Score every return request, approve the clean ones and send the risky ones to a person
Updated · By Ecomsellertool editorial team
Score each return request when it arrives against a written signal table: the customer's return rate, repeated item-not-received or wrong-item reasons, requests after the window, high-value items and worn-item notes from 3PL inspection. Clean requests follow your normal return rules. Risky ones are approved and watched, held for inspection before the refund, or sent to a person. A person makes every decline and every customer-level block.
- Shopify's self-serve returns let customers request returns from their order status page, and the merchant approves or declines each request based on its policies (checked November 20, 2026).
- Shopify Flow's Return requested trigger starts a workflow when a customer requests a return through self-serve returns; Flow's return actions are cancel, close and reopen, with none to approve or decline.
- In the Admin API, only returns with status REQUESTED can be approved or declined; decline reasons are final sale, return period ended and other, and the customer is notified only if the request says so.
- A Shopify refund cannot be canceled or reversed once started, which is why risky returns hold the refund until the 3PL has inspected the item.
- On Amazon FBA, Amazon handles returns on the seller's behalf, so the score there is for visibility, not for approving requests.
The step map: who runs each step
Trigger: A customer requests a return through Shopify self-serve returns (the Return requested trigger), or support logs a return or item-not-received request by hand, plus the 3PL's inspection result when a returned package is graded.
| Step | What happens | Who runs it |
|---|---|---|
| 1 | When a request arrives, rules read the order and the request: the items, their value, the reason and the customer's note, the delivery date against the return window, and the customer's order and return history for the last 12 months. | Rule |
| 2 | Rules work out the customer's return rate by value and by order, count item-not-received and wrong-item reasons, and run the signal table. Every matching row is written on the request with its value. | Rule |
| 3 | An AI model reads the customer's free-text note, and any photos the customer sent support, and flags when the note does not match the chosen reason, such as a defect reason with a note about changing their mind. It never approves, declines or refunds. | AI model |
| 4 | Requests with no matching row are approved under your normal return rules. Approve-and-watch rows are approved and tagged. Hold rows are approved with the refund held until inspection. Review rows wait in one queue. | Rule |
| 5 | A person reviews each queued request with the signals, the order and the customer's history on one card, and decides: approve, approve with the refund held, ask the customer a question, or decline with a reason and a message. | Person |
| 6 | When the 3PL grades the returned item, rules release held refunds on clean grades and send worn, wrong-item and empty-box grades back to the person who reviewed the request. | Rule |
| 7 | Each week rules list scored requests, decisions, held refunds released and declined, and any customer whose rate crossed a threshold, so the team can change the table and review any customer-level step. | Rule |
5 rule steps · 1 AI model step · 1 person step. A rule follows fixed logic you approve; an AI model handles messy input and its output is checked by rules; a person makes the calls listed below.
What stays with a person
- Every declined return request and the message that goes with it
- Every exception: a late return accepted, a partial refund, store credit instead of a refund
- Any customer-level step, such as ending free returns for one customer or refusing future returns
- Worn, wrong-item and empty-box inspection results before any refund decision
- The signal table, its thresholds and the fairness rules
Which tools can run it
- Shopify Flow: the Return requested trigger, Add customer tags and a Slack alert, when a person approves each request
- Shopify Flow or webhooks plus Zapier, Make or n8n, to count returns per customer, post a review card and hold refunds
- Custom code on your accounts, for the full signal table, 3PL grades, approvals and declines through the Admin API and the weekly report
Systems it connects
- Shopify returns and refunds
- Shopify Flow
- 3PL warehouse system or API
- Helpdesk or email
- Slack or a review board
Build size: Medium
The rules are simple counts; the work is a fair signal table, clean 3PL grades and keeping every decision on the record.
Most stores treat every return request the same way. Either a person reads each one, which is slow and inconsistent, or every request is approved and refunded on the carrier scan, which pays out for the worn dress and the empty box. This recipe sits between the two. It scores each request against the customer's own history and the order's facts, lets clean requests through your normal rules, and sends the few risky ones to a person with the reasons on one card. It is one page of our automation library.
This page covers the decision on the request. What happens after approval, the RMA, the 3PL inspection, the restock and the refund, is in our returns, restock and refund recipe. Fraud at checkout, before an order ships, is in the high-risk order review queue.
Which return signals should send a request to a person?
Write the table before you build anything, so "suspicious" means the same thing to everyone on the team. Each row names a signal, how rules detect it, an example threshold and the action. The thresholds are illustrative, not advice for your store: set yours from your own return history. Four actions are enough. Auto-approve follows your normal return rules. Approve + watch approves and tags the customer for the weekly list, with no change the customer sees. Hold refund until inspection approves the return but pays only after the 3PL grades the item. Person reviews waits in the queue.
| # | Signal | How rules detect it | Threshold example (illustrative) | Action |
|---|---|---|---|---|
| 1 | Clean request | Inside the window, normal reason, customer below every threshold, value under the limit | No other row matches | Auto-approve |
| 2 | Ordered two sizes, returns one | Same product in two sizes on one order; the request returns one of them | One size kept | Auto-approve (normal shopping, not a risk) |
| 3 | Rising return rate | Customer return rate by value over 12 months (formula below) | 30% to 50%, with at least 4 delivered orders | Approve + watch |
| 4 | High return rate | The same formula | Over 50%, with at least 4 delivered orders and 3 returns | Person reviews |
| 5 | Repeated "item not received" | Count of item-not-received refunds or reships for the customer, from your helpdesk tags or refund notes | A second claim in 12 months | Person reviews, with the carrier's delivery scan on the card |
| 6 | Repeated "wrong item" | The standardized return reason on each line item, counted across the customer's orders | A second wrong-item reason in 6 months on a SKU with few pick errors | Hold refund until inspection, with photos compared to the SKU |
| 7 | Request after the window | Delivery date against your return rule, on requests support logs by hand | Any day past the window | Person reviews (accept as an exception or decline) |
| 8 | High-value item | Line value on the request | Over $300 | Hold refund until inspection |
| 9 | Wardrobing signals | Event-wear or similar category, returned soon after delivery, plus 3PL notes such as tags removed, makeup or deodorant marks, odor or wear | Any worn note from inspection | Hold refund until inspection; a person decides on worn grades |
| 10 | Past inspection problems | The customer's earlier returns graded worn, wrong item or empty box | One in 12 months | Hold refund until inspection |
| 11 | Note contradicts the reason | The AI model compares the customer's note with the chosen reason | A defect reason with a note about changing their mind | Approve + watch; the model only flags |
Three rules keep the table fair. A single signal rarely sends a request to a person on its own; rows 4, 5 and 7 do because each needs judgment. No row declines anything or blocks a customer: those decisions stay with a person, every time. And the signals are only the customer's own behavior with your store and the facts of the order. Never score on a name, an age, a gender, a zip code, a neighborhood or anything that stands in for a protected trait. That would be unfair, and it is the kind of rule that creates legal risk.
How do you calculate a customer's return rate?
Use value, not just counts, because one returned coat matters more than three returned socks. Score only customers with enough history, so a first order with one return does not read as a 100% rate.
Customer return rate = value of items returned in the last 12 months ÷ value of items delivered in the last 12 months (scored only with at least 4 delivered orders)
Here is an illustrative worked example, not a real customer. A customer has 8 delivered orders in the last 12 months, worth $1,200 in items. They returned items from 5 of those orders, worth $660. Their return rate is $660 ÷ $1,200 = 55% by value, and 5 ÷ 8 = 62.5% by order. That is over the 50% example threshold in row 4, so their next request goes to a person. If your store's own average is, say, 15% by value, the reviewer sees that next to it. Then they read the history: if most returns were size swaps under row 2, the right call is to approve.
What does Shopify let a workflow do with return requests?
Customers request returns themselves. Shopify's page on self-serve returns says customers submit return requests from their order status page, and you approve or decline each request based on your policies (checked November 20, 2026). Shopify's return rules set the window (no returns, 14, 30 or 90 days, unlimited or custom), the final sale items, a restocking fee and the return shipping fee, and customers can request only items that are eligible under those rules. So the late and final-sale requests that reach the queue are mostly ones support logs by hand.
Shopify Flow can start the workflow but not finish it. The Return requested trigger starts a workflow when a customer requests a return through self-serve returns, and passes the return to it (checked November 20, 2026). The Add customer tags action can tag the customer for the watch list. But the Flow actions list has only cancel, close and reopen for returns, with no action that approves or declines a request (checked November 20, 2026).
| Job | In Shopify Flow | Where custom code or a workflow tool takes over |
|---|---|---|
| Start on a new request | Return requested trigger | Requests support logs by hand need their own start |
| Count returns per customer over 12 months | Not practical in a condition | A query over the customer's orders and returns |
| Tag a watched customer | Add customer tags | Nothing more needed |
| Approve | No built-in action | returnApproveRequest |
| Decline with a reason and a message | No built-in action | returnDeclineRequest, sent only after a person decides |
| Hold the refund until inspection | Not available | The 3PL grade decides when the refund runs |
The Admin API does the deciding. returnApproveRequest approves a request and sets its status to open, and works only on returns in the requested status (checked November 20, 2026). returnDeclineRequest declines one, and only non-refunded requests in that status can be declined. Its input takes a reason, a note of up to 500 characters for the customer, and a notify option that is off by default. The decline reasons are final sale, return period ended and other (all checked November 20, 2026). In this recipe, code calls the decline only after a person has chosen it and written or approved the note.
The request itself carries the facts the rows read. Each return line item has the quantity, a standardized return reason, a reason note and a customer note of up to 300 characters (checked November 20, 2026). The older return reason field is marked deprecated there, so build row 6 on the standardized reason. The customer note is what the model reads for row 11.
Why hold the refund instead of declining the return?
Because a hold costs the honest customer a few days, and a wrong decline costs you the customer. Shopify's page on refunding orders says you cannot cancel or reverse a refund once you start it, and you cannot create a return after a refund has been issued (checked November 20, 2026). So the money is the step to protect, not the request. Most risky rows approve the return, send the label and wait for the 3PL's grade. A clean grade releases the refund on the same rules as any other return.
The grade is where wardrobing and wrong-item claims show up. A dress with the tags cut and makeup on the collar, a cheaper item in the box, or an empty box all come from inspection, not from the request. Those grades go back to the person who saw the request, with the photos, and that person decides: full refund, partial refund, store credit or decline. The grading table itself, and how each grade maps to restock and refund, is in our returns, restock and refund recipe.
What about marketplace returns?
For Amazon FBA orders, there is no request to approve. Amazon's Fulfillment by Amazon page says Amazon handles customer service and returns on your behalf (checked November 20, 2026). The score still helps there as a report: which SKUs and which reasons come back most, so the listing or the packaging can change. We could not open Amazon's customer return policy page on November 20, 2026, so this page gives no Amazon return windows. For merchant-fulfilled marketplace orders, each marketplace's own return rules decide what you may refuse, and a claim that escalates goes to our A-to-z claim response queue.
What should the reviewer see, and what is kept?
One card per request: the matching rows and their values, the customer's return rate next to your store average, their last 12 months of orders and returns, the carrier's delivery scan for item-not-received claims, the model's flag if any, and any photos the customer or the 3PL sent. The reviewer picks a decision and a reason from a short list, so the weekly report can count them. If the customer writes in about the request, our support-ticket triage recipe routes the message back to the same card.
Keep the record on the order: the signals at the time, who decided, when, why, and the message sent. If the customer later files a card dispute instead of waiting for the refund, that record feeds our chargeback evidence recipe. Each week, look at how many scored requests a person approved anyway. If most of row 3 or row 4 end in approval, the threshold is too tight, and good customers are waiting for nothing. For more on what Flow covers alone, see our guide to what Shopify Flow automates.
Who builds a return abuse review queue?
Ecomsellertool builds workflow automation on your own Shopify store, 3PL accounts and helpdesk. We have built software on Amazon's seller APIs since 2017 and shipped 50+ tools for 100+ teams. We write the signal table and the fairness rules with you first, then build the scoring, the queue, the refund holds and the weekly report. Declines stay with your team. Send a project brief or read how our workflow automation service works.
The building blocks are work we have shipped:
- Refunds in the numbers that matter. For Lavley we built an Amazon analytics app whose insights module brings sales, inventory and advertising cost together to show profit after costs, with goals per brand and product type. A return score needs the same habit: refunds read next to sales, per customer and per SKU.
- Rules that run on triggers. For Sun Company we built rule-based automation for Amazon advertising: triggers that move search terms and change bids when their conditions are met. The signal table works the same way, with a person deciding anything that cannot be undone.
Sales, units, expenses, refunds and net profit in one view
Read the Lavley case study ↗The queue can run on Ecomsellertool Growth OS beside your returns and order workflows.
Get your return review queue built
Send us your return policy and the returns you worry about
Describe the workflow you want automated. We reply with a written plan: scope, timeline and a fixed price.
Frequently asked questions
How do I stop return abuse on Shopify without punishing good customers?
Score each return request against a short written table of signals, such as a high return rate over 12 months or repeated wrong-item claims, and let clean requests flow through your normal return rules. Risky ones get their refund held until the item is inspected or go to a person. Never decline or block a customer automatically.
Can Shopify automatically approve or decline return requests?
Not with a built-in Flow action. Flow's Return requested trigger can start a workflow and tag the customer, but its return actions are cancel, close and reopen. Approving and declining are Admin API mutations, returnApproveRequest and returnDeclineRequest, which an app or custom code can call for requests still in the requested status.
How do I calculate a customer's return rate?
Divide the value of items the customer returned in the last 12 months by the value of items delivered to them in the same period, and do the same by order count. Only score customers with enough history, for example at least four delivered orders, so one return on a first order does not look like a 100% rate.
What is wardrobing in ecommerce?
Wardrobing is buying an item, using it, usually once for an event, and returning it as unused for a full refund. The signals come from inspection: tags removed, makeup or deodorant marks, odor or wear. That is why a returns workflow holds the refund until the 3PL grades the item instead of refunding on the carrier scan.
Is it legal to refuse returns from serial returners?
It depends on where you and your customers are, and on what your published policy says. Consumer laws in some places set return rights a store policy cannot remove. This page is not legal advice: ask counsel before you refuse returns from a customer, and keep that decision with a person.
Should a return score use the customer's location or demographics?
No. Score only the customer's own behavior with your store and the facts of the order: returns, reasons, timing, values and inspection results. Never use protected traits or stand-ins for them, such as a name, a zip code or a neighborhood. Those signals are unfair and can create legal risk.
