By KYN AI Advisory Team — AI implementation specialists, Singapore
A WhatsApp reply desk reads each inbound message, labels what it is, routes it to the right person and writes the result to the CRM. The agent sorts and drafts. A named person sends anything that carries your name. This article describes how we design one: the labels, the routing rules, the handoff, the stop conditions and the checks that tell you it works.
It is a design pattern from our reply desk work, which sorts replies for an insurance brokerage at about S$0.01 per reply, and from WhatsApp-based reporting we run for a manufacturer. We state no accuracy figure for the WhatsApp version, since we have not published one.
What problem does a reply desk solve?
Enquiries arrive on WhatsApp at all hours, to personal phones, to a shared number, to whoever happens to be holding the device. Three things go wrong.
- Slow first response. The message sits until someone sees it.
- No owner. Two people reply, or nobody does.
- No record. The conversation lives on a phone. When the salesperson leaves, so does the history.
A reply desk fixes all three: every message lands in one queue, has one owner and is logged.
For the channel choice behind this, see WhatsApp Business app vs platform.
What are the steps?
- Receive. Messages arrive through the WhatsApp Business Platform into your system.
- Match. The system finds the contact by phone number. A known contact brings their history. An unknown number creates a new record.
- Read and label. The agent assigns one label.
- Extract. It pulls out structured details: product interest, quantity, location, timing.
- Route. Rules pick the owner.
- Draft. For common questions, it drafts a reply from approved wording.
- Review. The owner approves, edits or replaces the draft.
- Log. Everything is written to the record with actor and time.
Steps 3 to 6 are the agent. Steps 1, 2 and 8 are plain software. Step 7 is a person.
Which labels work?
Keep the list short, with clear meanings. Ten labels with fuzzy edges are worse than six sharp ones.
| Label | Meaning | Default action | |---|---|---| | New enquiry | A prospect asks about a product or service | Route to sales, draft reply | | Existing customer | Known customer, ongoing matter | Route to account owner | | Quote request | Asks for price, terms or a proposal | Route to sales, no draft of price | | Support | Fault, delivery or usage problem | Route to service | | Complaint | Dissatisfaction, refund or dispute | Stop, route to a named person | | Not relevant | Spam, wrong number, vendor pitch | Archive, log | | Unsubscribe | Asks to stop messages | Suppress in every channel |
The labels that matter most are the two that stop automation. A complaint should never receive a cheerful automated draft. An unsubscribe must take effect at once.
How does routing work?
Routing is rules, not model judgement. The model labels; the rules decide.
- By contact. A known customer goes to their account owner.
- By product or topic. Product A goes to team A.
- By territory or language. Indonesian goes to the Jakarta team; Chinese to Hong Kong.
- By time. Outside working hours the message is acknowledged by an approved template where your policy allows it, and queued for the next morning.
- By load. If the owner is away, the backup receives it.
Write the rules in a table that a manager can read and change. If the routing logic lives only in code, the business cannot adjust it.
See managing WhatsApp enquiries across a sales team for the team side of this.
What can the agent draft, and what can it not?
It can draft:
- a greeting that names the product the customer asked about;
- answers to frequent questions from a set of approved answers, such as opening hours, service areas and how the process works;
- a request for missing details, such as quantity or delivery address;
- a summary for the owner.
It cannot:
- quote a price outside an approved list;
- promise a date or a delivery;
- agree terms or discounts;
- reply to a complaint;
- send anything without approval to a contact who has not messaged first.
The approved answers are written by you. The agent chooses among them and fills in fields. It does not invent policy.
What stops the desk?
Write stop conditions as rules.
- Mentions of legal action, a regulator, injury or safety.
- A refund or chargeback request.
- A message in a language outside those you support.
- A voice note or image the system cannot read reliably.
- Low confidence on the label.
When a stop condition fires, the message goes to a named person with a one-line reason. The customer is not left silent: a person replies, and a short acknowledgement can go out if your policy allows it.
Hostile text is also possible. A message can include instructions aimed at the agent. The agent's account has no power to send, pay or delete, and its output is validated against a fixed structure. See prompt injection in business AI agents.
How do you respect consent and marketing rules?
Replying to a customer who messaged you is different from messaging someone cold. Keep the two flows apart.
- Inbound reply: the customer opened the conversation.
- Outbound first message: a person approves it, after checking the suppression list and the rules for that country.
In Singapore, marketing messages to numbers are governed by the PDPC's Do Not Call rules. Check the rule for the recipient's country in Hong Kong, Indonesia and Malaysia. For personal data handling more broadly, see PDPA for AI systems.
How do you know it works?
Measure what a person would measure.
- Label accuracy. Have a person label 100 real messages, then compare. Look hardest at the costly errors: a complaint labelled as an enquiry, an unsubscribe missed.
- First response time. Before and after, from your own timestamps.
- Unowned messages. Should be zero at the end of each day.
- Edit rate on drafts. High edits on one label mean the approved answers need work.
- Review time per 100 messages.
What does it cost?
Our reply sorting for the insurance brokerage costs about S$0.01 per reply, which is a figure for email replies on that system. A WhatsApp message is a similar size of task, so we expect a similar order of cost, but we have not measured it. WhatsApp's own messaging charges are separate. See WhatsApp Business pricing from 1 October 2026 for the platform side.
Where KYN fits
KYN builds reply desks as part of a sales system in the client's own cloud account. For an insurance brokerage we built outreach with one checkable first line and a reply desk that sorts every reply. For a manufacturer we send a morning report to each department on WhatsApp at about S$0.05 per report.
Agents have their own restricted accounts, every action is logged with the actor, and outbound messages, money, filings and deletions wait for a named person. The first 14 days are a review of your stack and interviews with your team; a full build ships in under 90 days. See sales and integrations.
Frequently asked questions
Can an AI reply to WhatsApp messages automatically?
For narrow, approved answers it can, if you set that policy. We keep a person in the loop for anything about price, terms, dates, complaints and first messages to new contacts. The agent drafts; the person sends.
Do I need the WhatsApp Business Platform?
For routing, logging and several users on one number, yes. The Business app suits one or two people on a phone. The platform connects to your system. Our comparison article sets out the difference.
What happens when the agent is unsure?
It sends the message to a named person with the reason. Low confidence is a stop condition, and it should be treated as normal, not as a failure.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
