Customer support on WhatsApp: routing, response times, AI answers and calls

Manpreet Sidhu looks after service at a water-purifier dealer in Chandigarh: eleven staff, six of them technicians on scooters, and one WhatsApp number on the sticker inside every purifier’s cover. Customers use it for everything. “Filter light is red.” A photo of a leaking joint. “Technician didn’t come.” A voice note in Punjabi at 10.40 pm. About 90 conversations a day reach two people at the office, who also answer the landline.

WhatsApp customer service is harder than it looks because the app feels informal and the work isn’t. Each chat is a ticket with a clock on it. This guide covers the rules Meta sets, then the decisions that are yours: who gets which chat, how fast, what a machine answers, and when to pick up a call instead.

It assumes your number is on the WhatsApp Business Platform with a shared inbox. If one or two people handle support from a single phone, the free Business app may be all you need, and app vs Platform will tell you when that stops being true.

The 24-hour window sets your real deadline

When a customer messages you, a 24-hour customer service window opens. It resets every time they write again. Inside it, your team can send ordinary free-form replies. Once it closes, you can’t. The only thing you can send is a template Meta has approved in advance.

So the latest possible reply isn’t a matter of policy. A complaint that arrives at 10 am on Saturday and is opened on Monday morning can no longer be answered with a typed message. You’d have to send a template along the lines of “We’re following up on your service request, reply here to continue”, wait for the customer to respond, and only then say what you meant to say.

Three habits follow from this:

• Clear the queue before any day off. Chats where the customer spoke last are the ones at risk.
• Get one “following up on your request” utility template approved now, before you need it. Meta decides the category, and anything promotional in the wording makes it marketing.
• Don’t ask a customer to “message again later”. Each new message from them resets the window, but relying on that is asking them to do your follow-up.

The glossary has the definitions if window, template and service message are new terms.

Replies now cost money after the first 1,000

Until the end of September 2026, free-form replies inside the window were free. From 1 October 2026, Meta charges for them. Each business phone number gets 1,000 free service messages a month, which reset monthly and don’t roll over. From the 1,001st, each delivered reply to an Indian number costs ₹0.115, before GST. Reactions, such as a 👍 on a customer’s message, stay free and don’t count toward the 1,000.

A reply from an automated flow or an AI agent counts the same as one typed by a person. The follow-up template you send after the window closes is billed as a template in its category, and a utility template to an Indian number is also ₹0.115.

For most support teams the bill is small. What it changes is the habit of splitting one answer into “Hi”, “one sec”, “checking” and then the answer. That’s four charged messages and three notifications for the customer. One complete reply is cheaper and reads better. Our news brief on the change has the detail.

Route by what the customer needs

The usual failure in a shared number is that everyone sees everything, so nobody owns anything. Routing fixes that, and it has two parts.

Sort the chat. Decide the three or four kinds of request you get and send each to a team. For Manpreet’s dealer: new purchase enquiries to sales, service and complaints to the service desk, payments and AMC renewals to accounts. A first message with reply buttons can do the sorting. WhatsApp allows up to three buttons on a message, so three teams fit neatly. If the customer types instead of tapping, a person reads the chat and assigns it.

Give it an owner. Within the team, one named person takes the chat and keeps it until it’s resolved. A returning customer goes back to whoever helped last, where that’s possible. Nothing makes a customer angrier than describing the same leak to a third person.

Two rules are worth writing down:

1. Every new chat is assigned within 15 minutes during working hours.
2. The unassigned list is empty when the office closes.

The step-by-step version, with teams, views and handover notes, is in our shared inbox setup playbook.

Response times you can keep

Customers on WhatsApp expect chat speed, and no small team can give it all day. What works is a target you publish to yourselves and actually meet, plus honesty with the customer when you can’t.

Our suggestion for a team of two to five:

Situation	Target
First reply, working hours	Within 15 minutes
First reply, outside hours	By 10 am the next working day, with an automatic note saying so
An update on an open complaint	Every working day until it’s closed, even if the update is “part ordered, arriving Thursday”
Any chat	Never past 24 hours with the customer waiting

The first reply doesn’t have to solve anything. “Got it, Mr Bhalla. I’m checking with the technician and will message you by 3 pm” buys you time, as long as someone messages by 3 pm.

Set the after-hours note to go once per conversation, not once per message. It’s a service message like any other, so it counts toward the 1,000.

There’s a trade-off here. A 15-minute target with two people means someone is watching the inbox all day and doing less of everything else. If that isn’t affordable, set 30 minutes and meet it. A slower promise kept beats a faster one missed.

Let a WhatsApp support bot take the repeat questions

Read a week of support chats and a pattern shows up. For a purifier dealer, a large share will be “what does the red light mean”, “how much is the AMC”, “when is my service due” and “which filter does my model take”. The answers exist already, in a manual or a rate card.

That’s the work for a WhatsApp support bot, which today means one of two things. A button menu handles the handful of requests with fixed answers. An AI agent reads what the customer typed and answers from material you give it, such as your website and your policy documents, which suits questions that arrive in free text and mixed Hindi and English. We compared the two in WhatsApp chatbot for business, and Meta’s Business AI vs an AI agent covers Meta’s own option.

For support specifically, four rules matter more than which kind you pick.

Keep it away from complaints. An agent can explain the red light. It shouldn’t be the one replying to “this is the third time your technician hasn’t come”. Send anything angry, anything about a refund and anything the customer has asked twice to a person.

Make the handoff carry the story. When the chat passes to a person, they should see what the customer asked and what the bot said, ideally as a short summary. If the customer has to start again, the bot made things worse.

Tell the customer what happens next. “I’m passing this to our service team, they reply between 9.30 and 6.30” is enough.

Read its conversations. Every one, for the first two weeks. An AI agent is only as current as the documents behind it, and a rate card from last year produces confident wrong answers.

RB Rohini Bhalla talking to AquaSure Service, Chandigarh

1. Rohini Bhalla:
   red light is blinking on purifier since morning 10:42 pm
2. AquaSure Service, Chandigarh:
   A blinking red light usually means the filter is due for a change. Your water is still safe to use for a few days. Would you like a technician visit? 10:42 pm
   • Book a visit
   • Talk to the team
3. Rohini Bhalla:
   it was changed last month only. this is second time 10:44 pm
4. AquaSure Service, Chandigarh:
   I'm passing this to our service team. They reply from 9.30 am and will have this conversation in front of them. 10:44 pm
5. Handed to Service desk with a summary

Some problems need a call

A customer describing a noise, an elderly parent who doesn’t type, a complaint that has gone three rounds in text: these go faster by voice. On the Platform, customers can call your WhatsApp number from the app on their phone, and with the right software your team answers in a browser.

The rules that matter for support, from Meta’s calling documentation:

• Calls a customer makes to you aren’t charged by Meta. Calls you make to customers are.
• You can’t ring a customer without their permission, which you request in the chat.
• A customer’s call opens or refreshes the 24-hour window, answered or not, so a missed call can be followed up with a typed message.

Decide who answers before you switch calling on. A call that rings on an empty desk is worse than no call button. WhatsApp Calling for Indian businesses covers requirements, permissions and billing.

What to measure

Four numbers, checked every Monday. Most inboxes let you get them by filtering and counting, and ten minutes by hand is fine.

Number	What you want	What it tells you when it’s wrong
Time to first reply, working hours	Inside your target for most chats	Too few people at the peak hour, or chats waiting unassigned
Chats where the window closed with the customer waiting	Zero	Weekend and holiday cover is missing
Chats reopened within a week of being resolved	Falling	Problems are being closed before they’re fixed
Service messages sent per number this month	Known, against the 1,000	Replies are being split, or a bot is too chatty

If you run an AI agent, add a fifth: of the chats it handled, how many were handed to a person, and why. A very low figure can mean it’s good. It can also mean it’s holding on to chats it should pass on.

A small staffing example

These numbers are illustrative, built for Manpreet’s dealer. Swap in your own.

• 90 conversations a day, 26 working days: 2,340 a month.
• Suppose 40% are repeat questions an AI agent or button menu closes. That leaves 54 a day for people.
• Suppose each of those takes about 6 minutes of a person’s attention across its messages: 324 minutes.
• Add 15 customer calls a day at 5 minutes each: 75 minutes.
• Total: 399 minutes, about 6 hours 40 minutes of handling a day.

One person gets perhaps six useful hours from an eight-hour day, and support doesn’t arrive evenly. If a third of the day’s chats come between 10 am and 1 pm, one person can’t hold a 15-minute first reply through that stretch and take calls. So the answer is two people on support during the day, with one covering Saturday. Without the bot, the same sum gives 90 × 6 + 75 = 615 minutes, and two people would be stretched every day.

The message bill for the same month, assuming five business messages per conversation including the bot’s:

• 2,340 × 5 = 11,700 service messages
• Less 1,000 free = 10,700 charged
• 10,700 × ₹0.115 = about ₹1,230, before GST

Staffing is the cost that matters. The message charge is roughly what one technician visit earns.

Where to begin

Pull last week’s chats and mark each one: answered inside 15 minutes or not, and repeat question or not. If fewer than half got a fast first reply, fix routing and cover before you add a bot. If more than a third are the same five questions, write those five answers down properly. They’re your quick replies today and your AI agent’s source material next month.