, someone may have an authorized label for a cushion and ask whether they can add a desk lamp to the same parcel. Unless the store’s process confirms both items are covered, don’t tell them to combine the returns. Likewise, a shipping label stuck to the original parcel is not evidence of return authorization. If a shopper reports a label won’t open or was already used, follow the store’s documented resolution path rather than promising a replacement.

In the lamp example, an illustrative reply could say: “To return the lamp by mail, request a label through Online Returns by June 3. The store reviews the request before authorizing a label.” The reply names the action, destination and deadline without claiming anything has been submitted. In a live flow, check that the destination actually starts a label request. A site link to an explanation of the policy is useful for reference, but it doesn’t start the request.

7. Route different rules and unresolved cases differently

A shopper who belongs under another published rule needs the right source, not a vague escalation. If the marketplace channel is confirmed, the bot can say: “This order follows our Marketplace Orders policy, not the website-order return process.” In the live chat, use a descriptive site link labeled Marketplace Orders that lands on the relevant page or section. Don’t send the shopper to the returns homepage if that means they must work out which rule applies all over again.

Send a different reply when the policy doesn’t settle the question. Suppose the shopper reports that the lamp arrived damaged and opening the package was necessary to check it. If no published rule covers that situation, say what the chatbot can establish and what staff need to review: “The standard lamp return requires an unopened package. The policy doesn’t say how to handle a lamp that arrived damaged and was opened to inspect it. I can direct you to our returns team for review.” This neither denies the request nor promises an exception.

Make the handoff useful. Where the support process allows it, carry forward the verified order, the selected item, the relevant dates, the shopper’s description and the rule that could not be applied. Keep reported details separate from verified records. If chat only provides a contact, say so; don’t claim a case has been filed. If the chatbot shares an interface with shopping tools, keep actions distinct there too: a spoken selection isn’t a confirmed cart change, and a cart change isn’t payment. In returns, asking for review isn’t authorization or a refund.

Reserve a direct refusal of a particular route for a case the policy clearly resolves. A website-order deadline that has passed is different from a delivery date you can’t verify. Give shoppers the next documented step in either case. Avoid adding a general “contact us for an exception” promise unless that is an option your staff actually offer.

8. Test what the chatbot says and where it sends people

Build test conversations from the policy conditions, not just polished sample questions. For each one, record the order facts the chatbot starts with, what the shopper says, any follow-up it should ask and the answer it should give. Include the expected destination and any action it must not claim to have taken. That lets someone reviewing the test tell whether a fluent answer is actually correct.

  • Regular request: An authenticated website shopper wants to bring an unopened lamp back on day 30. Expect the date, the participating-store requirement and what to bring. Do not expect a label prompt or another question about the sales channel.
  • Category condition: The shopper says a lamp was opened but never used. Expect the category’s unopened rule, not an approval based on lack of use. If they report damage on arrival, test that the conversation takes the applicable source or review path.
  • Different policy: The same lamp came from a marketplace. Expect the site destination for marketplace orders, not the website-order deadline or label process.
  • Missing evidence: The order shows dispatch but not delivery. Expect the bot to gather the relevant date or offer review, never to calculate the window from dispatch.
  • Outside the documented process: The shopper wants to put a second item in a parcel covered by a label for the first. Expect guidance based on verified coverage or a review path, not permission to combine items.

Run short, messy versions as well. Try “got it last month,” “bought online” and “the label’s there somewhere.” Check whether the bot asks only for the detail needed next instead of restarting the entire intake. Verify the destinations in the live interface, including whether the shopper lands on the promised section. When policy wording changes, update the expected answers and rerun the old boundary cases. Include at least one case where the old and new rules produce different outcomes so stale wording is easy to catch.

9. Measure policy accuracy before aiming for fewer handoffs

Review a sample of conversations against the policy that applied to each order. Mark whether the chatbot identified the channel and item, applied the right condition, used the correct dates and described the label status accurately. Give separate marks for the decision and the next step. A chatbot can correctly identify an eligible mail return and still get the conversation wrong by sending the shopper to the in-store policy or claiming a label already exists.

Track errors by what you can fix. A wrong window may point to a calculation problem, a missing delivery date or the wrong date field. Repeated questions about sales channel may mean the order record isn’t being consulted when it’s available. A wrong-site destination calls for a routing correction, even if the text of the reply is accurate. Count unnecessary questions alongside wrong answers: asking for a purchase date you already have wastes effort, but asking for a missing delivery date may be essential.

Measure handoffs by reason, not just volume. Separate cases the policy leaves open from those the bot could have answered using information it already had. A correct handoff on a damaged item is not the same failure as escalating a routine unopened return. If your support process lets you review what happened next, check whether staff had enough information to proceed and whether shoppers came back because the suggested route was wrong. Use those observations to improve the rule, the handoff or the site destination.

Assign someone to maintain the decision list and its tests whenever the published policy changes. Keep track of when wording takes effect if the store distinguishes between orders under old and new terms. After a change, verify the affected dates, product categories, routes and page sections together. The clearest measure of a good returns reply is whether a shopper can take the correct next step without having to reconcile the chatbot’s answer with a different rule on the site.

Try the EcomCloser demo

Related guides

Sources