AI Agent on WhatsApp: What It Resolves on Its Own and When to Hand Off to a Person

An AI attendant on WhatsApp works when it's given a small, well-defined job, and fails when it's told to "handle the customer." It answers well what's written down, collects well what's missing, and hands off well to a person when the topic goes off script. But it makes things up, promises what it can't, and irritates people when it's forced to decide what it doesn't know.

Almost every bad automated service experience comes from a design mistake, not a limitation of the technology: nobody said what the AI does, what it never does, and how it hands the conversation over. This article walks through those three decisions.

What to delegate to the AI

The golden rule: delegate what has a right answer you already know and can write down. Five types of tasks fit well.

Frequently asked questions

Business hours, address, payment methods, delivery time, how to get a duplicate invoice. Any question your team answers off the top of their heads, and always the same way, is a candidate. If you're pasting the same reply for the fifth time this week, the topic is ready for the AI.

Triage

Before a person takes over, someone has to find out who the customer is, what the problem is, and what's already been tried. The AI asks for the order number, the screen where the error appears, the registered email. By the time the human agent steps in, the conversation already has what they would have asked in the first three minutes.

After-hours service

Late nights, holidays, and the Monday morning peak are the moments the human team can't cover. The AI greets the customer, resolves what it can, and leaves the rest logged for whoever gets there first. The customer doesn't have to wait for business hours to know they were heard.

Opening a ticket

Many requests need a record: a technical problem, a complaint, a service request. The AI guides the conversation until it has the minimum information and opens the record with everything attached, instead of depending on someone rereading the conversation and transcribing it.

Follow-up

Conversations go cold. The customer said "let me check and get back to you" and disappeared. An automatic follow-up that reaches back out to people who stopped replying recovers conversations that would be lost to distraction, not lack of interest.

What not to delegate

The list below matters more than the previous one, because this is where the damage is.

• Emotional complaints. An irritated customer needs to feel there's a person on the other end. An empathetic automated reply usually makes things worse.
• Exceptions on price, deadline, or policy. Discounts, credits, and out-of-policy refunds are decisions for whoever answers for the outcome. The AI shouldn't promise anything that depends on authorization.
• Diagnosing a new problem. If the problem isn't in the knowledge base, the AI has no way of knowing the cause. Guessing in public creates rework and loss of trust.
• Sensitive legal, health, or financial matters. Any answer that could create an obligation or risk for the customer goes through a person.
• A customer who asks for a human. The request is an order. Negotiating ("can I try to help first?") more than once is the shortest path to a complaint.

Training: the AI is only as good as what you write

An AI attendant uses your business's knowledge, not general knowledge. That's why the real work is writing a knowledge base that answers what customers ask, in the tone your company would use.

Start with real conversations

Open the last 50 WhatsApp conversations and list the questions. Group them by topic. The five or six topics that account for most requests are your initial scope. The rest stays out, and that's a conscious decision, not a gap.

Write the answer the way your best agent would

Short answers, with the concrete detail up front. Compare:

Weak base	Good base
"We work during business hours."	"We're open Monday to Friday, 8 a.m. to 6 p.m., and Saturdays from 8 a.m. to noon. Outside those hours, leave your message here and we'll reply on the next business day."
"See our exchange policy."	"You have 7 calendar days after receiving your order to request an exchange. Tell me the order number and I'll open the request right now."

The values in the example are illustrative: use your own company's. What matters is that the base answers and routes, instead of sending the customer to look elsewhere.

Say what the AI doesn't know

Write refusal instructions: "If they ask about a discount, say a consultant will get in touch and hand the conversation over." The clearer the limit, the less the AI improvises. A good instruction also covers tone: formal or friendly, whether it uses the customer's name, whether it uses emoji.

Review, correct, and repeat

In the first few weeks, read the conversations the AI handled. Every bad answer is a flaw in the base or in the instructions, and the fix applies to every conversation that follows. It's a few minutes of work a day that pays for itself quickly.

The handoff to a human: the moment that decides everything

No customer complains because they talked to an AI. They complain when they realize they got stuck in it. Design the handoff as the main part of the service, not as an exception.

When to hand off

• The customer asks for a person, in any words.
• The topic is on the do-not-delegate list.
• The AI found no answer in the base after one or two attempts.
• The tone of the conversation signals irritation or urgency.
• The customer repeats the same question, a sign the answer didn't help.

How to hand off

Three precautions make the difference between a good handoff and a terrible one.

1. To the right team. Billing goes to billing, technical problems go to support. Sending everything to a general queue only pushes the problem down the road.
2. With the whole conversation. The agent needs to read what the customer already said. No "can you repeat the problem?"
3. With a clear expectation. The AI says it will pass the conversation to a person and gives a realistic response time. That time must actually exist: it's the first-response SLA, and it has to be tracked.

After the handoff

If the conversation needs work from another department, it should become a ticket or a task, with the history attached, rather than staying stuck in WhatsApp. The path is the same one described in how to organize WhatsApp support without it turning into a mess.

Follow-up without becoming spam

Good follow-up is rare and brief. One message to someone who stopped replying, after a reasonable interval, with a direct question ("Were you able to try the solution I sent?"). Two messages are the limit. Three is stalking.

Respect context: someone who asked not to be contacted gets no message. Also respect WhatsApp's rules for messages outside the conversation window, which on Meta's official channel use approved templates.

How to know if it's working

Four indicators are enough to evaluate an AI attendant in the first few weeks.

• Resolution rate without a human. How many conversations the AI closed without help, and whether the customer came back to ask the same thing.
• Handoffs by topic. If a topic almost always goes to a human, either the base is incomplete or it isn't a job for the AI.
• Satisfaction. The survey sent after a conversation closes, split between conversations handled only by the AI and conversations with a human.
• Time to first response. It should drop, especially outside business hours.

If satisfaction on AI-only conversations is far below human conversations, narrow the AI's scope. Handling fewer topics and handling them well beats handling everything and irritating people.

Where to start

Pick one topic, the most repetitive one in your service. Write its base, define the handoff rule, and let the AI handle only that topic for two weeks, with you reading the conversations. Then add the second topic. Teams that start this way reach reliable automated service in a few months; those who switch everything on at once spend the same time putting out fires.