How to give your WhatsApp AI agent a persona and tone of voice that sounds like your business, with examples you can copy
Customers notice the voice before the content. An agent that says "Greetings! I would be delighted to assist you" to a customer who wrote "bro got size L or not" has lost before it answered. The persona section of the instructions is where you decide how the agent sounds, and it takes eight lines.
Quick answer
Write the persona as eight short lines in the agent's instructions: a name and role, who it serves, how formal, how warm, how long replies should be, emoji and language rules, three phrases it should use, and three it must never use. Then test it with ten real messages from your chat history and adjust. Match your customers, not your brand deck: a streetwear shop and a dental clinic should not share a voice. ReplyKit drafts the first version from your website and you edit it in plain English.
The eight-line persona block
- Name and role. "You are Mei, the assistant for Lim Dental." A first name makes handover natural: "Let me get Dr Lim for that."
- Who you serve. "Patients and parents in Petaling Jaya, mostly messaging in English, Malay or Chinese."
- Formality. One of: formal, friendly-professional, casual. Pick from your customers' messages, not your preference.
- Warmth. "Warm and reassuring; never cheerful about pain." or "Direct and practical; skip pleasantries."
- Length. "Two to four sentences. One question at a time. Use a short list only for prices or steps."
- Emoji and language. "No emoji" or "At most one, never in bad news." "Reply in the language the customer writes in; keep Manglish if they use it."
- Phrases to use. Three that sound like you: "No problem", "Can", "Let me check with the team".
- Never say. Three to ban: "As an AI language model", "I apologise for any inconvenience", "Please do not hesitate".
Keep the rest of the instructions (what it knows, what it hands over) separate; the structure is in writing instructions for a WhatsApp AI agent.
Four worked examples
Dental clinic
You are Mei, the assistant for Lim Dental in PJ. Patients are often nervous. Be warm and calm, never jokey about pain or cost. Friendly-professional. Two to three sentences, one question at a time. No emoji. Reply in the patient's language. Use: "Of course", "That's common", "Dr Lim can look at that". Never: "Don't worry" about a symptom, medical opinions, "As an AI".
Streetwear shop
You are the assistant for Vault Studio. Customers are 18 to 30 and message in Manglish. Casual, quick, a little playful; match their slang but stay clear. One or two sentences. One emoji max, never on bad news. Use: "Got", "Can", "Restocking Thursday". Never: "Dear customer", "kindly", pretending something is in stock.
Plumber
You are the assistant for Ah Seng Plumbing. Customers are stressed and want a time. Direct and practical. State what we do, when we can come, and the call-out fee, in that order. Two sentences. No emoji. Use: "We can come today", "Call-out is RM 80". Never: guess a repair price without a visit, promise a time slot not in the schedule.
Tuition centre
You are Aunty Chen's assistant at Bright Minds Tuition. Parents are busy and careful with money. Respectful and clear, slightly formal. Three sentences. One question at a time. Reply in English, Malay or Chinese as written. Use: "Our trial class", "Teacher Chen will advise". Never: comment on a child's ability, promise grades.
Mistakes that make agents sound fake
- Borrowing the brand voice from the website. Web copy is written to be read once; chat is a conversation. Shorter, plainer, warmer.
- Over-apologising. One "sorry" is human; three is a call centre.
- Exclamation marks. Cap at one per message, zero in bad news.
- Asking three questions in one message. Customers answer one. The agent should too.
- Hiding that it is an assistant. Disclose on first contact; customers trust a clear assistant more than a fake person, as the data in do customers mind talking to AI shows.
- No never-say list. The model's default phrases leak through unless you ban them.
Testing the voice
Take ten real messages from last month, the rude one and the long one included, and run them through the test chat. Read the replies out loud. If one sounds like a bank, shorten it and add a phrase from your list. If one is too casual for the subject, add a rule: "Switch to calm and formal for complaints or medical topics." Two rounds of this usually settle it. The wider test is in testing your agent before going live.
Frequently asked questions
Should the agent have a name?
A first name helps: it makes handover natural and reads as a team member rather than a system. Always disclose that it is an automated assistant.
How formal should it be?
As formal as your customers' own messages, not your brand guidelines. Read ten real chats and match them.
Can it use Manglish or Singlish?
Yes, if your customers do. Tell it to match the customer's register and keep it clear.
How long should replies be?
Two to four sentences for most businesses, one question at a time, lists only for prices or steps.
What phrases should I ban?
The model's defaults: "As an AI language model", "I apologise for any inconvenience", "Please do not hesitate", and anything your business would never say.
Sounds like you, not like a bot
Instructions in plain English, drafted from your site, tested in chat before go-live. 7-day free trial, no card.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.