Using AI for Customer Support in Your Online Store: Practical Lessons

For an online store, customer messages arrive around the clock, and the questions repeat endlessly: Is this in stock? How long does shipping take? What's your return policy? Will this size fit me? That's what makes AI so tempting for customer support. But it's also where a single wrong answer can cost you a customer, start a dispute, or break a marketplace's rules. This post shares the principles we've found work well when using AI for customer support. It's experience and opinion, not a promise about the results you'll get.

The core principle: AI drafts, rules check, humans decide when needed

Don't think of it as "AI answers customers." Think of it as three layers:

1. AI writes a draft based on your store's real data.
2. A set of rules checks the draft and filters out anything that isn't allowed.
3. A human reviews or takes over in sensitive cases.

Each layer is cheaper than the one after it, so let the cheap layers do their work first.

Start with reply templates

Most customer questions repeat. Instead of letting AI write from scratch every time, build a library of reply templates written and approved by people:

• Group them by situation: pricing questions, stock availability, shipping costs, returns and exchanges, order tracking, post-purchase thank-yous, and follow-ups when a customer goes quiet.
• Give each template clear placeholders (customer name, product name, order number) that get filled in from real data.
• Keep the tone consistent with your brand: friendly, concise, and never overpromising.

That way, AI does the two things it's most valuable for: picking the right template for the context and lightly adjusting it so it sounds natural. In return, the important content, like your policies, always comes from a source you've approved, not from the model's memory.

Absolutely off-limits: AI inventing product specs

This is the most expensive lesson we've learned. AI models tend to fill gaps with something that sounds plausible. A customer asks whether a material is machine washable, the model has no data on it, and it may still answer with full confidence. If that answer is wrong and the customer follows it, the result is a damaged product, an upset customer, and sometimes a return and a bad review.

The rules we follow:

• Every spec (dimensions, materials, care instructions, ingredients, power ratings, warranty, country of origin) must come only from product data the store provides.
• If the data doesn't cover what the customer is asking, the right answer is: "Let me double-check that and get right back to you," followed by a handoff to a human. No guessing.
• AI must never make commitments on its own about prices, discounts, specific delivery dates, refunds, or warranties unless they're already in an approved policy.
• For sizing or fit advice, use only the store's own size chart, and make it clear that it's a guide.

The way to teach the model this is to put the product data in the prompt and state plainly: use only the information provided, and if something is missing, say so.

When to hand off to a human

Write down a list of situations the AI must not handle on its own, and have the system route them to a person as soon as they're detected:

• The customer is angry, complaining, threatening a bad review, or filing a complaint.
• Refund requests, returns outside your policy, or disputes about an order.
• Anything involving health, safety, or legal issues.
• Questions your store's data doesn't answer.
• Large amounts of money or important customers.
• The customer asks to talk to a real person. Respect that request.

When you hand off, include a summary of the conversation and the reason for the handoff so the person taking over doesn't have to make the customer start from scratch.

Check before you send

For outgoing messages, we run an automated check before anything is sent. A few simple but effective checks:

• Does the message contain any numbers, prices, or dates? If so, verify them against the real order or product data.
• Does it make any commitment outside your policy? Look for words like "guarantee," "definitely," or "refund," and pause for a human to review.
• Does it break the rules of the platform you sell on? Many marketplaces have rules about sharing outside contact details or steering customers to transact off-platform. Read the current rules of each marketplace you sell on.
• Does it expose any sensitive customer information? For example, accidentally sending someone else's address or phone number.
• Is the tone right? Not too long, not robotic, and not repeating the same canned line in two messages in a row.

Any message that fails a check is held for a human to review, not sent.

Measure to know if it's worth it

Don't just look at how many messages were answered automatically. Track the things that reflect real quality:

• The share of messages that had to be edited before sending.
• How often conversations were handed off to a human, and why.
• Complaints or returns tied to incorrect information.
• First response time compared with before you started using AI.

If the edit rate is high, that's a signal to improve your templates and product data, not to give the AI more "freedom."

A few common pitfalls

• Outdated or incomplete product data. AI is only as good as the data you give it. Keeping your data up to date matters more than which model you choose.
• Sending the same message in bulk. It makes the automation obvious, and marketplaces may flag it.

Takeaways

• Use three layers: AI drafts, rules check, humans decide when needed.
• Build a human-approved template library; let AI only choose and lightly adjust.
• Take product specs only from your store's data. If something is missing, say so and hand off to a human.
• Write down in advance the situations that must go to a human.
• Check every message before sending, and hold back anything that fails.
• Measure your edit rate and complaints, not just the number of automated replies.