The most common mistake with AI customer service in ecommerce is starting with the bot instead of the questions. A brand installs a chat widget, points it at the FAQ page, switches it on, and within a week shoppers are typing talk to a human in capital letters. The bot was not the problem. It was asked to handle everything, it had access to nothing useful, and nobody decided when it should step aside.
The better starting point is your inbox. For most online stores, a large share of support messages are variations of three questions: where is my order, how do I return this, and what size should I get. Those three are predictable, answerable and tied to data you already have. Get them right and the AI earns trust. Try to cover everything on day one and it loses trust fast.
Before you automate: sort a month of tickets
Export a month of support emails and chats and tag each one with a simple category. You do not need software for this; a spreadsheet works. Typical categories:
- Order status and shipping
- Returns and exchanges
- Sizing and fit
- Product questions (materials, ingredients, compatibility)
- Discount codes and pricing
- Complaints, damaged items and anything emotional
Count them. The top three categories are your first automation targets. The complaints category is your first list of things the AI should not handle alone.
Order status: answer with real data or not at all
An AI that replies orders usually ship in 3 to 5 business days to someone whose package is a week late is worse than no reply. Order status only works if the assistant can look up the actual order.
A good flow looks like this:
- The shopper asks where their order is.
- The assistant asks for the order number and the email used at checkout, and verifies both match before sharing anything. Never reveal order details based on an order number alone.
- It pulls the order status and tracking link from your store platform and replies in plain language.
- If the order is delayed past your stated window, it apologizes, shares what it knows and offers a human follow-up.
Example reply: Thanks, Maya. Your order #1042 shipped on Tuesday with tracking number ending 5531. The latest scan shows it at your local depot, so it should arrive within a day or two. Here is the tracking link. If it has not arrived by Friday, reply here and a member of our team will look into it personally.
Notice what it does: confirms the order, gives the real status, sets an expectation and names a next step.
Returns: explain the policy, then start the process
Returns questions are usually about eligibility and steps. The assistant needs your current return policy as its source, and ideally the ability to check the order date so it can say whether the item is inside the window.
- Confirm the order and item as with order status.
- Check the purchase date against your return window and any exclusions (final sale, personalized items, opened cosmetics).
- If eligible, send the shopper to your returns portal or start the request, and explain what happens next: label, drop-off, refund timing.
- If not eligible, explain why kindly and offer the alternative your policy allows, such as store credit or an exchange, or hand off to a person if the case is borderline.
Do not let the AI make exceptions or promise refunds outside the policy. Exceptions are a human decision.
Sizing: turn the size chart into a conversation
Sizing questions are where AI can genuinely beat a static page. Most shoppers never read a size chart closely. A short back-and-forth can do the reading for them.
Give the assistant your size charts per product, fit notes (runs small, relaxed fit, true to size) and, if you have them, garment measurements. Then let it ask one or two questions:
Example: Happy to help. What size do you usually wear in tops, and do you prefer a fitted or relaxed look? This shirt is cut relaxed, so most people take their usual size, or size down if they like it closer to the body.
Two rules keep sizing advice safe. First, the assistant should only use your data, never guess at measurements. Second, if the shopper is between sizes and your data does not settle it, it should say so honestly and mention your exchange policy.
Write the source material the AI will quote
An assistant can only be as accurate as what it reads. Most stores have policies written for lawyers and size charts buried in images, and neither works well as source material. Before launch, create a short, plain-language knowledge base the AI draws from:
- Shipping: processing time, carriers, delivery estimates by region, cutoff times for same-day dispatch, and what happens during holidays or sales.
- Returns and exchanges: the window, eligible and excluded items, who pays return shipping, how long refunds take, and how exchanges work.
- Sizing: size charts as text rather than images, fit notes per product, and how different product lines compare (a medium in one range versus another).
- Product facts: materials, ingredients, care, compatibility, and anything customers regularly ask before buying.
- Escalation contacts: support hours and what a shopper should expect after a handoff.
Keep one owner for this material. When a policy changes, the knowledge base changes the same day, or the AI will confidently repeat the old rule. A dated change log at the top of each document makes this easy to check.
It also helps to write a short style note: how formal the assistant should be, whether it uses the shopper's first name, which words your brand avoids, and how it signs off. The same answer lands very differently in a stiff tone versus your usual voice.
Guardrails that stop the annoyance
| Situation | What the AI does |
|---|---|
| Shopper asks for a human | Hands off immediately, no persuasion attempt |
| Damaged, missing or wrong item | Collects order number and photos, then hands off |
| Frustrated or upset tone | Acknowledges it, then hands off |
| Question outside its data | Says it does not know, offers a human |
| Refund exception or discount request | Hands off; never promises |
| Same question asked twice | Assumes its answer missed, offers a human |
Handoff should carry context. When a person picks up the conversation, they should see the order number, the question and what the AI already said, so the shopper never has to repeat themselves. Also be clear about who the shopper is talking to: label the assistant as automated, and give your support hours so people know when a human will reply.
A four-week rollout plan
- Week 1: Sort tickets, write or update the source content (shipping policy, return policy, size charts, fit notes), and connect order lookup.
- Week 2: Test internally. Have the team ask real questions from past tickets, including awkward ones, and fix wrong or clumsy answers.
- Week 3: Go live on order status only, or on one channel only. Read every conversation daily.
- Week 4: Add returns and sizing. Review handoff reasons and update the source content where the AI kept falling short.
After launch, keep a weekly review of a sample of conversations. Track how many are resolved without a handoff, how many are handed off and why, and any complaints about the bot itself. Those numbers tell you what to fix next.
Read the handoffs most closely. When the same reason keeps appearing, such as a product question the knowledge base does not cover or a policy the AI explains awkwardly, the fix is usually a few new lines of source content rather than a change to the tool. Over a few months, that steady tightening is what turns a basic assistant into one shoppers actually prefer to a contact form.
Building an assistant that knows your store
The difference between an annoying bot and a useful one is almost always setup: the right data connections, clear policies as source material, and handoff rules that respect the shopper. Our AI revenue concierge is built this way, covering your website chat and Instagram DMs with order lookups, policy answers and product guidance. If post-purchase questions keep flooding the inbox, pairing it with clear shipping and post-purchase email flows answers many of them before they are asked.