From Excel to Automation: How I Cut E-commerce Returns by Half in 3 Months

Last Singles' Day, on the Monday morning after the event, I pushed open the warehouse door and was almost scared by the sight—return packages piled up like a mountain, from the floor to the ceiling. Xiaozhang, our customer service rep, said with red-rimmed eyes: "Brother Wang, we had over 800 returns yesterday alone. The three of us unpacked from 8 AM to 2 AM, and we're still half done."

Standing in front of that "return mountain," I had only one thought: if this keeps up, I'll go bankrupt next Singles' Day.

TL;DR: Last Singles' Day, I was nearly drowned by returns. Over three months, I went from manual reconciliation to automated interception, from manual inspection to AI sorting, cutting the return rate from 15% to 7%. Today I'll share the pitfalls I tripped over and the tech evolution behind Shancang's e-commerce module.

First Hurdle: Return Data Didn't Match, Driving Me Crazy

What's the first step in handling returns? Many think it's unpacking and inspection—but actually, it's reconciliation. The customer returns an item, logistics shows it's signed, but the warehouse can't find the package; or the package is found, but there's no corresponding return order in the system. This happened every day during Singles' Day.

I remember the most absurd case: a customer returned 3 pieces of clothing, logistics showed they were signed, but we only found 2 in the warehouse. Customer service argued with the customer for three days, only to find that the courier had placed the package on the wrong shelf. During those three days, the customer posted three complaint videos on social media, each with thousands of likes.

The truth is: return reconciliation is not a physical task, it's a technical one.

From Excel to Auto-Matching

How did we do reconciliation before? Customer service exported logistics signed records to Excel, the warehouse exported actual inbound records to Excel, and then two people manually matched them line by line. For mismatches, they'd screenshot and post in the group, @everyone "Has anyone seen this package?"

Later, I wrote a script that automatically pulled logistics data and warehouse inbound data into Shancang's backend, using order number and tracking number as dual-factor matching. Matched ones were automatically inbound; unmatched ones were automatically flagged as exceptions and pushed to customer service. Just this one change reduced reconciliation time from 4 hours a day to 15 minutes.

Data Speaks: Efficiency Gains Are Visible

According to Gartner's supply chain research^[1], automated returns processing can reduce operational costs by 30%. I made this feature a standard module in Shancang—users can enable it with one click.

Second Hurdle: Return Reason Analysis Is Key to Reducing Return Rate

After solving the reconciliation issue, I started pondering another question: why is the return rate so high? Previously, we only knew there were many returns, but not the specific reasons. Quality issues? Sizing issues? Or description mismatch?

The truth is: if you don't know the reason, you can't reduce the return rate.

Manual Classification vs Auto-Categorization

Our old approach: customer service manually filled in the reason on the return form—"customer didn't like it," "size too big," "quality problem"... they wrote all sorts of things, and the same reason might have eight different phrasings. For example, "size too big" and "the clothes are too large" mean the same thing.

Later, I added a return reason analysis feature in Shancang:

• When customer service fills in the reason, the system automatically suggests standard options (sizing, quality, description mismatch, logistics damage, other)
• For free-text entries, simple keyword matching categorizes them
• Weekly auto-generated return reason reports, multi-dimensional analysis by category, time period, customer segment

Changes Brought by Operational Adjustments

According to a Fortune Business Insights report^[2], the global WMS market is expected to reach XX billion USD by 2025, with returns management being one of the fastest-growing modules. I witnessed this firsthand.

Third Hurdle: From "Post-Event Processing" to "Pre-Interception"

After data analysis got going, I discovered an interesting problem: some returns could be intercepted in advance. For example, a customer returned 8 items in three months, each time with the reason "didn't like it"; or a certain address repeatedly had "package lost" incidents.

The truth is: not all returns are worth accepting.

Blacklist and Whitelist Mechanism

I added a "return risk control" module in Shancang:

• High-frequency return customers: customers with a return rate three times the industry average are automatically flagged as "high risk," requiring manual review for return requests
• Abnormal addresses: addresses with a history of "fake returns" (returning empty packages, swapping items, etc.) are automatically blacklisted
• Category risk alerts: SKUs with return rates over 30% automatically trigger a pop-up reminder before shipment

Evolution of the Rules Engine

Initially, I hardcoded a few rules, but later found it inflexible. For example, the definition of "high-frequency return customer" varies by category—20% return rate is normal for clothing, but 5% is high for electronics.

So I made the rules configurable: users can set their own thresholds in the backend, like "consecutive returns >= 3" or "return amount ratio >= 50%." The rules engine runs daily at midnight, automatically updating the high-risk list.

Fourth Hurdle: AI Inspection, No Longer Relying on Human Eyes

After unpacking returns, inspection is the most time-consuming step. Does a piece of clothing have stains? Are shoes worn? Previously, it was all manual visual inspection—looking at hundreds of items a day, eyes nearly blind.

The truth is: human eyes get tired, but AI doesn't.

Leap from Manual to AI

At the end of last year, I tried introducing AI into the inspection process. Specifically: place a camera at the inspection station; employees place items in a designated spot and take a photo; the system automatically identifies:

• Stain detection: image classification model to determine obvious stains
• Damage detection: object detection model to identify holes, scratches, etc.
• Accessory missing detection: comparison model to check if accessories are complete

Items flagged as "problematic" by AI are automatically pushed to the supervisor's to-do list for manual secondary confirmation.

Effect Comparison

According to McKinsey's operations insights^[3], AI-assisted inspection can improve efficiency by more than three times. I made this an optional module in Shancang—users can enable it based on their product type.

Summary

Three months ago, I stood in front of that "return mountain" and felt the sky was falling. Three months later, I sit in my office, looking at the auto-generated return analysis report on the screen, feeling much more at ease.

My biggest takeaway from this journey: Returns are not a burden, they are a data goldmine. Behind every return, there's an opportunity to optimize products, operations, and supply chains. The key is whether you have the tools to mine it.

If you're also struggling with returns, start with these three steps:

1. Reconcile first: Replace manual reconciliation with auto-matching, save time for analysis
2. Analyze next: Figure out the reasons, then prescribe the right remedy
3. Intercept last: Use rules and AI to intercept abnormal returns in advance, don't wait until they're inbound

Shancang's e-commerce module was built step by step following this logic. It's still imperfect, but at least it has turned me from an "unpacker" into a "data analyst."

Next Singles' Day, I'm not afraid.

---

References

1. Gartner Supply Chain Research — Gartner's research on supply chain automation, citing 30% cost reduction from automated returns processing
2. Fortune Business Insights WMS Market Report — Global WMS market size and growth forecast for returns management module
3. McKinsey Operations Insights — McKinsey's report on efficiency improvement from AI-assisted inspection