On 15 January 2026, Meta amended the WhatsApp Business Solution Terms to prohibit general-purpose AI chatbots from operating on the WhatsApp Business Platform. ChatGPT-style assistants, Perplexity-style answer bots, and similar open-ended conversational agents are no longer allowed to talk to customers on WhatsApp. OpenAI publicly announced that ChatGPT would no longer be available on WhatsApp after 15 January 2026, and Perplexity, Microsoft, and several other providers shut down their consumer WhatsApp services.

The policy actually took effect earlier for new accounts. Meta updated the Business Solution Terms for new business accounts on 15 October 2025, three months before the deadline for existing accounts. So by the time this guide is published, the rule has been in force for at least seven months for new merchants and four months for everyone else.

Most established merchants we speak to still do not know it happened. Several apps in the Shopify App Store still ship features that pipe inbound WhatsApp messages directly into a large language model and pipe responses back. Those features are now non-compliant. If you are running one on your business number, your account is at risk of throttling, template pauses, or suspension.

The change does not mean WhatsApp is anti-AI. It means Meta has drawn a sharp line between two kinds of automation, and stores need to understand which side of the line their tooling sits on. This guide explains what changed, why Meta did it, what is still allowed, how to audit your own setup, and what enforcement has actually looked like since January.

What exactly changed under the updated terms

The WhatsApp Business Solution Terms now prohibit AI providers from distributing their AI chatbots on the WhatsApp Business Platform. The wording targets bots that "predominantly use or introduce open-ended or assistant-style conversations," "share chat data for AI model training or improvement," or "simulate broad AI assistants (for example, ChatGPT or Perplexity on WhatsApp)."

The practical interpretation: automated agents on WhatsApp must be task-bound. Acceptable uses include support agents that answer questions about a specific business's products, order tracking bots, booking and appointment bots, FAQ bots with a defined knowledge base, and product recommendation engines scoped to a single catalog.

Prohibited uses include general-purpose AI assistants that answer any question a user asks, AI companions or entertainment bots, knowledge bots that pull from the open internet, and any wrapper around a frontier large language model that accepts unrestricted user input.

The opt-in requirements also tightened. Before any business-initiated message goes out, the user must give explicit consent. Generic SMS or email opt-ins do not count for WhatsApp.

Why Meta did this

The official rationale is user protection. WhatsApp is a private messaging app for over 3.3 billion people, and Meta has consistently positioned it as a personal channel that businesses are guests on. General-purpose AI introduces unpredictability. If a bot can answer any question, it can also produce harmful, inaccurate, or off-brand answers in a setting where the user trusts the brand on the other end.

There is a second reason that often goes unsaid. AI assistants on WhatsApp generated significant inbound traffic that did not produce revenue for Meta. Customers were treating business numbers as free ChatGPT, sending unrelated questions and burning conversation windows that BSPs and businesses were paying for. Meta has explicitly stated that the new chatbot use cases placed unsustainable load on its infrastructure.

There is also a competitive dimension. Meta launched Meta AI as a consumer assistant inside WhatsApp, Messenger, and Instagram. By restricting third-party general-purpose AI on the API, Meta AI becomes the only assistant available on WhatsApp. Regulators in the EU, Italy, and Brazil have opened antitrust probes into this aspect of the policy, but the policy itself remains in force.

The third reason is regulatory. The EU AI Act, in phased enforcement through 2025 and 2026, requires risk classification of AI systems including chatbots. Meta would rather force partners into well-understood, narrow-scope agents than carry the regulatory exposure of letting any frontier model be deployed at WhatsApp's scale.

What is still allowed

Plenty. The ban is on general-purpose AI, not on automation, and not on AI as a category. Here is what you can keep running.

Task-bound support agents. A bot that answers questions specifically about your store's products, your shipping policy, your return process, your sizing, or your stock availability is allowed. The agent can be powered by a large language model under the hood; the constraint is on its scope. If the customer asks about something outside your business, the agent must decline or hand off to a human, not improvise an answer.

Order tracking bots. Bots that take an order number, look up status in your fulfilment system, and report shipping milestones are explicitly fine. This is one of the highest-volume use cases on WhatsApp and a major reason customers prefer it over email.

Booking and appointment bots. A bot that walks a customer through selecting a service, a time slot, and confirming an appointment is task-bound by definition.

Recommendation engines scoped to your catalog. A bot that asks about preferences and recommends products from your store, or runs a product finder quiz, is allowed. It produces predictable outputs from a defined input space.

FAQ bots with a defined knowledge base. If your bot's responses are drawn from your help centre, your product documentation, or a curated list of answers, you are within policy.

Lead qualification flows. Bots that ask a structured set of questions to qualify a lead and route it to sales are allowed.

Conversational checkout assistants. Agents that help a customer finalise an order, including answering product questions during the purchase, applying discount codes, or modifying cart contents, are allowed because they sit in a clearly defined commerce context.

What is no longer allowed

The disallowed cases are easier to recognise because they all share one property: the bot will respond to anything.

A WhatsApp bot that answers "what is the capital of Brazil," "write me a poem," or "what is the best ice cream brand" is in violation. Even if the bot is intended for customer support, the test is what the bot can do, not what it is supposed to do.

A WhatsApp bot that scrapes the internet for answers is in violation. The boundary on knowledge sources is your own business context plus information you have explicitly licensed for the bot to use.

A WhatsApp bot that role-plays a personality, acts as a companion, or simulates a celebrity is in violation. These were never the intended use case for the Business Platform and are now explicitly out.

A WhatsApp bot that wraps ChatGPT, Claude, or any other frontier model with no constraints on input or output is in violation, regardless of how the user reached it.

How to audit your current setup

If you have any AI feature live on your WhatsApp business number, run this audit this week.

First, ask one question: what is the broadest input my bot will accept and respond to substantively? If the answer is "any question," you have a compliance problem. If the answer is "any question about my products, orders, returns, shipping, and account," you are within scope.

Second, test your own bot. Open WhatsApp, message your business number, and send three deliberately off-topic messages. Ask about a current event. Ask for a recipe. Ask for help with a coding problem. If your bot answers any of these substantively, it is producing prohibited outputs.

Third, look at the logs. Most WhatsApp marketing platforms expose conversation logs in their admin panel. Skim the past month of inbound messages. If customers are using your bot for unrelated tasks (a surprisingly common pattern once people figure out a bot is AI-powered), that is exactly the behaviour Meta wants to stop.

Fourth, check what your provider says they support. Some platforms have already added scope restrictions to their AI agents. Others will eventually. If your provider's marketing still emphasises "ChatGPT-powered" or "fully conversational AI on WhatsApp," ask them directly whether their bots are configured to refuse off-topic queries.

How to rebuild a compliant AI agent

The replacement pattern is straightforward and arguably produces better customer experience than the open-ended approach it replaces.

Start by writing down the explicit scope of your agent. Two or three categories: products and inventory, orders and fulfilment, returns and exchanges, account and billing. Anything outside these categories should be handled by a polite handoff message that routes to a human or to a static help link.

Next, build a system prompt that constrains the agent to those categories. Modern frameworks like LangChain, LlamaIndex, and the major BSP platforms make this straightforward. The system prompt should tell the model what it is allowed to discuss and explicitly instruct it to decline anything else with a fixed handoff message.

Ground the agent's knowledge in your data, not the open web. Use retrieval-augmented generation (RAG) over your product catalog, your help docs, your shipping policy, and your order data. The model retrieves the relevant facts from your sources and answers from them. This pattern eliminates hallucination and stays within the policy boundary.

Add an explicit fallback. When the agent does not know an answer, it should not improvise. It should either ask a clarifying question, look up the relevant policy, or hand off to a human agent. A bot that says "I am not sure about that, let me get a person to help" is a compliant bot.

Test the boundaries with adversarial prompts before you ship. Try to get the bot to discuss politics, give legal advice, recommend competitors' products, or write code. If it complies with any of these, your scope restriction is leaky.

What enforcement has looked like since January

Meta's enforcement of the AI policy has been gradual since 15 January 2026, but the trajectory is clear. The first wave was warnings via the WhatsApp Business Manager. The second wave has been quality rating drops on accounts where bots were generating high block or "report spam" rates from off-topic interactions. We have not yet seen large-scale account terminations specifically for AI policy violations among Shopify-scale merchants, but the enforcement infrastructure is in place.

What is clear from the early enforcement pattern: Meta is not running adversarial probes against bots to test if they answer off-topic questions. Enforcement is reactive to user reports and quality signal degradation. So the practical risk is not "will Meta catch me," it is "will customers using my bot for off-topic queries report me as spam often enough to drop my Quality Rating."

The other piece worth knowing: Meta's reviewers are real people, not just automated systems. If a customer reports your bot for unrelated AI behaviour, a human will sample the recent conversation history. What they see has to look like a business agent, not a chatbot novelty.

Frequently asked questions

Does this mean I cannot use AI on WhatsApp at all?

No. AI is fine, the constraint is on scope. Task-bound agents that solve specific business problems for your customers are explicitly allowed. The ban is specifically on general-purpose AI that responds to anything.

What about voice agents on WhatsApp Business Calling?

The same scope restriction applies. WhatsApp Business Calling is a voice channel and voice agents on that channel must be task-bound for the same reasons text agents must be. Open-ended voice assistants are not permitted.

Can I still hand off from a bot to a human?

Yes, and you should. A bot that hands off to a human for anything outside its scope is the compliant pattern. Most modern WhatsApp platforms have a built-in handoff mechanism that flips the conversation to a human agent inbox.

Are Meta's own AI assistants exempt from this rule?

Meta AI inside WhatsApp is a consumer feature, not a business feature, and operates under a separate policy framework. The Business Solution Terms apply specifically to bots on the WhatsApp Business Platform operating on business phone numbers.

What happens if my account is flagged for AI policy violations?

The escalation typically goes warning, then quality rating drop, then template pause for the offending automation, then in serious cases account restriction. Some cases have resolved within 48 hours when the merchant disabled the offending bot and submitted an explanation through Direct Support.

My BSP platform offers an AI feature. How do I know if it is compliant?

Ask them directly two questions. First, is your AI agent scoped to my business context (products, orders, support topics) or is it open-ended? Second, does it have explicit refusal behaviour for off-topic prompts? A compliant provider will answer both clearly. An evasive answer is itself a signal.

Can I use AI for outbound message generation?

Yes. Drafting marketing copy, generating product descriptions, or composing template content with AI is unrestricted. The policy applies to the live bot that talks to customers in real time, not to the back-office tools you use to prepare messages.

Do I need to disclose to customers that I am using AI?

Meta has not mandated a disclaimer at the start of every conversation, but the broader expectation under WhatsApp's policy is that you should be honest if a customer asks whether they are talking to a person. AI tools should answer that question truthfully if asked.