AI Chatbots on WhatsApp: Understanding Meta's Current Rules

Building an AI chatbot on WhatsApp is still allowed for your company, even after Meta prohibited general-purpose AIs on the official API. What has changed is the approach: the bot needs to exist to serve your customer about your business, not to answer anything to anyone. In this guide, I show you where that line is, how to set up customer service, and its cost.

TL;DR

• Section 4.7 of the WhatsApp Business Platform terms prohibits AI providers when AI is the primary functionality of the service. A company's own customer service bot is an ancillary functionality and remains permitted.
• You can use ChatGPT, Claude, or Gemini behind your bot. The condition: conversation data cannot train the provider's models.
• As of October 1, 2026, Meta also charges for responses within the 24-hour window: $0.0085 per message in the United States.
• The real risk of blocking comes from an open scope, lack of human handover, and declining quality score, not from the use of AI itself.

What Meta's Rule Prohibits, in Its Own Words

The confusion began in October 2025, when Meta announced that, effective January 15, 2026, general-purpose AI assistants like ChatGPT, Perplexity, and Luzia would no longer be able to operate within WhatsApp via the commercial API. The headline became "Meta prohibits AI on WhatsApp." That's not what the text says.

The text currently resides in Section 4.7 of the Meta Terms for the WhatsApp Business Platform, last updated September 23, 2026. It states three things:

1. Who is affected: "providers and developers of artificial intelligence or machine learning technologies," as determined by Meta in its sole discretion.
2. When: when these technologies are the primary functionality, "rather than incidental or ancillary," made available for use.
3. What's still allowed: the company may retain an AI provider as its solution provider, provided that platform data is not used to "create, develop, train, or improve" AI systems or models. Fine-tuning for the company's exclusive use is permitted.

In other words: the rule targets those who sell AI as a product within WhatsApp. A store, school, or clinic that uses AI to serve its own customers is on the other side of the line, and Meta stated to the press, in the announcement, that this use would not be affected.

And the FTC's Stance?

In Brazil, the antitrust regulator CADE suspended the application of the new terms in January 2026 and upheld the measure in March, still without a final ruling. For companies looking to set up an AI chatbot on their own WhatsApp, this dispute changes little: it concerns general-purpose assistants. The complete timeline is in the post about automating WhatsApp customer service with AI without getting banned.

Company AI Chatbot on WhatsApp: What's Still Allowed

The question I ask every client before connecting a language model to the API is simple: if your AI chatbot on WhatsApp disappeared tomorrow, would your customer service still make sense? If the answer is yes, AI is ancillary. It accelerates an existing service.

The table summarizes our interpretation of the terms in operation:

Use Case	Status	Reason
Bot answering questions about store products, delivery times, and orders	Permitted	AI is ancillary to the company's customer service
Ad lead qualification and scheduling visits or consultations	Permitted	Serves the company's own customers about its own service
Student support for an online course (access, certificates, deadlines)	Permitted	Institution's service, AI as a tool
Bot using GPT, Claude, or Gemini via API with company's own key	Permitted, with condition	The provider cannot train models with conversation data
"Ask our assistant anything" offered as a product	Prohibited	AI becomes the primary functionality
Reselling access to an AI model via a WhatsApp number	Prohibited	This is precisely the case of an AI provider
Company bot that accepts conversations on any topic	Gray area	In practice, it becomes a general assistant and exposes the account

The last row is where most companies stumble without realizing it. No one plans to create a general-purpose assistant. But a lazy prompt, such as "you are a friendly assistant for company X," allows the model to respond to cake recipes and homework questions. Then your AI chatbot on WhatsApp starts functioning like a generic AI with a logo.

The "Ancillary" Test: 5 Questions Before Deploying Your Model

Since Meta decides what is primary or ancillary at its sole discretion, there's no certificate of compliance, only a defensible design. These are the questions we use to validate an AI chatbot on WhatsApp before publishing.

1. Can the bot say "that's not my area"?

The system prompt needs to define the scope: company products, services, policies, and data. Outside of that, the bot politely declines and redirects to relevant topics.

2. Is the company speaking, not the model?

The customer interacts with the "Store XYZ," not "ChatGPT." The name, tone, and signature belong to the brand. If your communication sells the model as an attraction, you're pushing the project to the wrong side of the rule.

3. Do responses come from your knowledge base?

The model should respond based on your content: catalog, price list, regulations, FAQ, order status. This technique is called RAG (retrieval-augmented generation): the system retrieves the relevant passage and provides it to the model along with the query. "Memorized" responses are where fabrications originate.

4. Is a human agent just a tap away?

Human handover is the transfer of the conversation from the bot to an agent. It needs to be triggerable by keyword ("agent," "human"), by detected intent (complaint, cancellation), and by failure (the bot didn't understand twice in a row).

5. Does your contract with the AI provider prohibit training with your data?

This is the explicit condition in Section 4.7. Check the commercial terms of the provider's API (not those of the consumer application) to see if the data sent is used to train models, and keep that evidence. If the provider doesn't guarantee this in writing, switch providers.

How to Set Up an AI Chatbot on WhatsApp Without Risk of Blocking

This is the sequence we follow at Agathas Web to put an AI chatbot on WhatsApp into production. It applies to a store with one agent or an institution with ten.

Step 1: Start with the Official API, Never the QR Code

Tools that connect to WhatsApp Web via QR Code violate the application's terms and are the most common cause of banned numbers. AI plugged into them multiplies the risk. The foundation is the official API, with the account verified in Business Manager. If you're not yet familiar with the architecture, the complete guide to the WhatsApp Cloud API explains the components.

Step 2: Write the Knowledge Base Before the Prompt

Gather what a new agent would need to read on their first day: catalog, hours, return policies, deadlines, and real questions from conversation history. Concise and updated text is worth more than an 80-page PDF. Define who updates it.

Step 3: Confine the Scope in the System Prompt

The prompt defines who the bot represents, what topics it accepts, when it transfers to a human, and what it never promises (discounts, out-of-policy deadlines, diagnoses). Responses should be one or two messages: those who receive five consecutive bubbles tend to block the number.

Step 4: AI for Inbound, Human for Queued Conversations

The AI chatbot on WhatsApp excels in inbound interactions, when the customer initiates the conversation. Proactive messages outside the 24-hour window require an approved template and opt-in, so don't let the AI send anything on its own. Also define who receives transferred conversations and what the bot says after hours: a bot that promises "I'll transfer you" and then disappears is the shortest path to a complaint.

Step 5: Test with External Users

Ask three external people to try to break the AI chatbot on WhatsApp: off-topic questions, discount requests, insults, audio, photos. Adjust the prompt and knowledge base until failures turn into polite refusals or handovers.

Step 6: Monitor Your Quality Score Weekly

Meta assigns a quality score to the number based on blocks and reports; a drop is the first sign that the bot is irritating users. If more than half of conversations end up with a human, the knowledge base is weak.

How Much Does It Cost to Operate an AI Chatbot on WhatsApp in 2026

The cost has three layers, and confusing them is the most common budgeting pitfall.

1. Message Delivery by Meta. As of July 2025, billing is per message delivered. And, according to the official pricing page for non-template messages, starting October 1, 2026, service messages (responses within the 24-hour window) began to be charged, at the same rate as utility and authentication messages. In the United States, the reference rate cited by Meta is $0.0085 per message, without volume tiers. The 72-hour window opened by a Click to WhatsApp ad remains free of delivery cost. Details in the post about cost per conversation on the official WhatsApp API.

2. The AI Model. If you use your own model (GPT, Claude, Gemini) with your key, you pay the provider per token, separately from Meta. If you use Meta's own agent, the Meta Business Agent, the charge is unified: $2.00 per million tokens, billed since August 1, 2026, covering AI and delivery, including within the free windows.

3. The Platform. Inbox, agent queue, reports, AI orchestration. This is the tool's monthly fee.

Meta itself publishes an estimate for 10,000 AI messages in the United States:

Scenario (10,000 messages, United States)	Cost per message	Estimated Total
Service + Third-party AI, low complexity	~$0.02 to $0.03	~$285
Service + Third-party AI, high complexity	~$0.09 to $0.10	~$985
Meta Business Agent, high complexity	~$0.04 to $0.05	~$400 to $500

Two takeaways emerge from this table. First: context size dictates cost, so a lean knowledge base and short history reduce model costs. Second: compare the cost per resolved interaction, not per message. An AI chatbot on WhatsApp that resolves an issue in three messages is cheaper than a "cheap" one that needs twelve and still transfers.

Meta Business Agent, Platform with Own AI, or Custom Development

There are three paths to having an AI chatbot on WhatsApp via the official API. None is superior in all aspects.

Criterion	Meta Business Agent	Platform with Own AI (BYO)	Custom Development
Model choice	Meta's only	GPT, Claude, Gemini, DeepSeek	Any
AI billing	Per token, combined with delivery	Directly with provider, per token	Directly with provider
Inbox and human queue	Limited	Included	Needs to be built
System integration (ERP, Moodle, CRM)	Limited	Via API and webhooks	Full
Time to go live	Days	Days	Weeks or months

Meta's agent suffices for small operations and repetitive questions. The platform with its own AI serves SMEs that want to choose the model, control token costs, and have the human team on the same screen, which is the case for most of the clients we serve. Custom development is worthwhile when the bot needs to query and record in proprietary systems, such as Moodle enrollment or duplicate bill requests.

When NOT to Implement AI on Your WhatsApp

Not every customer service scenario should include AI, and stating this before a contract saves the client money. These are cases where an AI chatbot on WhatsApp hinders more than it helps:

• Low volume and consultative customer service. With 15 conversations per day, each one a negotiation, an agent with quick responses resolves issues better.
• Debt collection and negotiation. The wrong tone from a model, in this context, turns into a formal complaint. Use utility templates and human agents for negotiation.
• Health, legal, and financial services requiring individual guidance. The bot can schedule and triage. It cannot provide guidance.
• Non-existent knowledge base. Without reliable content, AI fabricates. First organize the information, then automate.
• No one to manage it. An unmanaged AI chatbot on WhatsApp becomes outdated in weeks: prices change, and it continues to provide old information.

Common Pitfalls We Observe

• Using the consumer version of the AI provider, whose terms do not guarantee the prohibition of training.
• Budgeting only for the platform's monthly fee and forgetting Meta's delivery costs and tokens.
• Ignoring the WhatsApp API changes calendar for 2026-2028: prices and rules change several times a year.

How Agathas Web Solves This

We have developed systems and maintained servers since 2008, from critical Moodle environments to integrations with the official WhatsApp API. From this operation, Voyia, our customer service platform for the official WhatsApp API, was born. It was designed precisely for the scenario in this post: ancillary AI, with the company in control.

What's Included:

• Connection of your number to Meta's official API, with support for verification and migration from the WhatsApp Business App or another platform.
• Multi-agent inbox with queue and transfer, also available on iOS and Android apps.
• AI integration in the BYO (bring your own key) model: you connect Claude, GPT, Gemini, or DeepSeek, pay the provider directly, and Voyia orchestrates the conversation, maintains context, and transfers to a human when needed.
• Meta fees passed on without markup, billing in US dollars, and data on servers in the United States.

How it Works in Practice: You sign up, we connect your number, and you configure the knowledge base, prompt, and flows using the documentation and dashboard videos. If you prefer, we can set up the flow and scope prompt under a separate quote.

How to Subscribe: AI integration is available in the Professional ($69/month, up to 3 agents) and Business ($149/month, up to 5 agents) plans, with no long-term commitment. The Professional plan has a setup fee, which will be communicated by the sales team before subscription. See details and contact the team on the Voyia page.

Conclusion: Meta's Rule is a Filter, Not a Prohibition

Meta's rule didn't close the door on AI in customer service. It closed the door on those who use WhatsApp as a showcase for a generic assistant. An AI chatbot on WhatsApp with a business-specific scope, its own knowledge base, human handover, and a contract prohibiting training with your data is on the right side of the line.

Next step: list the 20 most frequent questions that come to your WhatsApp, separate those a bot can resolve independently, and estimate the monthly volume. With this, the conversation about platform and cost becomes objective. If you'd like to do this exercise with us, explore Voyia and speak with our team.