Make the chatbot leave price, colour and delivery date blank in customer replies
A customer asks about colour, price and delivery at once, and a sales ledger usually answers only part of it. Let AI write the wording, and take every number from the ledger.
It is nine in the evening and a customer messages your shop page: “Is the A thermos flask still available in blue, how much is it, and can you deliver tomorrow afternoon?” Before you ask AI anything, the first job is to open your sales ledger and find the row for flask A.
The ledger row and the message drafts below are made-up examples. Suppose the row for flask A reads:
| Product | Price | Stock | Estimated delivery |
|---|---|---|---|
| Thermos flask A | 320,000 VND | 4 | 1–2 days |
Read it closely and you see the ledger answers only part of the question. The price is there: 320,000 VND. Stock is there too, but it is a total of 4 units, not split by colour, so the ledger does not answer “is it still in blue?” And “1–2 days” is a general estimate; it says nothing about whether delivery to this customer’s address can happen tomorrow afternoon.
Three questions, three different levels of certainty, and no matter how smooth a message sounds, it cannot change that. So split the work: AI writes the words, the ledger supplies the numbers.
Ask AI for the wording and leave blanks for the numbers
Paste the customer’s question into an AI assistant together with a narrow instruction. Remove the customer’s name, phone number and address from what you paste.
Write a short, friendly reply to this customer. Do not fill in the price, the quantity in stock, the colour availability or the delivery date yourself. Wherever a number is needed, leave a blank in the form [CHECK: PRICE], [CHECK: BLUE], [CHECK: DELIVERY DATE]. Do not use the words “definitely” or “guaranteed”.
(This is the prompt translated from the Vietnamese original. Write it in whatever language you and your customers actually use.)
The draft you hope to get back looks like this:
Hello, I’ll check flask A for you. The price is currently [CHECK: PRICE]. The blue one is currently [CHECK: BLUE]. For delivery tomorrow afternoon, I’ll look at your address and get back to you: [CHECK: DELIVERY DATE].
The square brackets make the gaps visible. A draft as smooth as “Yes, blue is in stock, 320,000 VND, delivered tomorrow!” sounds far nicer, but then you can no longer tell which of those numbers is real.
Fill each blank from the ledger, then send
Go back to the row for flask A and deal with the brackets one at a time:
- Price: the ledger says 320,000 VND. Fill it in.
- Blue: the ledger only has a total of 4 units. Look at the shelf or ask whoever looks after the stock. If you don’t know yet, keep the line “I’m checking blue and will let you know”.
- Delivery date: “1–2 days” is an estimate. Write “tomorrow afternoon” only after you have asked the courier or know today’s order cut-off time. If not, write “expected 1–2 days, I’ll confirm once I have your address”.
The message you send could be: “Hello, flask A is 320,000 VND. I’m checking the blue one on the shelf and will tell you in a few minutes. Delivery is expected in 1–2 days, and I’ll confirm exactly once I have your address.” This promises less than the customer hoped for, but every number in it traces back to a row you have just opened.
Test it when the ledger is missing a column
To find out whether your instruction is tight enough, delete the stock column from the made-up data and see whether the draft invents “in stock” or “almost sold out”. If the draft says “in stock” when the ledger records nothing of the kind, don’t send it. Add the line “if data is missing, write [DATA MISSING] and stop” to the instruction and try again. If it keeps happening, take AI out of the numbers: use a fixed message template, fill in the numbers yourself, and ask AI only to polish the wording.
One rule for a small shop
Before you press send, find every number, date and any word that sounds like a promise (“definitely”, “right away”, “arrives tomorrow”). Each one must point to a ledger row you have just opened. If it can’t, delete it or change it to “I’ll check and get back to you”.
This is not worth the effort if your ledger is slow to update or messy: AI writes quickly, but you still lose time correcting the numbers. If you use real customer data, follow your shop’s own policy and the terms of the tool you are using.
Sources
- Prompt engineering guide — OpenAI — The primary guide on writing clear instructions and giving a model context; the article applies this principle to limit what the AI is allowed to write.
- Artificial Intelligence Risk Management Framework: Generative AI Profile — NIST — The primary document on generative AI risk management; used to cross-check the principle of checking outputs against the context of use.