An online store gets the same questions every week. Where is my order, how long does delivery take, can I return this, will a medium fit me. The answers are written down somewhere already, usually in your head and in six hundred previous replies sitting in your sent folder.
The cost is not the typing. It is the interruption. Each question arrives while you are packing, photographing or reconciling, and each one takes you out of the job you were doing and puts you back three minutes later.
This guide builds a support setup that answers the repeat questions from your own material and hands everything else to you with the context already gathered. Budget about three hours to set it up.
What you are building
Three parts. A single answer bank written in your own wording, an AI drafting step that answers from that bank and nothing else, and a routing rule that sends anything unusual straight to a person. Repeat questions get a reply within minutes at any hour. Anything about a specific order problem, a refund decision or a complaint lands with you.
Tools: the helpdesk or shared inbox you already answer from, your store platform for order and shipping status, an automation tool such as Zapier, Make or n8n, and an AI assistant such as ChatGPT or Claude.
Step 1: find out what you are actually asked
Open your last two hundred customer messages and sort them into buckets by hand. It takes about an hour and it is the only hour in this build you cannot skip, because everything later depends on the list being real rather than what you assume people ask.
- Shipping: when will it arrive, where is it now, do you ship to my country, how much is postage.
- Returns and exchanges: how long do I have, who pays, what condition, how do I start one.
- Sizing and fit: measurements, what to order between sizes, how a fabric behaves after washing.
- Product questions: materials, care, compatibility, what is back in stock and when.
- Everything else: a damaged parcel, a wrong item, a payment problem, a complaint.
The first four buckets are the automation. The fifth is not, and pretending otherwise is how a store ends up with an upset customer explaining a broken vase to a machine three times.
Step 2: write the answer bank once
The AI does not know your returns window, your postage bands or whether your shirts run small. It has to be told, in one document, in your words.
Write one short answer per question in the first four buckets. Use the exact policy, the exact number of days, the exact link. Where a sizing answer depends on the product, write the rule you use when a customer asks you directly, for example that a particular line runs narrow and somebody between sizes should take the larger one. That is the knowledge the model cannot guess and the reason your version will read better than a generic bot.
Step 3: the drafting prompt
The safety of this whole build sits in one instruction: answer from the bank, or say you cannot. Models are helpful by default, which here means they will invent a delivery window rather than admit the document does not contain one.
You are drafting a reply for a customer of an online store. Use only the answer bank below and the order details provided. ANSWER BANK: [paste it]. ORDER DETAILS: [paste the fields from the store, or write NONE]. CUSTOMER MESSAGE: [paste it]. Rules: answer only what the bank covers. If the message asks anything the bank does not answer, or mentions damage, a wrong item, a payment problem or a complaint, reply with the single word ESCALATE and nothing else. Never invent a delivery date, a price or a policy. Under a hundred and twenty words, no exclamation marks, no emoji, plain sentences. Sign off as [your name or the store name].
The single word ESCALATE is what the automation routes on. Ask for a sentence instead and you will spend an evening writing a rule that tries to guess what the model meant.
Step 4: wire it up
- Trigger on a new message arriving in your helpdesk or shared inbox.
- Look the customer up in your store platform by email address and pull the most recent order: status, tracking link, items and the date it shipped. If there is no order, carry on with NONE.
- Call the AI assistant with the prompt, the answer bank, the order fields and the message text.
- If the reply is ESCALATE, assign the ticket to yourself with the order details attached as an internal note, and send the customer one short line saying a person is looking at it.
- Otherwise save the draft on the ticket. For the first two weeks, approve every draft by hand.
- Log every reply in a sheet with the bucket, the customer and whether you edited it.
That log is how you decide what to send automatically later. When a bucket has run for two weeks without you changing a single draft, move that bucket to automatic sending and leave the rest in drafts. Shipping status is usually the first to earn it. Sizing is usually the last, because that is the one where being wrong costs you a return.
Step 5: close the loop on the questions themselves
Every repeat question is a gap in a product page. The support queue is the cheapest product research you will ever have, and it is already written down.
Here is a list of customer support messages from the last month, with the bucket I assigned each one. Give me: the ten questions asked most often, in the wording customers used; the three product pages that generate the most questions and what is missing from them; five lines I could add to a shipping or returns page that would remove whole categories of question; and anything two or more customers were confused by in the same way. Use only the messages given. Do not invent examples and do not estimate numbers I have not provided.
Act on that list once a month. Every question you answer properly on the page is one the automation no longer has to draft, and one fewer reason for somebody to abandon a basket.
What changes
The repeat questions stop interrupting the day. The hard messages arrive with the order history already attached, so you start from the facts instead of hunting for them. And the queue turns into a list of things worth fixing rather than a list of things to get through.
Connecting a helpdesk, a store and an assistant like this is the standard build in Workflow 05: Build Automation Workflows, the AutomateFirst module about wiring the tools you already pay for so the repeatable half of the job runs in the background. This guide is that build in full, free. The playbook has six more for one payment of $297, with a 30 day refund.
Workflow 05: Build Automation Workflows
Six more workflows, written exactly like this one
The AutomateFirst playbook covers seven repetitive jobs end to end, with the prompt library, PDF checklists and lifetime access. One payment of $297, no subscription.
Other guides