AI Chatbots for Business Websites and WhatsApp: How to Scope One

The problem a chatbot is meant to solve

A physiotherapy clinic gets the same eight questions every day. What are your timings? Do you do home visits? How much is the first session? Is there parking? Many of these arrive on WhatsApp after 9pm, when the front desk has gone home. By morning, some of those people have booked somewhere else.

That is the honest case for an AI chatbot. It is not "AI for your business" in general. It is a narrow tool for answering repeat questions instantly, at any hour, and passing the serious enquiries to a person.

Modern chatbots are built on large language models, which means they handle loosely worded questions far better than the old menu-driven bots ("Press 1 for timings"). They also make mistakes in new ways. This guide covers both, and how to scope a bot so the mistakes stay small.

What AI chatbots are good at, and where they fail

The pattern is simple. The bot is reliable when the answer exists in material you gave it and does not depend on the individual asking. It is unreliable the moment it has to fill a gap.

The failure that costs the most

Language models are built to produce a fluent answer. Left unchecked, a bot asked "Do you offer EMI on the 12-month course?" may say yes because it sounds plausible. Now a customer has a screenshot of your business promising something you do not offer.

The fix is a rule, set during build and tested before launch: answer only from approved material, and when the answer is not there, say so and offer a person. Every bot Mecverse builds works this way. When it does not know, it collects a contact instead of guessing.

How to scope a chatbot: three jobs, written down

Most chatbot projects that disappoint were never scoped. Someone said "add a chatbot" and the bot got the whole website as its material. Write a one-page scope around three jobs instead.

Job 1: Answer FAQs from approved material

List the 20 to 30 questions customers actually ask. Your front desk or WhatsApp history is the best source. For each, note the correct answer and where it lives (price list, policy document, services page). That list becomes the bot's material and its test set.

It does not need to be polished. A list in a notebook or in someone's head is a fine starting point, as long as someone writes it down once.

Job 2: Capture leads

Decide exactly what the bot should collect, and when:

• Fields: usually name, phone number and the requirement in one line. A coaching institute might add the student's class. Every extra field loses some people.
• Trigger: when the visitor asks about pricing, booking or anything the bot cannot answer.
• Destination: where the lead goes (email, a WhatsApp group, a CRM) and who owns follow-up.
• Consent: tell people what you are collecting and why. India's Digital Personal Data Protection Act applies to names and numbers collected this way.

Job 3: Hand off to a human

Define the handoff rules before launch:

1. The bot does not have the answer in its material.
2. The customer asks for a person, in any wording.
3. The message is a complaint, refund or anything emotional.
4. The question needs advice or live data the bot cannot see.

Then decide what the customer is told. "Our team will reply by 10am tomorrow" is far better than a vague "someone will get back to you". Make sure that promise matches reality.

Website chat vs WhatsApp: what changes

The same assistant can run on both channels, but the plumbing and rules differ.

On WhatsApp, replies inside the customer service window (24 hours from the customer's last message) can be free-form, which is where a chatbot does its work. Messages your business starts outside that window must use a pre-approved template. Meta's pricing documentation explains the current categories and rates.

Meta business verification is the step that most often slows a WhatsApp launch. Our verification guide covers the documents needed and the common rejection reasons.

What an AI chatbot costs to run

Be wary of anyone who quotes a single flat price without asking how many conversations you expect. The running cost depends on usage, and it breaks into a few parts.

1. Setup and training. A one-time fee to gather your material, build the bot, test it against real questions and connect it to your channels.
2. Model usage. Each reply uses an AI model, which is billed by volume. More conversations, and longer ones, cost more.
3. Hosting or platform. Where the bot runs, plus any chat or WhatsApp platform subscription.
4. WhatsApp message charges. Meta bills per message by category and country. Service replies to customer messages are free; Marketing, Utility and Authentication templates are charged. Our WhatsApp API pricing page breaks down how these costs stack up.
5. Your team's time. Someone answers the handoffs and keeps the material up to date. This is the cost people forget.

This is why Mecverse quotes chatbots as a one-time setup and training fee plus a monthly fee covering hosting and conversation volume, priced against your expected enquiries rather than a published flat rate. To get a useful quote from any provider, bring two numbers: roughly how many enquiries you get a month, and what share arrive on WhatsApp versus the website.

Your next steps: a launch checklist

Before you speak to any provider, or before you build one yourself, get these down on paper.

1. List the 20 to 30 questions customers ask most, with correct answers.
2. Write the "never say" list.
3. Choose lead fields and where leads go.
4. Write the handoff rules and the reply-time promise.
5. Pick the channel: website, WhatsApp or both.
6. Estimate monthly enquiry volume per channel.
7. Test the bot with your FAQ list, your "never say" list and a few deliberately odd questions before it meets a customer.
8. Read transcripts every week for the first month and correct wrong answers.
9. Name one person who updates the material when prices or timings change.

Two numbers tell you if it is working: how many conversations end with a correct answer or a captured lead, and how many wrong answers you find in the weekly review. If the second number is not falling, the material or the rules need work.

If you would rather have it built and tested for you, see how we approach an AI chatbot for website and WhatsApp. We collect your material, show you the answers before launch, and monitor the first weeks.