Most articles about WhatsApp automation talk about what you can automate. This one covers the opposite. Some conversations get expensive when a machine answers them, and the damage never shows up in the response time report. It shows up in the customer who cancels three months later without quite saying why.
One rule runs through everything below. AI can answer what has already been decided. Anything that needs a decision goes to a person. These are the situations where that rule matters most.
1. Anything involving money
Prices, discounts, penalties, invoices, refunds, plan changes. Any number with a currency sign.
AI tends to be helpful to a fault. Faced with "can you give me a discount?", the most likely model reply is some form of yes, or an "I'll check" that the customer reads as a promise. A wrong number written in a chat becomes a screenshot, and the screenshot becomes an argument.
What AI can still do: acknowledge the request, gather the history (contract, deliveries, earlier agreements) and draft a reply for you to approve.
2. Deadlines and new scope
"Can you deliver by Friday?" "Could you add one more screen?" These look like small questions and they decide your week.
In agencies, software houses and consulting, scope that grows without a conversation is the biggest leak in your margin. An automatic "sure, no problem" turns a request into a commitment. The person who owns the account should answer these, even if it takes a few hours longer.
3. An angry customer
Complaints, chasing something late, "this is unacceptable". An upset customer wants two things: to be heard by someone accountable, and to see a plan.
AI usually misses in one of two directions. It either over-apologizes with scripted empathy that makes things worse, or it defends itself with facts in a cold way. Either way, the customer realizes there is nobody real on the other side at the exact moment they need someone.
What AI can still do: pull the history of the problem (when it started, what was already tried) so you can reply with facts.
4. A mistake that was yours
The system went down, a delivery shipped broken, you forgot something. Telling a customer about your own error is one of the most delicate moments in support. It means owning it once, talking about impact and the fix, and not groveling.
Automated text about your own error tends to slide into over-confession, one apology after another, or into vague language ("we experienced an issue"). That message deserves two minutes of a human's time. We cover the wording habits that give a bot away in the tells that make AI text read like a bot.
5. "Am I talking to a bot?"
When a customer asks whether a person is on the other end, they already suspect it. Any automated reply here is bad. Denying it would be a lie. Confirming it in a canned way confirms the worst impression.
This question goes to you, and the answer is the truth, in your own words.
6. Ambiguous messages
"Did you get it?" "Saw that?" "So where are we?" Without clear context, AI guesses, and a guess in customer service becomes a misunderstanding. A person would ask back. A good system detects the ambiguity and hands the chat to you.
7. Sensitive and personal topics
Illness, grief, a customer's money trouble, a fight between their partners. None of that fits a bot flow, and customers remember those conversations for years. A person answers.
8. The first conversation of a complex sale
For simple products, a qualification bot works. For custom services, the first conversation decides whether the lead trusts you. Diagnostic questions, real curiosity and noticing what the lead did not say are hard to automate without sounding scripted.
What is worth automating
To balance things out, here is what usually works well with AI:
- Reading and organizing. Transcribing voice notes, summarizing a group, sorting the queue by who has waited longest. See how to transcribe WhatsApp voice messages.
- Answering what is already decided. Status of work in progress, information that is already in the history, confirming you received something.
- Drafting everything else for you to approve, so the decision stays with you and the typing does not.
- Remembering. Following up with a customer who went quiet, catching a task that slipped.
How to split it in practice
The safest setup is to turn automation on per conversation, not for the whole number. A routine technical support chat can run on autopilot. The chat with your customer's co-founder stays in approval mode, always.
Inside the automated chats, set triggers that hand the conversation back to you. Mentions of price, deadline, discount, cancellation, complaint or "is this AI?" are good candidates. When a trigger fires, the AI does not answer on the merits and lets you know.
In WACrew, an open-source project, every chat starts in Confirm mode, so nothing goes out without your OK. Autonomous mode is switched on chat by chat, and delivery stops on its own if you reply from your phone or the customer writes again before the message goes out. If a customer asks whether it is AI, it does not lie and does not pretend to be you. It stops and calls you.
FAQ
Does automating support make the customer experience worse? It depends on what you automate. Organizing the queue and answering status questions makes it better. Handling complaints or negotiating prices alone usually makes it worse.
How do I know if my automation has gone too far? Read a week of automated replies and ask, for each one, whether you would have said that. If the answer is no more than once or twice, tighten the triggers.
Chatbot or AI, which one is less risky? A menu chatbot is predictable and limited. AI is flexible and can make things up. With human approval on the important chats, AI gets more done. The page on the WhatsApp AI agent compares the approaches.



