WhatsApp chatbot for business: rule-based flows, AI agents, and what each costs

Mohit Bansal runs two optical stores in Indore with eight staff and one WhatsApp number printed on every receipt. Most days the number gets the same four messages, over and over: what time is the eye test, how much are progressive lenses, are my glasses ready, do you take this insurance card. His staff answer between customers, so a question sent at 1 pm gets a reply at 4.

He asked a vendor for a WhatsApp chatbot and got two demos that looked nothing alike. One showed a menu of buttons. The other answered a typed question about lens coatings in full sentences. Both were called chatbots. Both were right, and the choice between them matters more than which vendor he picks.

Two kinds of WhatsApp bot

A rule-based bot follows a path you draw in advance. The customer writes in, the bot sends a menu, the customer taps “Eye test timings”, the bot sends the timings and offers the next set of buttons. Nothing is written on the fly. If the customer types something the path doesn’t expect, the bot either repeats the menu or passes the chat to a person.

An AI chatbot, usually called an AI agent now, reads what the customer typed and writes an answer from material you give it: your website, your price list, your policies. It can handle “do u have blue light glasses for kids under 2000” without a button for it. It can also get things wrong, which is the trade you’re making.

WhatsApp gives both kinds the same building blocks. Meta’s interactive messages allow up to three reply buttons, each with a label of at most 20 characters, or a list of up to ten rows, each with a title of at most 24 characters. For longer forms, such as booking an appointment with a date and a branch, WhatsApp has Flows: form screens inside the chat. Our glossary covers both terms.

What button flows are good at

Button flows are cheap to run, predictable and easy to check. You know exactly what the bot will say, because you wrote it. That makes them right for:

• Sorting a chat to the right person: sales, service, accounts.
• Collecting a few details in order: name, branch, preferred day.
• Answering a small set of questions with fixed answers, such as timings and addresses.
• Confirming or cancelling a booking with one tap.

SJ Sakshi Jain talking to Bansal Opticians

1. Sakshi Jain:
   hi 1:04 pm
2. Bansal Opticians:
   Hi Sakshi, welcome to Bansal Opticians. What can we help with? 1:04 pm
   • Book an eye test
   • Is my order ready?
   • Talk to the store
3. Sakshi Jain:
   Book an eye test 1:05 pm
4. Bansal Opticians:
   Which store suits you? 1:05 pm
   • Vijay Nagar
   • Rajwada
5. Sakshi Jain:
   Vijay Nagar 1:05 pm
6. Tagged: eye test · Vijay Nagar · sent to the Vijay Nagar front desk

Where they break is free text. Indian customers type in English, Hinglish and everything between, with spelling that no keyword list predicts. A flow that only understands taps will push those customers back to the menu, and some of them give up. Keep menus to three or four choices and always include a “talk to a person” button.

What AI agents are good at

An AI agent earns its keep when customers ask open questions and the answers already exist somewhere: on your website, in a rate card, in a returns policy. It removes the “please select from the options below” loop, and it answers at 11 pm when the store is shut.

It also needs more care than a menu.

• It’s only as current as its material. If the price list on your website is from March, so are its answers.
• It shouldn’t guess. A good agent says it doesn’t know and passes the chat on. Test that it does.
• It writes, so it can write too much. Four messages where one would do costs you money, as the next section shows.

Some agents can also take actions, such as booking a slot in your calendar or checking an order in your system. Many can’t, and the ones that can need connecting to your software. Ask the vendor to show it working on your data, not on a demo store.

Which one fits

If most of your chats are…	Start with
The same three or four requests, with fixed answers	Button flows
Routing to the right team or branch	Button flows
Typed questions about products, prices or policies	An AI agent
A mix of both	Buttons first, with an AI agent answering what the buttons don’t cover

Our view: most small businesses should start with buttons for the top two or three requests and add AI once they can see which typed questions keep coming. You’ll lose a little on the late-night questions in the first month. You’ll also know exactly what your bot says while you learn what customers ask.

The handover is the part customers remember

Whichever kind you choose, the bot will meet chats it can’t handle: a complaint, a bulk order, a customer who just wants a person. What happens next decides whether the bot helped.

Settle these before you switch anything on:

1. Which phrases or buttons send the chat to a person, and which person or team gets it.
2. Whether that person sees what the customer already said, so the customer doesn’t repeat it.
3. What the bot tells the customer while they wait, especially outside working hours.
4. How a person turns the bot off for one chat, so it doesn’t interrupt them.

A bot with no way out is worse than no bot. Customers who get the same menu three times stop messaging and start calling.

The 24-hour window, and what each reply costs

Bots work inside WhatsApp’s customer service window: the 24 hours after a customer’s last message. Inside it, the bot can send any reply, including buttons and lists. Once it closes, neither the bot nor your staff can send a free-form message. You need an approved template, billed by category; our template examples cover those.

Replies inside the window are no longer free. Since 1 October 2026, Meta charges for service messages after the first 1,000 per business phone number per month. For Indian numbers the rate is ₹0.115 per delivered message, before GST, the same as a utility template. A bot’s reply counts exactly like a person’s. Our news brief on the change has the detail.

To see how that adds up, take an example. Say Mohit’s number handles 1,200 chats a month. The button flow sends about four messages per chat (greeting with menu, answer, follow-up menu, handover note), and staff send two more on half the chats:

• Bot: 1,200 × 4 = 4,800 messages
• Staff: 600 × 2 = 1,200 messages
• Total: 6,000, of which 1,000 are free
• Meta’s charge: 5,000 × ₹0.115 = ₹575 a month, before GST

That’s modest. It doubles if the bot sends eight short messages where four would do, so count the messages in a flow before you publish it, and prefer one message with buttons over two plain ones.

Two things stay free. Reactions, such as a 👍 on a customer’s message, are free and don’t use the 1,000. And when a customer arrives from a Click-to-WhatsApp ad on their phone and you reply within the window, Meta opens a free entry point window that can stay open for up to 7 days, during which messages are free. See how the ad window works.

Meta’s own Business AI

Since May 2026, Indian small businesses can switch on Meta’s Business AI inside the WhatsApp Business app. For a business run from one phone, it may be all you need. We compared it with a Platform agent in Meta’s Business AI vs an AI agent, so we won’t repeat it here. The short version: if one or two people answer every chat, try Meta’s option first.

How to build a WhatsApp chatbot without code

Most providers now let you build either kind without a developer. The steps are much the same everywhere.

1. Export or scroll through last month’s chats and write down the 20 questions that come up most, in the customers’ own words.
2. Split them: fixed answers go into a button flow, open questions go to the AI agent, and anything involving money or a complaint goes to a person.
3. For an AI agent, gather the material it should answer from: website pages, price lists, policies. Fix anything out of date first.
4. Draw the flow: greeting, menu, answers, and a “talk to a person” option at every step.
5. Set the handover: who gets which chats, during which hours, and what the customer sees meanwhile.
6. Test with the real questions from step 1, including the misspelled and Hinglish ones. Fix every wrong answer before a customer sees it.
7. Switch it on for part of the day first, such as after store hours, and read every bot conversation for the first week.

What a chatbot costs, line by line

“Free WhatsApp chatbot” usually means a provider’s free tier. Meta’s charges still apply on the Platform. Expect three lines on the bill:

Cost	Who charges it	How it’s counted
Messages	Meta	Per delivered message: ₹0.115 for service replies after 1,000 per number per month; templates by category
Software	Your provider	Monthly plan: seats, numbers, contacts
AI usage	Your provider (or Meta, for Meta’s agent)	Credits or tokens per answer

The third line is the one to question. Ask how usage is counted, how many credits a typical answer uses, what happens when credits run out mid-month, and whether testing uses them too. If the provider runs Meta’s Business Agent for you, Meta itself has charged $2 per million tokens for it since 1 August 2026, which Meta puts at roughly 4 to 5 US cents a message.

Where to start this week

Take your last 100 WhatsApp chats and sort them into three piles: same question, same answer; open question with a known answer; needs a person. The biggest pile tells you which bot to build first. If the third pile is the biggest, you don’t need a bot yet. You need a shared inbox so those chats reach the right person faster.