Two Ways In

There are two ways to put an AI model behind a WhatsApp number. The official one is Meta's WhatsApp Business Platform, usually through the Cloud API: the business registers a number, receives messages through webhooks and replies through the API. The unofficial one uses a library such as Baileys, which its README describes as a WebSocket-based TypeScript library for the WhatsApp Web API. It links to an ordinary WhatsApp account the same way WhatsApp Web does, so no developer account, business verification or message templates are needed. The model call is the same in both cases. What differs is who allows the connection, and what it costs.

The Unofficial Route and Its Ban Risk

Baileys is not affiliated with WhatsApp. Its maintainers state that they do not condone uses that violate WhatsApp's Terms of Service, and they discourage bulk or automated messaging. The WhatsApp Terms of Service are explicit: users must not send impermissible communications such as bulk messaging, auto-messaging and auto-dialing, and must not use the service for non-personal purposes unless authorised. An auto-reply bot on a personal account falls inside that prohibition, so the number it runs on can be restricted or banned. The practical rules follow from that: use a spare number, keep the bot to people who expect it, cap how often it replies, and do not build customer service that a ban would take offline.

The Official Route: What It Costs in 2026

On the official platform, Meta's pricing page says that since July 1, 2025 it charges per message, by category and market, and only for messages from businesses to users; messages from users to businesses are free. Free-form replies, which is what an AI answer is, can only be sent inside a 24-hour customer service window that opens and resets with each user message. Reaching out first requires a pre-approved template. As of October 1, 2026, Meta charges for these service replies again, with a free tier of 1,000 service messages per business phone number per month that resets monthly and does not roll over. Conversations that start from a click-to-WhatsApp ad open a free entry point window of up to seven days. For a small support bot that mostly answers inbound questions, that free tier may cover a lot; a busy one should model its costs per market.

Meta's Rule on AI Providers

The official platform also restricts what kind of AI bot is allowed. In October 2025 WhatsApp added a section on AI providers to its business terms, reported by TechCrunch, effective January 15, 2026. It prohibits providers of large language models and general-purpose AI assistants from using the platform when the AI is the primary functionality being offered. Meta told TechCrunch that businesses using AI to serve their own customers, such as a travel company's support bot, are not affected. Europe is the exception in progress. Under regulatory pressure, Meta allowed general-purpose chatbots back in Europe from March 2026 for a fee of €0.0490 to €0.1323 per non-template message. On June 9, 2026 the European Commission ordered interim measures, requiring Meta to restore access for rival general-purpose AI assistants on the terms in place before October 15, 2025 while its investigation continues.

Which Route Fits

For a business answering its own customers, the official platform is the only route that will not put the number at risk. The AI-provider rule does not target that use case, and the cost depends on reply volume and market. For a product whose main offer is a general-purpose assistant on WhatsApp, the AI-provider rule applies, and its status depends on the region. Read the current terms before building. For a personal experiment, such as learning how the pieces fit or testing prompts with friends, an unofficial library on a spare number is the quickest way in, provided you accept that the number may be banned and keep the bot away from anything that matters.

Guardrails Worth Copying Either Way

Whatever the route, the same controls keep an AI bot cheap and predictable: an allowlist of who it answers, hourly caps per sender and overall so one chat cannot spend the whole model budget, a single reply in flight per chat, short in-memory history, and no tools that let a message trigger actions. Keep API keys and the WhatsApp session out of version control, and tell people that their messages go to a third-party model provider. Our open-source WhatsApp AI bot experiment shows these controls in a small Baileys and DeepSeek project, meant for a spare number rather than customers. Source review: October 29, 2026. Pricing and policy change often, so check Meta's pages before you commit.