Does the AI that replies to your customers use your business information — or make it up?
Short answer: It comes down to one thing: where the AI reads from. An AI that reads from the internet or social posts will make up an answer when it cannot find one — confidently. An AI that can only read what the owner typed, and that is blocked from answering when that information is missing, cannot invent a price because it has no price to invent. How to find out: ask it something you deliberately left out.
Why does an AI make things up?
Not because it is malicious — because it is built to answer. When asked, it looks for the most plausible answer; when the real information is missing, "most plausible" becomes "looks most correct". "Yes, cash on delivery is fine" sounds plausible. "Refund within 7 days" sounds plausible. Neither may be your policy.
How many vendors talk about this risk?
We read 162 active ads in the WhatsApp AI/chatbot category in Malaysia (Meta Ad Library, 2 September 2026). Not one mentions the risk of the AI answering wrongly. All of them say "AI replies 24 hours". None says "AI stays quiet when it doesn't know" — perhaps because the second line does not sound as good. But it is the only line that matters.
What is the difference between "reads from the internet" and "reads from the owner's information"?
| Reads from the internet / social posts | Reads only from the owner's information | |
|---|---|---|
| What it knows | A lot — including things that are wrong about your business, old posts, competitors' prices | Only what the owner typed — prices, stock, policies, one product per entry |
| When it cannot find an answer | Answers with the most plausible thing | Reaches the edge — and what it does at that edge has to be tested |
| Owner's control | None | Full — change the entry, the answer changes |
How does Merpati stop the AI from making things up?
"We use advanced AI" is not an answer. This is what actually happens:
- The AI is only given the owner's information. Not the internet. Not other customers' chats. Every reply is generated from entries the owner typed.
- When no matching information exists, the AI is not called at all. This is not an instruction to the AI ("please don't make things up") — it is a guardrail in the system. If the customer's question finds no match, a fixed reply goes out: it will check with the owner. The AI never gets the chance to be creative.
- Numbers are checked. If a reply contains a price or figure that does not exist in the owner's information, it is held back. The same for opening hours: if the hours were never typed, the AI cannot say "we open at 9am".
- The AI cannot claim to have done something that did not happen. "I've set your booking" only goes out if that booking genuinely exists in the records.
We write this not to be admired, but so that you ask the same questions of every tool you compare — and see who can answer specifically.
How do you test it yourself in 2 minutes?
- Enter 5 products with prices.
- Leave out your refund policy.
- Ask: "can I get a refund if it doesn't fit?"
If it replies "Yes, within 7 days" — close it. If it replies "Let me check with the owner first" — continue to the full five tests.
Run that test in Merpati's test area
No message reaches a customer. Sign up at merpati.my with a Google account, enter 5 products, and ask a question you did not provide an answer for.
Sources
- Meta Ad Library, Malaysia, 162 active ads in the WhatsApp AI/chatbot category, collected 2 September 2026 — 0 ads mention the risk of the AI answering wrongly. Raw data kept in Merpati's internal records.
- Public Threads posts by @kairelkeflee, @askzoye — collected 2 September 2026.