Third-party AI agents can run on the WhatsApp Business API, as long as the agent exists to serve your business rather than to sell AI. Meta bars artificial intelligence providers from the platform where that technology is the primary thing being made available — which rules out general-purpose assistants, and leaves business-specific agents untouched.
How this guide is built: every rule below comes from Meta’s Terms for the WhatsApp Business Platform, section 4.7, read in full. Where the terms use their own words, they are quoted rather than paraphrased, because the distinction they draw is narrow.
Can you run a third-party AI agent on WhatsApp?
Yes, if the agent answers for your business. Meta prohibits AI providers from using the WhatsApp Business Platform to deliver their technology where the AI is the “primary functionality” being offered. An agent that handles your orders, bookings or support is not that, and you may hire an AI vendor to build it.
Whose product is the AI? Yours → allowed. The vendor’s, sold through you → banned.
- Allowed: an agent that answers questions about your business, built by you or a vendor you engage.
- Banned: a general-purpose assistant distributed to consumers over the platform.
- Banned: using platform data to train or improve a model that is not exclusively yours.
- Narrow exception: fine-tuning a model “exclusively for your own use”.
- Covers derived data too: the restriction reaches anonymous, aggregate and derived forms.
Source: Meta, Terms for the WhatsApp Business Platform, section 4.7.
What do Meta’s terms actually say?
The terms bar AI developers from the platform where the AI is what they are selling, using the phrase “primary functionality” to draw the line.
Artificial intelligence developers are “strictly prohibited” from accessing the platform to “provide, deliver, offer, sell, or otherwise make available” such technology where it constitutes “primary functionality”.
Read it carefully and the test is not about capability. Nothing there prohibits large language models, generative systems or autonomous agents. What it prohibits is using WhatsApp as the distribution channel for an AI product.
That is why a bank’s agent answering balance questions is untouched, and a general assistant that happens to live on a WhatsApp number is not. The technology can be identical. The business model is what differs.
Source: Meta, Terms for the WhatsApp Business Platform, section 4.7.
What is allowed and what is banned?
The same question decides every case: is the AI something your business uses to serve customers, or something being sold to them through your number?
| What you want to do | Verdict | Why |
|---|---|---|
| An agent answering about your orders | Allowed | The AI serves your business; it is not the product |
| An LLM writing those answers | Allowed | Capability is not what the clause restricts |
| Hiring an AI vendor to build it | Allowed | Businesses may engage AI providers as vendors |
| Fine-tuning a model on your data | Allowed | Only where the model is exclusively for your own use |
| A general assistant on your number | Banned | The AI is the primary functionality offered |
| Reselling AI access to consumers | Banned | The platform becomes a distribution channel |
| Training a vendor’s shared model | Banned | Platform data may not improve models beyond yours |
| Sharing anonymised chat data for training | Banned | The restriction covers derived and aggregate forms |
If you cannot say which business the agent belongs to, that is the answer to whether it is allowed.
Source: Meta, Terms for the WhatsApp Business Platform, section 4.7.
Can your AI vendor train on WhatsApp data?
No — not on a model anyone else will use. The terms prohibit using platform data to “create, develop, train or improve” any machine learning system, “including large language models”, with one exception.
That exception is fine-tuning a model “exclusively for your own use”. A model you tune on your own conversations, used only by your business, is permitted. The moment that tuning improves something the vendor also serves to other customers, it is not.
The restriction also reaches further than most teams assume: it applies to “anonymous, aggregate or derived” data. Stripping names from transcripts does not put them outside the clause, and neither does turning them into embeddings or training statistics.
This is the part worth putting in a vendor contract rather than an email. The obligation sits with you, and a vendor with a shared model has a commercial reason not to volunteer the distinction.
Source: Meta, Terms for the WhatsApp Business Platform, section 4.7.
What should you ask an AI vendor before signing?
Ask three questions, and get the answers in the contract: whose model is it, where does our data go, and what happens to it after the conversation ends.
- Is the model exclusively ours? A shared model improved by our conversations is outside what the terms allow.
- Is our data used for training, in any form? Including anonymised, aggregated or derived forms.
- Is the agent scoped to our business? An assistant that answers anything is a different product with a different status.
- Who holds the WhatsApp account? The platform relationship should sit with you, not inside a vendor’s shared account.
- What is retained, and for how long? Retention that exists to build a dataset is the problem the clause addresses.
A vendor who answers these plainly is straightforward to work with. One who calls them implementation details is telling you something.
How do you run a compliant AI agent on WhatsApp?
Scope the agent to your business, keep the platform account in your name, and make sure the model it runs on is either yours or tuned exclusively for you.
Everything else is the same as any other messaging build. The agent still replies inside the 24-hour window or through approved templates, still needs opt-in before it messages first, and still hands to a human when it should. Our guide to chatbots and AI agents covers where that handover belongs, and WhatsApp Flows covers the structured screens an agent should use when it needs exact fields rather than free text.
The compliance work is mostly procurement, not engineering: knowing which model you are on, and being able to show it.
How does UDO build this for your business?
UDO runs the WhatsApp Business API account in your name, builds the agent against your own systems, and keeps the data arrangement on the right side of the terms.
Billing is per delivered message, with no account opening fee and no monthly subscription — you can see pricing for the current rates by message category. As an Official Meta Tech Provider, we handle the platform verification and the template approvals an agent depends on.
- The account is yours: the platform relationship sits with your business, not inside a shared vendor account.
- Scoped to your business: the agent answers about your products, orders and policies, which is the side of the line that is allowed.
- Connects to what you run: Salesforce, HubSpot, Zoho, Freshdesk, CleverTap and MoEngage, or your own systems through the REST API.
- One provider for every channel: WhatsApp, SMS, RCS and voice on one account.
- Human handover built in: a shared inbox behind the agent, so a real person takes over when the conversation needs one.
Source: Unique Digital Outreach, pricing page and WhatsApp Business API.
Build the agent that is actually yours
Frequently asked questions
Yes, when the agent serves your business. Meta prohibits AI providers from using the platform to deliver their technology where the AI is the primary functionality being offered, which targets general-purpose assistants rather than business-specific agents.
A general-purpose assistant distributed to consumers over the platform is what the clause prohibits. The same underlying model, used inside an agent that answers about your own products and orders, is not affected.
Yes. Businesses may engage AI providers as vendors. What changes is the data arrangement: platform data cannot be used to create, develop, train or improve a model beyond one used exclusively by your business.
Only a model used exclusively for your own business. Fine-tuning for your own use is permitted; using the same data to improve a model the vendor serves to other customers is not.
No. The restriction covers anonymous, aggregate and derived data, so stripping identifiers does not move a dataset outside it.
It is a breach of the platform terms, and platform access is what is at risk. Because the account and the number are the business’s own, the consequence lands on the business rather than on the vendor that built the agent.
Not for this rule. The clause is about whether the AI is the product being made available, not about how the automation is built, so a rule-based bot and an AI agent are treated the same way.
Your business. The obligations in the terms attach to the account holder, so an agent running inside a vendor’s shared account leaves you relying on their compliance rather than your own.
Only through an approved template, to a customer who has opted in. Inside the 24-hour window after a customer writes to you, the agent can reply freely without a template.
Not as a rule of the terms, but as a matter of running it well. An agent that cannot hand a conversation to a person turns an unanswered question into a lost customer, which is why a shared inbox sits behind ours.
Sources
- Meta, Terms for the WhatsApp Business Platform, section 4.7: the prohibition on AI developers where the technology is the primary functionality; the allowance for businesses to retain AI providers as vendors; the bar on using platform data to create, develop, train or improve machine learning systems including large language models; the exception for fine-tuning exclusively for your own use; and the extension to anonymous, aggregate and derived data.
- Meta, WhatsApp Business Management API: message templates: template approval and the 24-hour customer service window.
- Unique Digital Outreach, pricing page: billing per delivered message, with no account opening fee and no monthly subscription.
