AI WhatsApp business: how to use the API for customer support

For a small business, WhatsApp is often where customers already are. Ofcom reported that 76% of UK adults used WhatsApp in the three months before its 2023 research, making it the most widely used messaging app in the country. The same pattern holds in many other markets, which is why customers increasingly expect to ask a question, change a booking or chase an order in the same app they use for friends and family.

The difficulty is that a shared inbox on one phone does not scale. Messages arrive at night, staff are busy, and replies slow down just when a customer is most likely to buy or leave. An AI agent connected to the WhatsApp Business API can answer the routine questions straight away, hand the awkward ones to a person, and keep a record in your systems. This guide explains how that works, what the rules allow, and where small businesses go wrong.

Quick answer

An AI WhatsApp business setup connects an AI agent to the WhatsApp Business Platform Cloud API, so it can read incoming messages, answer from your approved information, book appointments and pass complex cases to staff. Meta now charges per message, service replies inside a 24-hour window are free, and general-purpose chatbots are banned.

What the WhatsApp Business API actually is

Most small businesses start with the WhatsApp Business app, a free app on a phone where one or a few people reply by hand. The WhatsApp Business API is a different thing. It is a programmatic connection that lets software send and receive messages on your business number, rather than a person typing into a handset.

Meta describes the Cloud API as something that enables you to programmatically message and call on WhatsApp. Your system sends messages by making requests to Meta's servers, and it learns about new customer messages through webhooks. Meta's documentation says webhooks inform you of incoming messages and the status of outgoing messages, along with account changes such as template quality updates. In practice, a customer writes to your number, Meta passes the message to your server, your software decides what to do, and your reply goes back the same way.

That software is where an AI agent sits. A simple auto-reply tool only matches keywords. An AI agent reads the meaning of the message, looks up what it needs in your calendar, order system or knowledge base, and writes a reply in your tone. If you want the broader picture of how this differs from a scripted bot, our guide to AI support agents covers the difference in detail.

You do not connect directly to Meta in most cases. Businesses usually go through a Business Solution Provider or build against the Cloud API themselves. Either way, the phone number needs to be registered for the API, which means it can no longer be used in the ordinary app on a handset at the same time, so plan for how your team will see and join conversations. A shared team inbox, built into the agent's handover tools, normally replaces the phone.

What an AI agent can do on WhatsApp

The most useful jobs are narrow, repetitive and tied to a clear outcome. In our experience the same handful of tasks account for most of the value for small businesses.

The common thread is that each job ends in an action in a system you already use, not just a pleasant reply. A chatbot that chats but cannot check a booking or update a record only moves the work somewhere else.

The rules that shape what you can build

WhatsApp is not an open channel. Meta sets rules that decide what an AI agent can send, when, and to whom. Getting these wrong leads to blocked messages or a suspended number, so they are worth understanding before you design anything.

The 24-hour customer service window

When a customer messages you, a 24-hour customer service window opens. Meta's documentation explains that service messages are free-form messages you can send during that window, and that once it closes you can only send pre-approved template messages. If the customer writes again before the timer runs out, it restarts.

For an AI agent, this shapes the whole design. Inside the window it can hold a natural conversation. Outside it, it can only send a template, such as a reminder or a follow-up, that Meta has approved in advance. Agents that ignore this rule fail silently when they try to reply to a customer who wrote the day before.

Opt-in and templates

Businesses must get permission before sending templates. Meta's documentation states that you must obtain user opt-in before sending message templates, and that the opt-in must make clear your business name and what you intend to send. Collect it at a natural point, such as a booking form or checkout, and record when and how it was given. A customer who messages you first has started the conversation, but you should still not assume that covers later marketing.

Pricing

Meta changed how it bills in 2025. Its pricing page states that Meta charges on a per-message basis from 1 July 2025, replacing the older conversation-based model. Templates fall into marketing, utility and authentication categories, and the same page says service conversations are free for all businesses. It also says utility templates sent inside an open service window are free.

There is another useful detail for businesses that run advertising. Meta describes free entry point windows, which open when someone messages you through a Click to WhatsApp ad or a Facebook Page button. These remain open for 72 hours, and during that time you can send any type of message at no charge. Rates differ by country and category and change over time, so always check the current rate card on Meta's pricing page rather than relying on a number quoted in an article, including this one. You will also pay a provider or build cost on top, which our guide to how much an AI agent costs breaks down.

No general-purpose chatbots

WhatsApp has restricted a particular kind of AI use. The WhatsApp Business Solution Terms say that providers and developers of AI technologies are strictly prohibited from accessing or using the WhatsApp Business Solution where such technologies are the primary functionality being made available, as opposed to incidental or ancillary. In plain terms, you cannot run an open-ended assistant that answers anything from anyone. You can run an agent whose job is to support your own customers with your own services.

That distinction matters when you scope the project. An agent that books appointments, answers questions about your products and tracks orders is ancillary to your business. A bot whose entire purpose is to be a general AI assistant on WhatsApp is not. Keep your agent focused on your services, and it fits comfortably inside the rules.

Privacy, disclosure and human oversight

A WhatsApp conversation contains personal data: names, phone numbers, addresses, and often more. The same data protection duties apply as in any other channel, and AI introduces a few specific questions.

The first is telling people they are talking to an AI. In the EU, the European Commission states that providers of AI systems that interact directly with people must design them so that individuals are informed that they are interacting with an AI system, and says Article 50 of the AI Act applies from 2 August 2026. The exception for cases where it is obvious should be read narrowly, so the safe course is a clear first message that says an AI assistant is answering, with an easy route to a human. Even outside the EU, this is good practice and builds trust.

The second is automated decisions. Most WhatsApp agents do not make significant decisions about people, but some touch on eligibility, pricing or access. Under UK GDPR, the ICO says safeguards for solely automated significant decisions include giving people information about the decision and enabling them to obtain human intervention, which it describes as a review of the decision with the possibility that you could change it, as well as the ability to contest it. If your agent could ever decline a request or set a price, design a human review step from the start.

The third is data handling. Decide what the agent may store, how long you keep transcripts, who can read them, and which third-party model providers see the content of messages. Our guide to GDPR and custom AI agents walks through these questions for UK and EU businesses. Regulation differs across markets, so a business serving customers in several countries should check the rules where its customers live, not only where it is based.

Where WhatsApp AI projects go wrong

Most failures come from the same few mistakes, and none of them are about the AI model.

How Klevere approaches WhatsApp agents

Klevere builds custom AI agents for small and mid-sized businesses, with 500+ AI agents deployed across 50+ projects and 12 industries, and a 98% client retention rate. For WhatsApp work, we start with the conversations your team already has rather than with a tool. We read a sample of real messages, group them into the jobs that repeat, and decide which ones an agent should handle fully, which it should prepare for a person, and which it should never touch.

From there we connect the agent to the systems that hold the answers, such as your calendar, CRM, accounting package or order system, so replies are based on live data. We write the handover rules with your team, set up the template messages and opt-in capture the platform requires, and add a plain disclosure that the customer is speaking to an AI assistant. Every agent has a review period where a person checks transcripts, and we adjust the instructions before widening what it can do.

We are also direct about when WhatsApp is not the right first step. If your customers rarely use it, or your enquiries are mostly complex and one-off, a different channel or a lighter workflow may pay back sooner. You can see how we think about this in our overview of custom AI agent development, and our AI vs human customer support guide helps decide what to automate at all.

A sensible rollout plan

A small business can reach a safe first version in stages, without a large project.

Treat the timings as a guide rather than a promise, because a clean data source and a clear owner matter more than the calendar.

Frequently asked questions

Can I use AI with the WhatsApp Business app instead of the API?

The ordinary WhatsApp Business app is built for a person replying by hand and offers only simple automations such as greeting and away messages. To connect an AI agent that reads and writes messages programmatically, you need the WhatsApp Business Platform. That means registering a number for the API, usually through a provider or a custom build, and moving your team to a shared inbox.

Is it legal to use an AI chatbot on WhatsApp?

Yes, provided you follow Meta's rules and local law. Meta's terms bar businesses whose main product is a general AI assistant, but a support or booking agent for your own services is acceptable. Law adds duties around consent, data protection and, in the EU, telling people they are speaking to an AI. Check the rules in each market you serve.

How much does a WhatsApp AI agent cost to run?

There are three parts: Meta's per-message charges for templates, the provider or platform fee, and the AI and build costs. Service replies inside the 24-hour window are free on Meta's side. Rates vary by country and message category, so check Meta's current pricing page and ask any provider to itemise their own fees separately.

Do customers have to agree before an AI agent messages them?

If you send template messages, such as reminders or promotions, Meta requires opt-in that states your business name and what you intend to send. When a customer messages you first, you can reply freely inside the 24-hour window. Keep a record of how and when consent was given, and offer an easy way to opt out.

Will an AI agent replace my staff on WhatsApp?

Usually it replaces the repetitive part of the work rather than the people. The agent handles routine questions and bookings, and staff handle complaints, negotiation and judgement calls with better context. Businesses that do this well often find the same team can answer faster and spend more time on the conversations that matter.

Can the same agent handle phone calls and WhatsApp?

It can share the same knowledge, rules and connections to your systems, which saves duplicated effort. The channels behave differently, though. Voice needs speed and short answers, while WhatsApp suits links, photos and a conversation that continues over hours. Our guide to AI voice agents covers the phone side.

If you are weighing up WhatsApp as a channel for your own business, the quickest way to find out what is worth automating is a free AI audit. We will look at your real enquiries and tell you plainly which ones an agent should handle and which it should leave alone.