I Built an AI Customer Support System for Returns and Exchanges

For the first two years of running my DTC fashion brand in San Diego, returns and exchanges were eating me alive. My team and I were spending 14 hours a week — almost two full working days — answering the same six questions over and over. Where's my order? Can I return this? Can I get a different size?

So I built a smart assistant to handle it. Not a chatbot. Not some off-the-shelf tool. A custom system designed around my specific products, my brand voice, and my business rules. It handles the repetitive 80% of customer support while keeping me in the loop for the decisions that actually matter.

Here's what it does and why it works.

The Problem Wasn't Hard Questions — It Was Constant Interruption

Most customer emails took 2-3 minutes to answer. The real cost was everything around them. You're in the middle of reviewing a new product or updating your website, and a return request comes in that needs a fast response. So you stop everything, pull up the order, check if the return is eligible, look up whether the exchange size is available, write a reply, attach a shipping label. Three minutes of typing. Fifteen minutes of lost focus.

And when you're doing this 15-20 times a day by hand, mistakes happen. You confirm an exchange for a size that's actually out of stock. You forget to check if the item was bought within the return window. Every mistake creates a follow-up thread that costs more time than the original request.

With handmade products, this is especially painful. I can't just grab another unit off a warehouse shelf. Each piece has production lead time. A promise I can't keep is a broken promise to a customer who already had a problem.

How the System Actually Works

Think of it like hiring a really sharp assistant who reads every incoming email, figures out what the customer needs, pulls up their order information, and writes a draft reply — all before I even see it.

The assistant sorts each email into a category: return request, exchange request, order status question, complaint, or something else. It gets this right about 94% of the time. When it's not confident, it flags the email for me to handle personally. No guessing.

Here's the critical design decision: the AI writes draft replies but never sends them on its own. Every response goes into a review queue where I can approve it with one click, make a quick edit, or handle it myself. This adds maybe 8-10 seconds per email. In exchange, I keep 100% control over what my customers actually see.

Why does this matter? My brand makes handmade fashion. The tone needs to feel personal, not robotic. A bad auto-reply damages trust more than a slow reply. A customer who gets a clearly canned response about their one-of-a-kind piece feels dismissed.

In practice, 78% of drafts get approved without any changes. Another 15% need minor tweaks. Only 7% need to be handled from scratch.

I also built a self-service returns page on the website. A customer types in their order number and email, sees which items are eligible for return, picks a reason, and gets an instant shipping label. Under two minutes, no waiting for a reply. This alone eliminated about 60% of return-related emails. Those customers never needed to contact us at all.

The Part That Actually Earns Its Keep

When a customer wants to swap a Medium for a Large, the system doesn't just say "sure." It checks real-time inventory. If the Large is in stock, it reserves it and drafts a confirmation. Simple.

If it's not in stock, things get interesting. The system looks for alternatives — same style in a different color, similar styles in the right size. And because I make handmade products, it also checks the production schedule. If I have a batch finishing in six days, it can offer to hold the exchange: "We'll have your size ready by Thursday. Want us to set it aside for you?"

Before this system, my exchange completion rate was 64%. After: 89%. The difference comes almost entirely from accurate inventory checking and suggesting smart alternatives instead of just saying "sorry, out of stock."

One more important line I drew: the AI handles store credit automatically, but refunds always require my approval. Store credit stays in the business — if there's an error, I can fix it. A refund is real money leaving the account, and the failure modes are ugly: duplicate refunds, refunds on fraudulent claims, refunds when store credit was the right answer. The principle is simple. Automate where the cost of a mistake is low. Keep humans in the loop where a mistake costs real money.

Results After 90 Days

Here are the hard numbers:

• Response time: dropped from 4.2 hours to 38 minutes
• Weekly support hours: down from 14 to 4.5 (68% reduction)
• Exchange completion rate: 64% to 89%
• Returns handled without any email: 61% (up from 0%)
• Cost: a fraction of what a part-time support hire in San Diego would run ($2,200-2,800/month)

The number I care about most: repeat purchase rate among customers who went through a return or exchange jumped from 22% to 34%. Faster, more accurate service brings people back.

What didn't work perfectly? International returns are still mostly manual — too many country-specific rules to automate safely. And damaged-item claims that need photo review are still human-only. I've tested AI image analysis for this, but honestly, it's only about 70% accurate. Not good enough to trust.

The system also got better over time. Classification accuracy started at 89% and climbed to 94.2% over twelve weeks as I corrected mistakes and fed those corrections back in.

Thinking About AI for Your Business?

If you read this thinking "we're spending way too much time on this stuff" — I'd like to hear about it. I do free 30-minute discovery calls where we look at your operations and figure out where AI could actually move the needle. No slides, no pitch deck. Just an honest conversation about what's worth automating and what isn't.

Book a Discovery Call