Run a Shopify store with one or two people and the support inbox fills up with the same handful of questions: where's my order, can I return this, what size should I get, which payment methods do you take. That repetition is exactly what AI is good at; the trick is setting it up so it doesn't start inventing policies you never agreed to.
Step 1: Write down your real policies before you let AI near a customer
The mistake most small stores make is asking an AI to "just answer customer questions" straight away. It will happily answer, and sometimes invent a return window or shipping promise you don't actually offer. Do this in the other order: write your shipping timelines, return rules, and size chart into one shared doc first, then instruct the AI to answer only from that doc, and to say "let me check with the team" for anything it isn't in.
Step 2: Use AI to draft replies, not send them blind
While your team is still reading every message, feed the customer's question and your reference doc into ChatGPT or Claude and have it draft a reply for a human to check before it goes out. This is where most of the time actually gets saved: you're no longer re-reading the policy doc and wording the reply from scratch for every ticket, just approving or lightly editing a draft.
Step 3: Wire the routine cases so they don't need your attention every time
Once the draft-then-approve step is working, Zapier can trigger on a new support email or contact-form submission, call the AI to draft a reply from your reference doc, and drop it into a queue for a quick one-click approval. Only the most clearly repetitive, lowest-risk questions (order status, standard return policy) are worth routing to auto-send without a human glance; everything else stays in the approval queue.
Step 4: Draw a hard line for what always goes to a person
Set these categories to route straight to a human, no AI draft involved:
- Anything about a specific order gone wrong (missing, damaged, wrong item)
- A customer who is clearly upset or escalating
- Refunds or disputes above whatever dollar threshold your team sets
- Any question your reference doc doesn't cover
That boundary matters more than maximizing how many tickets AI touches: the time you free up should go toward the conversations that actually need a person.
What to put in your support reference doc
The doc only works if it covers the questions you actually get. A starting checklist for a small Shopify store:
- Shipping: processing time before dispatch, delivery windows for each region you ship to, which carriers you use, and what happens when tracking stalls.
- Returns and exchanges: the window, condition rules, who pays return postage, items that can't be returned, and how refunds are issued.
- Sizing and fit: your size chart, how the cut runs, and what to suggest when a customer is between sizes.
- Payments: the methods you accept and what to say when a payment fails.
- Order changes: whether address or item changes are possible after ordering, and until when.
- Things you never promise: delivery dates you can't control, discounts outside a running campaign, refunds before an item is back with you.
Write each rule the way you'd say it to a customer, and put a "last updated" date at the top. When a policy changes, update the doc before anything else, because the AI drafts will follow whatever it says.
A prompt for turning past replies into a first draft of the doc
If your policies mostly live in your sent folder, AI can gather them into a draft. It shouldn't decide them.
Below are [NUMBER] support replies that [STORE NAME] has sent to customers.
Extract every policy statement they contain (shipping times, return rules, sizing advice, payment methods, order changes) into a list grouped by topic.
For each statement, quote the reply it came from.
If two replies contradict each other, list both under "CONFLICTS" and don't choose between them.
Don't add any policy that isn't stated in the replies.
Replies:
[PASTE PAST REPLIES, WITH CUSTOMER NAMES, ADDRESSES AND ORDER NUMBERS REMOVED]The CONFLICTS list is the useful part. Every item on it is a decision the owner makes once, so the AI never has to guess between two answers again. For a reply-drafting prompt to use once the doc exists, see the template in ChatGPT vs Claude for support replies.
Worked example: an exchange request at a fictional store
Take a fictional two-person Shopify shop selling linen bedding, "Northbay Linen", as an example.
Input: a customer writes, "I ordered a duvet cover in the wrong size. Can I swap it, and how long will it take?" The reference doc says exchanges are accepted within the store's return window for unused items in their original packaging, and that the customer pays return postage.
What the AI returns: a draft confirming the exchange is possible, restating the unused-and-packaged condition, explaining that return postage is on the customer, and adding "your new cover will arrive within three days of us receiving the return."
What a person changes: the doc never mentions a turnaround for exchanges, so that last sentence goes. The owner replaces it with "we'll send the new size as soon as your return arrives and email you the tracking link." Then she adds an exchange turnaround line to the doc so the next draft doesn't need the same fix. That loop, where each correction becomes a doc update, is what makes the drafts better week by week.
Common mistakes when automating Shopify support
- Turning on auto-send before the drafts are reliably right. Stay on draft-then-approve until you're rarely editing the routine categories.
- Letting the doc drift. A sale, a new carrier or a changed return window that isn't in the doc means the AI keeps quoting the old rule.
- Pasting more customer data into the AI than the reply needs. The order number, items and status are usually enough.
- Hiding the way to a person. Every reply should make it easy to ask for a human, especially in the auto-sent categories.
- Counting tickets touched instead of time saved. The goal is more attention on the hard conversations, not a higher share of AI replies.
FAQ
How much time does this realistically save a 1-2 person team? Most of the savings come from cutting the time spent finding the right policy and wording the reply, not from eliminating messages entirely. A team that reads and approves every draft still saves real time; a team that also auto-sends the lowest-risk category saves more, at the cost of needing the reference doc to stay accurate.
What if a customer notices they're talking to AI? Nothing to hide: most stores are upfront that first-response drafts may be AI-assisted and that a human reviews anything that needs judgment. The reference-doc-only rule is what keeps the AI from promising something the store can't back up, which matters more to the customer than which tool wrote the first draft.
Should I use a helpdesk app or just Zapier and a shared inbox? For one or two people, the draft-and-approve setup above can start with your existing inbox and Zapier. Once you want to work tickets alongside order details in one place, or add a live chat widget, a dedicated helpdesk becomes worth comparing. See Gorgias vs Tidio for Shopify support.
How often should the reference doc be reviewed? Whenever a policy, carrier or promotion changes, and as a habit whenever you catch yourself editing the same kind of draft twice. A repeated fix usually points at a missing or unclear line in the doc.
Can AI answer "where is my order?" on its own? It can draft the reply, but it only knows what you give it. If the order status and tracking link come from your store data in the same step, it's one of the lowest-risk categories to consider for auto-send. If a parcel has stalled or looks lost, that's an order gone wrong, and it goes to a person.
For drafting, ChatGPT and Claude both work well here; pick whichever your team already has a subscription for. To wire the whole thing into a queue, see Zapier. Once your listings are also up to date, turning one product photo into a full marketplace listing covers the other half of a small store's AI workflow. For a closer look at how ChatGPT and Claude compare specifically on support replies, see ChatGPT vs Claude for support replies. If most of your support actually happens on WhatsApp instead of a Shopify inbox, see automating WhatsApp Business support with AI. For the full picture of every tool and playbook on this site grouped by scenario, see how to build an AI tool stack for cross-border ecommerce.
