Why AI customer service is so annoying — and when it isn't
Short answer: AI customer service becomes annoying when it does three things — answers questions it does not know, repeats the same line without understanding the question, and refuses to hand the customer over to a human. The cause is not that it is an AI; the cause is a loop the customer cannot escape. The AI that does not annoy people is the one that knows when to stop.
What do customers actually hate?
We read hundreds of public comments about AI customer service before writing this. The one with the most agreement — 3,348 likes — was this story:
"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, 2026
Notice: she did not say "I want a human". She said the AI was not listening.
"PLEASE give customers a clear way to reach a human when AI cannot solve the problem. AI + humans. Not AI instead of humans."
— @edwinavoiceofai, Threads
And from an online shop owner:
"Shoppers do not resent automation. Nobody has ever closed a tab in anger because a machine answered them quickly and correctly. What they resent is the loop."
— @askzoye, Instagram, 31 August 2026
Why does the AI end up in a loop?
| Cause | What the customer experiences | Who can fix it |
|---|---|---|
| The AI was told to "answer every question" | Wrong prices, wrong stock, promises the business never made | The owner — remove that instruction |
| The AI reads from the internet or social posts | Confident answers that are wrong for this particular business | The system — restrict the source to the owner's information |
| No exit | "I understand your frustration. Is there anything else I can help with?" | The system — stop when asked |
Two of these three causes are not the AI's fault. They are design faults.
How does an AI that does not annoy people behave?
- It answers precisely from the owner's information. Prices, shipping, policies — whatever the owner typed. Not "please refer to our catalogue".
- It stays quiet when it does not know. "Let me check with the owner first" — and that message actually reaches the owner.
- It stops when asked. The customer types "can I talk to a person" — the AI stops immediately and sends a summary to the owner. No second round.
- It remembers the customer. Someone who asked about size M last week does not have to start from scratch.
Notice what is missing from the list: "smart". Every AI today is smart. What separates them is whether they know their limits.
How does Merpati handle those three causes?
Merpati is built around three guardrails, not three promises:
- The AI reads only the information the owner typed. When nothing matches, the AI is not called — a fixed reply goes out and the owner is told which question could not be answered.
- When a customer asks for a human — "talk to a person", "is the owner there?", "human please" — the AI stops for that conversation and the owner receives a summary: who, what they want, what has already been discussed. The full sequence is here.
- If the AI admits uncertainty twice in a row to the same customer, the third time it goes quiet and notifies the owner. Nobody wants to hear "I'm not sure" three times.
How should you evaluate any AI customer service tool?
Do not ask "is this AI smart". Ask every tool, including Merpati, these three questions:
- Where does it read from?
- What does it do when it cannot find an answer?
- When a customer asks for a human, what exactly happens?
If the answer to the third question is "it tries to persuade them one more time" — that is the loop.
Try it with your own number
The fastest way to test those three questions is not reading. Sign up at merpati.my with a Google account, connect the WhatsApp number you already use, and ask the AI something you did not enter. See whether it guesses, or stays quiet.
Sources
- Public comments on Threads and Instagram, collected 2 September 2026 — @memstephens (3,348 likes), @edwinavoiceofai, @askzoye (31 August 2026). Quotations verbatim; like counts as at the collection date.
- Meta Newsroom, 3 June 2026 — about.fb.com/news/2026/06/meta-business-agent