A customer types "I want to talk to a person". Does the AI stop, or try to persuade them again?
Short answer: In Merpati, when a customer types "I want to talk to a person", "can I speak to the owner", or "human please" — the AI stops immediately for that conversation, sends one confirmation message to the customer, and sends a summary to the owner: who, what they want, what has already been discussed. The AI does not resume until the owner switches it back on for that conversation. There is no round of persuasion.
Why is "one more try to persuade" the most damaging thing?
Because the customer has already decided. Every AI message after that is a message that is not listening — and that is what people remember:
"Maintenance guy: What do you think of our new AI customer service system? Me: I don't like it. It interrupted me… MG: yeah, a lot of our customers miss a human greeting. Me: That's not what I said."
— @memstephens, Threads (3,348 likes)
"Can I please talk to a human PLEASE"
— public comment, Threads, 2026
What exactly happens when a customer asks for a human?
| Step | What happens | Why |
|---|---|---|
| 1 | The AI detects the request — including "talk to admin", "I would like to speak with a human", "is the owner there?" | It understands meaning, not exact word matches. A system looking for the word "talk" misses "talking" — a case we once had to fix. |
| 2 | The AI stops for that conversation, immediately | The customer's next messages are not answered by the AI |
| 3 | The customer receives one message: "Your message has been forwarded to the admin, who will reply during working hours" | Once. Not "is there anything else I can help with?" |
| 4 | The owner receives a summary — on WhatsApp (the admin number they set) and by email | "Ms Ros asked about exchanging a size, ordered yesterday, wants a refund if exchange isn't possible" — not just "handoff occurred" |
| 5 | The conversation is flagged "needs you" in the inbox | The owner replies when ready, and switches the AI back on when they choose |
When else should the AI stop even if the customer did not ask?
- Refunds and money. The AI can explain the policy if it was written. But "I want a refund now" — the owner is notified, even while the AI continues to answer the policy. Money = human.
- Questions not in the owner's information. The AI says it will check with the owner — and the owner is told the question, so they can add the information.
- A customer turned down twice in a row. If the AI has already said "not sure" twice to the same person, the third time it goes quiet and notifies the owner.
Does this mean the AI is weak?
The opposite. An AI that stops at the right moment is the reason customers do not notice they are talking to an AI — every time they need a human, they get a human. Meta itself, when launching Business Agent, made "Let you decide when a team member steps in" one of its five headline promises. Everyone agrees it matters. What differs is how.
Test it from another phone
Sign up at merpati.my, set your admin number, and from another phone type "I want to talk to the owner". See what arrives — and what does not.
Sources
- Public Threads posts by @memstephens (3,348 likes), @edwinavoiceofai, and the comment "Can I please talk to a human PLEASE" — collected 2 September 2026, quoted verbatim.
- Meta Newsroom, 3 June 2026 — about.fb.com/news/2026/06/meta-business-agent