Module 1 · Start safely
A person sends the reply, not a machine
In short
AI may draft chat replies, but a person sends them. Automatic replies that promise prices, stock, shipping or refunds without a check can bind your shop to a wrong promise. Decide who checks and who sends.
After this lesson you can
- Tell an AI draft reply apart from a message sent in your shop's name
- Reject automatic replies that promise prices, stock or delivery times without a person checking
- Decide who in your small team checks and sends customer messages
Terms in this lesson
Understand
In an online shop, a chat reply is a promise. When the admin writes 'in stock, ships today', the buyer pays because of that sentence. When the admin writes 'full refund available', the buyer claims a full refund. A message from your shop's account is read as your shop's word, whoever or whatever drafted it.
That is why we separate two things. A draft is text prepared by AI or a template, not yet sent, and allowed to be wrong. A message is text already sent to a customer, in the shop's name, and it must be right. Between the two there is always one person who reads, checks the facts, then presses send. This principle is called human in the loop, and its basis is covered in the humans decide lesson.
The problem is that a language model does not know your stock today, the shipping cost to the buyer's city, or your shop's refund policy, unless you give it to it. Asked to reply to 'still available?', the model answers confidently, because that is the reply shape it has seen most often. A confident answer with no data is invention.
Many business chat apps and marketplaces offer automatic replies, and some now use AI. Such automation is safe when its content promises nothing that changes. A greeting, opening hours and 'we got your message, the admin will reply before 8 pm' are examples of safe automatic replies. Their content is fixed, and you checked it once when setting it up.
What is not safe is an automatic reply that states a price, stock, shipping cost, arrival time, discount or refund decision. All of these change every day and need checking against your shop's records. If a machine answers, nobody checks. Get it wrong once, for example stock is gone but the reply says available, and you must cancel the order, and on a marketplace seller cancellations can lower your shop's rating.
In a small shop, whoever holds the shop phone may be the owner, a child, a sibling or one admin. Write down who may send messages in the shop's name and who decides binding matters: refunds, exchanges, special discounts. The admin may use AI drafts for ordinary questions, but for money decisions they ask the owner's approval first. This simple rule stops two people answering the same customer with different promises.
Also keep a short note for binding decisions: date, order code, what was promised and who approved it. If the customer comes back a week later, whoever holds the phone can see what was promised.
Before
The prompt
Reply to this buyer's message: 'Is the green robe in size L still available? If I order now, can it reach Surabaya by Friday?'
What came back
Hi! The green robe in size L is in stock. Check out now and it's guaranteed to reach Surabaya by Friday. Order before it runs out!
After
The prompt
Goal: I need a draft reply to a stock and arrival-time question, which I will check and send myself. Reader: a buyer in the chat app asking before ordering. Material: stock of green robe L per my records today: ‹stok›. The courier I usually use to East Java estimates 2 to 4 working days. I ship daily at 3 pm. Limits: do not guarantee an arrival date; give the estimate and say timing depends on the courier; do not state stock if ‹stok› is empty. Shape: at most 50 words, polite, no excessive exclamation marks. Check: mark the parts I must verify before sending.
What came back
Hi, the green robe in size L is currently available [check stock]. If the order comes in before 3 pm, we ship today. The courier estimates 2 to 4 working days to Surabaya [check courier], so we cannot promise Friday. Shall we process it for you?
What changed
- The first draft guarantees stock and an arrival date with no data at all.
- The second draft uses stock from your records and the courier's estimate, then marks what to check.
- You are still the one who sends; the check marks come out once you have verified.
Common mistake
- Mistake
- Switching on an AI-based automatic reply that answers every buyer question, including price, stock and refunds, while the shop is busy.
- Why it hurts
- The busiest time is exactly when stock changes fastest. The machine can promise goods already sold out or a refund that breaks policy, and the buyer holds a screenshot of that promise.
- Do this instead
- Use automatic replies only for greetings and a notice that the message was received. Prepare templates for common questions, and let a person fill in the figures and send.
Try
Sorting card
Sort these replies: safe to send automatically, or must a person check and send?
Shipping cost to the buyer's address and the total due
A notice that the shop is closed for Eid, with the reopening date
Greeting: thanks for reaching out, the admin replies from 8 am to 8 pm
A refund decision for a damaged parcel
A link to the shop's product catalogue
An 'in stock' answer for the product asked about
6 cards left
Your small shop has an owner and one admin. Open each card to see who should hold the decision.
The admin, using the product fact sheet. AI drafts are fine if their content is matched to the fact sheet.
The owner. A discount changes margin; the admin passes it on after the owner agrees.
The owner, based on evidence and the written policy. The admin records the decision and the order code.
The owner checks the new text before it goes on, because it will be sent to everyone.
Take home
Before pressing send
- I read the whole draft, not just the first sentence
- Price, stock, shipping and delivery time are matched to today's records
- There is no guaranteed or certain for anything that depends on the courier
- Refund or discount promises are approved by the person in charge
- The shop's automatic replies hold only fixed content
- I record binding decisions with their order code
What went wrong for us
AI drafts, a person sends
- Context
- In apps we run, AI can draft messages, such as messages to a team or draft replies to candidates. Technically it would be very easy to have those drafts go out directly.
- What tends to go wrong
- The habit we keep is this. An AI draft is never sent on its own. A person reads it, checks its facts and presses send; messages to many people at once are not registered as an action AI may run.
- What it costs if ignored
- A sent message is read as the sender's word. A wrong promise about price, stock or money cannot be taken back once read.
- Root cause
- A model writes confidently even when it lacks the data needed, and automatic sending removes the only checkpoint.
- How we do it
- Every feature of ours in which AI writes a message stops at a draft (for example, the logistics app's AI draft tool says "not sent" and the person sends from Communications). The send button is in the hands of the person who holds the account.
- The rule we keep
- The design question: at which point does a person read before a message goes out? If there is none, the feature is not switched on.
The lessonAI may draft the reply; a person sends it and answers for it.
What it means for your workUse automatic replies only for fixed content, and let a person send answers about price, stock and money.