Once an independent store is up and running, the questions support answers every day turn out to be highly repetitive: how long shipping takes, whether returns and exchanges are possible, which size to pick, which payment methods you accept. These are the best questions to hand to AI first, so that people can spend their time on the cases that genuinely need judgement.
Step 1: Turn your FAQs into a knowledge base before letting AI answer anything
Asking AI to simply "help answer customer questions" is risky: it may invent policies you don't have (promising a return window you don't actually offer, for example). The right order is the reverse. First gather the real information, such as your shipping policy, return and exchange rules and size chart, into one document and keep it in Notion as a knowledge base. Then require the AI to answer only from that knowledge base, and to say plainly "a member of our team will confirm this for you" whenever it doesn't know.
Step 2: Use AI to draft replies, or connect it to your support tool
If your support team still handles messages by hand, feed the customer's question together with the knowledge base into ChatGPT or Claude and have it write a draft reply, which an agent checks before sending. This step saves the time spent digging through policy documents and working out how to phrase the answer. If message volume is high enough that you need replies to go out automatically (with no human in between), first list every type of question the knowledge base doesn't cover and set those to hand over to a human automatically, rather than letting the AI force an answer.
Step 3: Decide which questions must go to a human
We recommend routing the following straight to a person, without the AI handling them on its own:
- Individual cases involving a specific order (missing items, wrong items, damage)
- Messages where the customer is clearly unhappy, and complaints
- Refunds or disputes involving larger amounts
- Questions the knowledge base doesn't cover, or that the AI isn't confident about
Writing this boundary down clearly matters more than chasing how many messages the AI can answer automatically. The time you save should go into exactly these conversations that need a person.
AI customer service prompts you can copy
The more specific your instructions, the less likely the AI is to improvise. You can use the prompt below as it is; just replace the parts in square brackets with your own details:
You are the online customer service agent for [STORE NAME]. Your tone is friendly and concise, and you reply in the language the customer writes in.
Answer only from the knowledge base below. Do not add any policy, timeframe or promise that is not in the knowledge base.
If the answer is not in the knowledge base, or the question involves a specific order, a refund amount or a complaint,
reply only: "Let me check this with a colleague and get back to you as soon as possible." Then add [HANDOVER] on a new line at the end of the reply.
Knowledge base:
[PASTE YOUR SHIPPING POLICY, RETURN AND EXCHANGE RULES, SIZE CHART, PAYMENT METHODS, ETC.]
Customer message:
[PASTE THE CUSTOMER'S MESSAGE WORD FOR WORD]The [HANDOVER] tag at the end is very practical: an agent can see at a glance which drafts can go out as they are and which ones they need to take over.
AI can also help you build the knowledge base itself, but only to tidy up the format, not to fill in content:
Below are scattered customer service chat logs and policy notes from our store.
Please organise the questions customers ask repeatedly into a "question + standard answer" format.
Answers may only come from the original text I have provided. List any question that has no clear answer in the original text separately at the end and mark it [TO BE COMPLETED].
Original text:
[PASTE CHAT LOGS AND POLICY PAGE TEXT]The items marked [TO BE COMPLETED] are the ones where you need to confirm the policy yourself and then add it to the knowledge base.
Example: how a fictional yoga-wear store uses this process
Take a fictional independent store selling yoga wear as an example. Suppose its knowledge base says: orders ship within 1–2 working days; unworn items with tags intact can be returned or exchanged within 30 days of delivery; sizing follows the size chart on the product page.
A customer writes: "I usually wear an M. Will the waist on these leggings be on the small side? Also, my order from last week still hasn't arrived."
With the prompt above, the AI's draft will roughly come in two parts: for the sizing question, it quotes the size chart and suggests the customer compare their waist measurement against it; for the order question, it replies "Let me check this with a colleague and get back to you as soon as possible" and adds the [HANDOVER] tag.
When the agent takes over, they usually change two things:
- Make the sizing advice specific: how stretchy the fabric is, and whether to size up. If the knowledge base covers it, let the AI quote it; if not, the agent answers and adds the answer to the knowledge base while they're at it
- Look up the order manually: check the shipping status in the back office, write what they find into the reply, and send it all together
In this example, AI saves the time spent looking up the size chart and wording the reply; checking the order and offering a concrete solution is still done by a person.
How to write knowledge base entries so the AI answers accurately
- One topic per entry: write "return and exchange conditions" and "how long refunds take to arrive" as separate entries, so the AI doesn't mix them up when quoting
- Spell out scope and exceptions: which products can't be returned or exchanged, and which countries or regions have different shipping costs. List exceptions separately
- Use the customer's wording as the heading: "How long until I get my order?" is closer to what customers actually write than "Shipping time information", and it's easier for the AI to match
- Add an updated date: when a policy changes, you can tell at a glance which entries are out of date, which makes regular checks easier
- Leave out anything you're unsure of: it's better to leave a gap and let the AI hand over to a human than to write an answer that's only "roughly" right
Common pitfalls with AI customer service on an independent store
- Dumping your whole website into the AI as the knowledge base: marketing copy and policies get mixed together, and the AI may quote a limited-time free shipping promotion as if it were a permanent policy. Only put settled rules in the knowledge base.
- Testing only normal questions before going live: deliberately ask a few awkward questions the knowledge base doesn't cover, and see whether the AI dutifully hands over to a human instead of making up an answer.
- Keeping the knowledge base only in Chinese when customers ask in English: AI can translate, but details such as size units and date formats are easy to get wrong, so agents should pay particular attention to them when checking.
- Setting up handover with nobody to pick it up: if you haven't agreed who handles [HANDOVER] conversations and within what time, customers end up waiting even longer and become even less happy.
FAQ
How detailed does the knowledge base need to be? Covering the 10–20 questions your support team is asked most often is enough to see results. After that, keep adding to it based on the new questions agents actually receive, rather than aiming for something huge and complete from day one.
Could wrong AI answers lead to customer complaints? If you strictly follow the rule of "answer only from the knowledge base, and hand over to a human when unsure", the risk mainly comes from the knowledge base itself being inaccurate or out of date. That's why someone needs to check and update it regularly, rather than writing it once and forgetting about it.
If customers ask in different languages, do I need a knowledge base in each language? Not necessarily. You can maintain the knowledge base in a single language and have the AI reply in the customer's language, so a policy change only needs updating in one place. For high-frequency questions in your main markets, though, we recommend agents spot-check a few translated replies to make sure key information such as return conditions and sizing hasn't been mistranslated. For choosing a translation tool, see DeepL vs ChatGPT for product translation.
How do I know when the knowledge base needs adding to? Keep an eye on conversations tagged [HANDOVER]. If the same question keeps getting handed over, it has already become a common question. Write a standard answer for it and add it to the knowledge base, and next time the AI can draft the reply directly.
Do agents still need to read AI drafts word for word? At first, we recommend checking every one to confirm the AI really is answering only from the knowledge base. Once it has been running for a while, you can focus on sentences about timeframes, return conditions and amounts, and just skim the rest, which is mostly tone.
For building the knowledge base, see Notion; for generating replies, you can use ChatGPT or Claude, and there's little difference between them for customer service. If you want to link enquiry handling and customer service automation into one complete workflow, see how to auto-reply to B2B sales enquiries with AI. The same idea of "fix the rules first, then let AI apply them at scale" also works for email marketing; see how to run email campaigns with AI and Klaviyo.
If you haven't settled on a support tool yet, start with Gorgias vs Tidio for Shopify support; to find out where these common questions come from in the first place, see how to analyse customer reviews in bulk with AI.
