WhatsApp chatbot for business: how the API works and what it costs

What a WhatsApp chatbot actually is

A WhatsApp chatbot is a program connected to a business phone number that reads incoming messages and answers automatically. It can be a simple menu ("reply 1 for hours, 2 for prices"), a guided flow that collects data for a booking, or an assistant that uses a language model to understand free text and answer from your own information.

In countries where WhatsApp is the default way to talk to businesses, such as Mexico, Brazil, India or much of Europe, a bot is often the difference between answering a customer in seconds and answering hours later. In the US it matters most for businesses that serve those markets or customers who prefer WhatsApp over SMS.

WhatsApp Business app vs the WhatsApp Business Platform (API)

There are two different products with similar names, and choosing the wrong one is the most common mistake I see.

The WhatsApp Business app is free and runs on a phone. It has quick replies, labels and an away message, and it is perfect for a small business that answers by hand. It is not built for real automation or for several people answering from different computers.

The WhatsApp Business Platform, also called the Cloud API, is what serious chatbots use. Your messages arrive at a server through a webhook, your code decides what to answer, and several agents can work from a shared inbox. You can connect it to your booking system, CRM, inventory or an AI model.

How WhatsApp API pricing works

Meta's official pricing documentation explains the model. Effective July 1, 2025, Meta charges on a per-message basis for template messages: messages your business starts using a pre-approved template, grouped into marketing, utility and authentication categories.

Marketing templates are always charged. Utility and authentication templates are charged when they are sent outside an open customer service window. When a customer writes to you, a 24-hour customer service window opens; inside it, all non-template messages are free, and utility templates delivered within the open window are free too.

There is also a free entry point: when a user starts the conversation from a Click-to-WhatsApp ad or a Facebook Page call to action, a 72-hour window opens in which you can send any type of message at no charge.

Actual prices depend on the recipient's country and are published by Meta as downloadable rate cards in USD and other currencies. Meta updates them periodically, so always check the current rate card instead of relying on numbers from a blog post, including this one.

• Customer writes first and your bot answers within 24 hours: free messages.
• You send an appointment reminder (utility template) outside that window: charged.
• You send a promotion (marketing template): always charged.
• Customer arrives from a Click-to-WhatsApp ad: 72 hours of free messaging.

What a chatbot can do well

The best WhatsApp bots handle a narrow, repetitive job very well. Good candidates:

• Answering frequent questions: hours, location, prices, availability, shipping.
• Taking orders or reservations with a guided flow and confirming them.
• Sending appointment confirmations and reminders, and handling reschedules.
• Qualifying leads: asking budget, location and timing before a salesperson steps in.
• Order status lookups connected to your system.
• Collecting documents or photos and routing them to the right person.

What a chatbot should not do on its own

Language models can be wrong with great confidence. A bot that invents a price, a policy or a medical answer creates real problems. I design assistants so they only answer from your approved information, say "I don't know" when the answer is not there, and hand the conversation to a person for complaints, refunds, sensitive data or anything that requires judgment.

Meta also has commerce and messaging policies about what businesses can send and how users must opt in. Promotional messages need consent, and some industries have extra restrictions. Building those rules into the flow from day one is cheaper than fixing a blocked number later.

How I build a WhatsApp chatbot

My process is short and focused on the first measurable win:

• Pick one use case with high volume and clear rules, for example order taking or appointment reminders.
• Collect the real questions customers ask and the answers your team already gives.
• Set up the Business Platform number, webhook and message templates, and get them approved.
• Build the flow and, if it adds value, an AI layer limited to your own knowledge base.
• Test with real conversations, including edge cases and hand-offs to a human.
• Launch with supervision and adjust with the first weeks of conversations.

What it costs to build

The build cost depends on three things: the channel setup and verification, the integrations (booking, inventory, CRM, payments) and whether you need an AI model or a rules-based flow is enough. A frequently-asked-questions assistant is the simplest starting point; a bot that takes orders and writes them into your system is a small software project.

Then there are running costs: Meta's message charges, hosting for the webhook, and AI model usage if you use one. For most small businesses the running cost is driven by how many marketing and reminder templates they send, not by the bot replying to customers, because replies inside the customer service window are free.

Measuring whether your chatbot works

A bot is only useful if it resolves conversations. Before launch, agree on a few numbers to watch in the first weeks: the share of conversations resolved without a person, how many were handed off and why, response time, and how many bookings, orders or qualified leads the bot produced.

Reading a sample of real conversations every week is just as important as the numbers. It shows you which questions the bot misunderstands, which answers are missing from its knowledge base and where customers get stuck.

Where my experience comes from

I spent months at Outlier evaluating AI model responses for programming tasks, checking technical correctness, logic errors and instruction following. That work shapes how I build assistants: assume the model can be wrong, limit what it can claim, and test with real inputs. I also built Menu Engine, an MVP that uses AI to categorize menu items and prices with a human review step, which is the same pattern a good business chatbot needs.