AI customer service: how an online store answers fast without losing the customer

In an online store, most customer service messages are the same ten questions in different words: when will it arrive, how much is shipping, how do I exchange a size, will this work for me. Answering them by hand takes hours every week, and the customer waits. AI customer service for online stores promises instant answers at any hour, but it can also drive customers away with vague replies or no way to reach a person. This guide covers which questions AI can close alone, where it works better (website chat or WhatsApp), how to hand a chat to your team without losing the customer, and how to measure it. It is part of the guide to AI for business.

What is AI customer service

AI customer service is automated support powered by a language model. The model reads the customer's message, looks for the answer in the business's own information and writes a reply in natural language. Unlike a menu bot that says "press 1 for shipping", an AI agent understands free text, typos and two questions in one message.

The important distinction is where the AI takes its answers from:

• The model's general knowledge: the answer sounds convincing, but it may invent a policy.
• Only the business's knowledge: it is limited to what you wrote, and that is exactly its strength. It tells the customer what your business actually commits to, and admits when it does not know.

Automated customer support: what AI closes

The rule is simple. AI closes questions whose answer is written down and the same for every customer. Questions that require checking one customer's details, a decision about money, or sensitivity go to your team.

Look closely at the "specific order" and "stock" rows. Many tools answer them confidently without any access to the data, and the result is worse than "let me check with the team". Ask every vendor directly where their tool gets order status and stock from.

Website chat vs WhatsApp

The two channels serve different moments in the customer's journey:

• Website chat: meets the customer before the purchase, while they are on a product page making up their mind. The questions are about specs, sizing, shipping and returns, and a chatbot that answers right away can keep them on your site. How to choose and install one is covered in the guide to choosing a chatbot for your website.
• WhatsApp: meets the customer before and after the purchase, with a delivery question, a size exchange or a request for an invoice. In markets where WhatsApp is the default messaging app, it is where customers turn first.

AI customer service on WhatsApp

For an AI agent to reply on WhatsApp you need the WhatsApp Business API, because the free app has no bot. Meta's rules apply too:

• The 24-hour window. You can reply freely only within 24 hours of the customer's last message. After that, you can reach out only with an approved template.
• Business use only. Since January 2026, Meta does not allow general-purpose AI assistants on the API. An agent that answers the business's own customers about its products and services is allowed.
• A path to a person. Meta expects automated experiences to offer a clear way to reach a human.

The differences between menu bots, chatbots and AI agents on WhatsApp are covered in the WhatsApp bot guide.

Handoff from AI to your team

The handoff is where most AI customer service setups fail.

A customer who wrote three messages to a bot and then has to explain everything again to a rep is angrier than one who simply waited. The same goes for a customer left not knowing whether anyone will reply. Four principles:

1. The bot says what is happening. "I don't have access to order details, so I'm passing this to the team; they'll check and reply here" beats "I'll check and get back to you", which nobody will do.
2. The team sees the whole conversation. The rep continues the same chat and does not ask again what the customer already wrote.
3. The bot steps aside. Once a rep joins the chat, the bot stops replying there, so the customer does not get two conflicting answers.
4. Someone owns it. Handed-off chats must land where your team actually looks, not in an inbox nobody opens.

Handoff in StoreChart

StoreChart's WhatsApp AI agent replies from the shared WhatsApp inbox, so your team sees every reply it sends. Its knowledge comes from a scan of your site, answers extracted from past chats that you approved, and content you add.

• A mode per conversation: off, supervised or auto. In supervised mode the agent replies until someone from your team writes in that chat, including from the WhatsApp Business app on a phone, and then it stops in that conversation.
• A question about an order: the agent answers what your policy says, such as delivery times, says plainly that it has no access to that specific order, and passes the chat to your team.
• A question about stock right now: the same. The agent does not guess; it passes the chat on.
• A flag in the inbox: a handed-off conversation is flagged as needing attention.
• Handoff settings: handoff keywords, a minimum confidence level and working hours.
• When there is no answer: the agent sends a fallback message, or stays silent and leaves the chat to your team.

Example: a clothing store

Here is what that looks like in a conversation, at a hypothetical clothing store:

Customer: I ordered on Sunday, when should it arrive? And can I exchange the size if it doesn't fit?

Agent: Orders ship within 2 business days and usually arrive within 3 to 5 business days. You can exchange a size within 14 days of receiving it. As for your specific order, I'm passing it to the team; they'll check and update you here.

Team member: I checked, your order shipped yesterday and arrives tomorrow.

What happened in that chat:

• The agent answered both policy questions from its knowledge.
• The team member checked the order itself, with everything written before in front of her.
• The customer did not have to explain her question again.

The figures in this article's examples are illustrative only and do not reflect actual customer data.

Bubble, the website chat, works differently: it has no live handoff to a rep. When it has no answer, it points the visitor to your store's WhatsApp or phone, and your team takes it from there.

Measuring AI customer service

The number of conversations a bot handled says little on its own. These are the measures worth tracking:

• First response time. How long a customer waits for the first reply, before and after.
• Conversations closed without a person. The share of chats that ended without your team stepping in, and without the customer coming back with the same question.
• Chats handed to the team, and why. If most are about the same topic, the agent's knowledge is missing something.
• Unanswered questions. In StoreChart they are saved as knowledge gaps, and an answer you write there goes into the agent's knowledge.
• Complaints about answers. Every wrong answer is a fix in the knowledge, not just an apology.

In the first weeks, read conversations, not just reports. Most fixes come from reading ten real chats.

Common AI customer service mistakes

• Launching without current knowledge. An agent answering from an old shipping page promises customers things your business no longer does.
• Hiding that it is a bot. A customer who finds out halfway through that they are talking to a machine feels misled.
• Leaving no route to a person. Even the best agent does not know everything.
• Letting AI decide about money. Refunds, compensation and discounts stay with your team.
• Forgetting your other channels. Inquiries and prospects arriving from different places belong in one place, which is the job of customer relationship management.