01 the business
Fashion and footwear retailer
50 staff, 25,000 orders a month
Each return means opening the order, checking dates and reasons, emailing for photos and making a label. In busy weeks customers wait days. Some refunds go out before the item is back, and some are forgotten.
02 before
The problem
Each return means opening the order, checking the date and reason against the policy, emailing for photos and creating a label. Busy weeks leave customers waiting days for an answer. Refunds sometimes go out before the item is back, or not at all.
- Customers wait days for a return label.
- Refunds are issued before items arrive.
- Nobody can see which products are returned for faults.
03 the build
What the AI does
- Step 1: Reads the return request from email, chat or your returns portal and finds the order in Shopify.
- Step 2: Checks the purchase date, item condition and reason against your written returns policy.
- Step 3: Asks for photos when the reason is damage or a fault, and logs them against the order.
- Step 4: Creates the return label in ShipStation and tells the customer exactly what to do next.
- Step 5: Prepares the refund, exchange or store credit once the warehouse marks the item as received.
- Step 6: Tags repeat fault reasons by product so buying and quality teams can see them.
paused: waiting for a person to approve
How it was put together
- Read return requests and matched them to orders.
- Checked purchase date, condition and reason against the written policy.
- Asked for photos for damage claims and created labels in the shipping tool.
- Prepared refunds only after the warehouse marked items as received.
04 human in the loop
Where a person stays in charge
A team member approves every refund above a set value, every exception to policy and any case where the item has not been received.
- Refunds above a set value, policy exceptions and unreceived items need a person.
- Repeat-return patterns are flagged for review, not auto-refused.
- Every decision is logged against the order.
05 rollout
From kickoff to live in 5 weeks
- Weeks 1–2Write the policy as clear rules and connect shop, returns and warehouse systems.
- Weeks 3–4Process returns in suggestion mode and compare with the team.
- Week 5Automate standard returns; keep exceptions with people.
06 after
What changes
Customers get a clear answer quickly and refunds match what came back. Your team handles the unusual cases, not the routine ones.
Assume 25,000 orders a month, a 15% return rate and 7 minutes of handling per return by hand.
25,000 × 15% = 3,750 returns. 3,750 × 7 minutes = about 437 hours a month of handling.
What to measure
- Time from request to label.
- Refunds issued before receipt.
- Returns by reason and product.
- Customer satisfaction on returns.
07 evidence
What the research says
US retailers expected 16.9% of 2024 sales to be returned, about $890 billion, and 93% said return fraud and abuse is a significant issue.
Across 5,179 customer support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and lower-skilled agents.
NBER, Generative AI at Work (Brynjolfsson, Li and Raymond) (opens in a new tab) · 2023
08 lessons
Mistakes to avoid
- Vague policies. Rules need to be written down before they can be applied consistently.
- Refunding on label scan instead of receipt.
- Ignoring the reasons. Fault patterns are worth more than the time saved.
09 questions
Common questions
- How big is the returns problem?
- US retailers expected 16.9% of 2024 sales to be returned, about $890 billion, and 93% called return fraud and abuse a significant issue.
- Does it work with a returns portal like Loop?
- Yes. It reads requests from the portal and writes decisions back.