Returns processing: 48 hours to real-time
How a fast-growing UK e-commerce brand automated 78% of their returns — cutting resolution times, freeing up staff, and boosting customer satisfaction by 31%.
The Client
A fast-growing UK e-commerce brand processing around 50,000 orders per month. They sell consumer goods across their own Shopify storefront and third-party marketplaces — a category where returns are a fact of life, not an edge case.
With a return rate of around 12%, the operations team was handling roughly 6,000 return requests every month. And every single one was being reviewed manually.
The Challenge
The returns process was straightforward in theory but painful in practice. Each return request required a staff member to:
- Verify the order exists and is within the return window
- Check the reason against the returns policy (damaged, wrong item, changed mind, etc.)
- Cross-reference with delivery tracking to confirm receipt
- Decide whether to approve, request photos, or escalate
- Issue the refund or exchange and update inventory
Each request took 8–15 minutes of manual work. With 6,000 per month, that consumed 3 full-time employees. But the bigger problem was speed: customers were waiting up to 48 hours just to hear whether their return had been approved.
In an era of instant everything, 48 hours felt like an eternity. Trustpilot reviews were starting to mention it. Customer support tickets about return status made up 34% of all inbound queries.
The ops team was stuck in a vicious cycle: the more the business grew, the worse the returns backlog became — and hiring more people wasn't a scalable answer.
The Solution
We built an AI agent pipeline that sits between the customer-facing returns portal and the operations team. The system handles the routine cases automatically and routes the genuinely tricky ones to a human — with all the context they need to decide quickly.
How it works
When a customer submits a return request, an AI agent classifies it by reason, urgency, and complexity. Natural language understanding handles the messy reality of free-text customer descriptions — "it arrived smashed" gets correctly classified even without selecting "damaged item" from a dropdown.
The system pulls order data from Shopify, checks delivery status via carrier APIs, and validates against the returns policy. Is it within 30 days? Was it actually delivered? Does the product category allow returns? All checked in seconds.
Straightforward cases — valid order, within policy, standard reason — are approved instantly. The customer gets a return label and refund confirmation within minutes of submitting. No human touch required.
Edge cases — high-value items, repeat returners, suspected fraud patterns, ambiguous reasons — are flagged for human review. But the agent pre-fills the decision context: order history, customer lifetime value, photo analysis (if provided), and a recommended action. The human reviewer spends 2 minutes instead of 15.
Every human decision on an escalated case feeds back into the model. The system gets smarter over time — the auto-approval rate started at 65% and climbed to 78% within the first two months as it learned the team's patterns.
Technology Stack
The Results
The remaining 0.7 FTE still handles escalated cases — but they're now dealing with genuinely complex situations rather than rubber-stamping obvious approvals. The team reports that the work is more interesting and less repetitive.
Perhaps most tellingly, the Trustpilot score improved by 0.4 stars within three months. Customers noticed the difference before anyone told them the process had changed.
"Returns used to be the thing nobody wanted to deal with. Now it mostly just… happens. Customers get instant answers, the team focuses on growing the business, and the edge cases that do need a human actually get proper attention instead of being rushed through a backlog. It's transformed our operations."
Project Timeline
Mapped the full returns workflow, codified the returns policy into machine-readable rules, and analysed 3 months of historical return decisions to establish baseline patterns.
Built the classification and validation pipeline, integrated with Shopify and carrier APIs, developed the escalation logic, and created the operations dashboard.
Ran in shadow mode for one week — processing returns in parallel with the manual team. Tuned confidence thresholds, then switched to live auto-processing with human oversight on escalations.