Field notesCustomer & commerce
WhatsApp customer service with AI: what to automate and what not to.
Customers treat WhatsApp like a conversation with a person, so automation has to be quick, honest about being automated and one message away from a human. Automate lookups such as order status first; keep complaints, payment problems and anything sensitive with people.
veridive6 min read
The first three messages arrive within a minute: “merhaba”, then “siparisim nerde”, then “2 gündür bekliyorum”. No order number, no Turkish characters, and an expectation of a reply before the customer puts the phone down.
That is how people write on WhatsApp: the way they write to a person, because the same app holds their family and friends. Email is a letter and web chat is a help desk, but WhatsApp is a conversation, and automation there has to behave like one: quick, honest that it is automated, and one message away from a human.
The playbook is short. Automate lookups whose answer comes straight from a system, starting with order status. Keep complaints, payment problems and anything sensitive with people. And make the step from assistant to person effortless, with the conversation attached.
Why is WhatsApp different from email and web chat?
Four differences shape the design.
- Expectations are personal. Customers expect a fast reply and a human tone, and they return to the same thread days later expecting it to remember them.
- Messages arrive in bursts. Several short messages, voice notes, a photo of a damaged product. Wait a few seconds for a burst to finish and treat it as one turn, instead of answering each fragment.
- Input is informal. Turkish typed without Turkish characters, abbreviations, English terms inside Turkish sentences.
- The platform has its own rules. When a business may message first, which message templates need approval, what content is allowed. They change, so check the current rules with your provider instead of building on assumptions.
Which requests should be automated first?
The first candidates share one property: the answer is a fact in a system. The model’s job is to understand the request and find the right record; the reply is built from system fields such as status, carrier and tracking link, so it can’t invent a delivery date.
| Request | First route | Why |
|---|---|---|
| Order status | Automate, from order and carrier data | The answer is a system fact |
| Delivery changes | Automate rescheduling after an identity check; address changes go to a person | A new address can redirect goods |
| Store hours and locations | Automate, from the store system | Public, stable data |
| Return steps | Automate the steps and the return code; refund decisions stay with people | Steps are policy; refunds cost money |
| Complaints, damaged goods | Draft a reply for an agent | Needs judgment |
| Payment problems, anything sensitive | A person, flagged | Money, card data, trust |
Agree the list with the service owner, and run each request type as a suggestion to agents before it answers customers directly; when agents rarely change the suggestion, the case for automating it is made. Chatbot or agent assist covers that sequence.
How do you verify who you are talking to?
A WhatsApp number that matches the number on an order is a signal, not proof: phones are shared, and numbers change hands. Match the check to what you are about to reveal:
- Public information such as store hours: no check.
- Order status: a number match plus one more thing the customer knows, such as the order number or the email address on the order.
- Changes that move goods or money: a stronger check, or a person.
Reveal the minimum, such as “your order ending in 47 is with the carrier” rather than the full address. Never ask for card numbers, passwords or full identity numbers in chat, and decide in advance what happens when a customer sends them anyway.
How should the handover to a person work?
A person must be one message away at all times. Recognize the ways customers ask: “temsilci”, “müşteri temsilcisi”, “yetkili”, “agent”, and the same words typed without Turkish characters. Hand over unasked when the customer repeats a question, sounds frustrated or reaches a case the assistant may not handle. The person replies in the same thread and sees everything the assistant saw, so the customer never repeats themselves. Handing a customer from bot to person details what that handover should contain.
Take an illustrative example: a home goods retailer’s flow for “where is my order?”.
- Recognize. “siparisim nerde” is read as an order-status request. The assistant says it is automated and asks for the order number.
- Verify. The WhatsApp number matches the order, and the order number confirms it.
- Look up. The carrier shows the parcel at a regional hub, past its promised delivery date.
- Answer with facts. The reply states the status from carrier data and apologizes once, without promising a new date the system doesn’t have.
- Hand over. A late parcel is a trigger, so the case goes to an agent with the thread, the verified order, the carrier events, what the assistant said and why it handed over. The customer reads: “Bir çalışanımız bu konuşmaya buradan devam edecek.” (A colleague will continue this conversation from here.)
The agent opens a query with the carrier and replies in the same thread, starting from what the customer already knows.
Customers forgive a bot that says it is one. They don’t forgive one they can’t leave.
What about personal data in chat history?
Chat threads collect more than you asked for: names and addresses, product photos with faces in the background, identity cards sent unprompted, voice notes. Take these questions to your data protection officer before launch:
- What is the legal basis for processing chat content with a model, and where is it processed?
- Does sending it to a model provider count as a transfer abroad under KVKK or GDPR?
- How long are chat histories, transcripts and model logs kept, and who can read them?
- What happens to sensitive data a customer sends without being asked?
- How are access and deletion requests handled across the platform, your CRM and the logs?
Tell customers at the start that they are talking to an automated assistant, and how to reach a person. It is the honest default; any wording your markets require is a question for counsel.
How do you keep the tone right in Turkish and English?
Write for a phone screen: short messages, one question at a time, no letter-style openings. In Turkish, choose “siz” or “sen” once and keep it for the whole conversation; “siz” is the safe default where the brand hasn’t decided. Understand Turkish typed without Turkish characters, but always reply with correct spelling. Answer in the customer’s language, and follow if they switch. The opening line can do two jobs at once: “Merhaba, ben otomatik asistanım. İstediğiniz an ‘temsilci’ yazarak bir çalışanımıza ulaşabilirsiniz.” (Hello, I’m an automated assistant. Type ‘temsilci’ at any time to reach a colleague.) Write these rules down in a tone of voice guide, with examples in both languages.
Tag real conversations
Export a sample of WhatsApp conversations and tag them by request type. The lookups, usually led by order status, are your automation candidates; the rest show where agents need help. Start with suggestions to agents on those lookups, measure how often agents change them, and automate only what passes. That is the shape of a first project in customer operations, and for retailers selling through their own shop and marketplaces it connects to commerce intelligence.
This note is general information, not legal advice.
Ask an assistant about this note