BlogConversational AI

WhatsApp AI assistants for business: use cases and limits

A WhatsApp AI assistant is permitted, and works well, when it serves your own customers: bookings, order status, lead qualification and reminders. The limits that shape the build are Meta’s. Free-form replies are allowed only within 24 hours of the customer’s last message, anything later needs a pre-approved template, messages are billed individually, customers must opt in, and general-purpose AI chatbots are barred as a product.

What Meta allows and what it restricts

Everything in this section is as the official pages read at the time of writing. Meta revises these documents often, so follow the links before you commit to a design.

The AI provider rule

The Meta Terms for WhatsApp Business Platform, last updated on 23 September 2026, have a section headed AI Providers (4.7). In substance: providers and developers of AI and machine learning technologies, with large language models, generative AI platforms and general-purpose AI assistants given as examples, may not use the platform to offer those technologies when they are the “primary (rather than incidental or ancillary) functionality” on offer. Meta decides which side of that line a service falls on.

Three further points from the same section matter to an ordinary business. You may retain an AI provider as your solution provider. Data from the platform may not be used to train or improve AI models, with an allowance for fine-tuning a model that is for your exclusive use. And the section says such services may be made available to businesses in certain countries, which Meta lists separately. Competition regulators have challenged the rule, so the geography keeps moving.

In practice, a clinic’s booking assistant or a retailer’s order assistant is on the right side, because the AI is incidental to the service the business provides. A number that lets anyone chat with a language model about anything is the case the rule targets. Keep your assistant scoped to your own business and have it decline general requests. That is sound design anyway.

The customer service window and templates

When a customer messages or calls your business number, a 24-hour timer starts, and each new message from them resets it. Meta calls this the customer service window. Inside it you can send free-form replies: text, media, lists, up to three reply buttons. This is the only place a generative assistant can speak in its own words.

Outside the window you can send only message templates: pre-written messages with variables, each assigned to a category (marketing, utility or authentication) and reviewed by Meta before use. Review can take up to 24 hours, and a template that draws complaints or low read rates can be paused.

The consequence people miss: your assistant cannot say “I’ll message you tomorrow when the part arrives” and then do so in free text. Tomorrow’s message is a utility template you wrote and had approved weeks earlier.

Opt-in, escalation and sensitive data

The WhatsApp Business Messaging Policy, also updated on 23 September 2026, sets three rules that land directly on an assistant:

  • Opt-in. You may contact people only if they gave you their number or username and opted in to messages from your business. How you collect that is your responsibility.
  • A way out of automation. Automation is allowed inside the window, provided customers have a clear and direct route to escalate, such as transfer to a human agent in the chat, a phone number or an email address.
  • Sensitive data. Do not ask for full payment card numbers, financial account numbers or personal ID numbers in chat, and do not exchange health information where local regulation prohibits it.

What does it cost in Meta fees?

Meta charges per delivered message, by category and by the recipient’s country. The older per-conversation model ended on 1 July 2025.

For an assistant, the category that matters is service: the free-form replies inside the window. Until 30 September 2026 those were free. Since 1 October 2026 each business phone number gets 1,000 free delivered service messages a month and pays for the rest, and utility templates sent inside an open window are charged again. Marketing, utility and authentication templates each have their own rate, with volume tiers for utility and authentication only. Rates vary by country, so we will not quote any here.

Two details are worth knowing. If a customer arrives from a click-to-WhatsApp ad in the mobile app and you reply within the window, your messages to them are free for up to seven days. And Meta now sells its own configurable assistant, Meta Business Agent, billed by tokens instead of messages. Whether to use it or build your own is the usual build or buy question, with the added note that Meta excludes some sectors, among them finance and health.

The design lesson is blunt. Once the free tier is used, every bubble your assistant sends is a billable message. An assistant that splits each reply into three short messages costs three times as much as one that sends a single complete reply. Your language model and hosting bills come on top.

Use cases that work

Take a dental group with three branches.

Bookings. The patient writes “can I come in Thursday afternoon”, the assistant checks the diary through a tool and offers the free slots as a list, and the patient taps one. This suits WhatsApp because it is short and ends in an action.

Reminders. A utility template goes out the day before with “Confirm” and “Reschedule” buttons. A tap opens the service window, and the assistant handles the rescheduling in free text.

Order and status lookups. “Is my aligner ready” is a read-only query against a system you already have.

Lead qualification. Someone taps an ad, the assistant asks three or four questions (treatment, branch, timing), and hands the lead to the front desk with a summary. The free window after an ad click makes this cheap.

What works poorly: long advisory conversations, anything that needs a signature or a card number, cold outreach to people who never opted in, and open-ended chat.

How do you build one?

The architecture is small. The discipline is in the edges.

Webhook. Meta posts incoming messages to an HTTPS endpoint you register. You verify the endpoint once by echoing a challenge, then validate the signature header on every request using your app secret. Acknowledge with a 200 at once and do the work on a queue. Failed deliveries are retried for up to seven days, which means duplicates, so deduplicate on the message ID. There is no API to fetch old webhooks, so store what you receive.

Identity. Do not key your users on phone numbers. WhatsApp now supports usernames, and Meta’s documentation on business-scoped user IDs says a customer’s phone number may be left out of webhooks once they adopt one. Store the scoped ID and treat the phone number as optional.

Session state. For each user keep the conversation history, the time the window closes, what has been collected so far, consent, and whether a human has taken over:

{
  "user_id": "business-scoped user ID",
  "phone": null,
  "window_closes_at": "2026-10-05T09:12:00Z",
  "opt_in": { "utility": true, "marketing": false, "source": "booking form" },
  "collected": { "branch": "Riverside", "treatment": "check-up" },
  "human_has_taken_over": false
}

People send three short messages in a row. Wait a couple of seconds and combine them before calling the model, or it will answer the first fragment.

Model with tools. The model gets a short set of tools (check availability, create booking, look up order) and the rule that slots and statuses come only from tool results. Have it return a structured reply, text plus optional buttons, which your code turns into the right WhatsApp message type. Connecting the model to more of your systems is standard agent tooling.

Message design. WhatsApp replies are read on a phone, between other chats. Keep each reply to one message of a few lines, put choices on buttons or a list instead of asking the customer to type, and show the typing indicator while the model works so that the wait does not look like silence. Customers write in the language and spelling they use with friends, so test with real messages, not tidy ones.

Human handoff. A flag on the session stops the assistant, the conversation appears in a shared inbox with a summary, and the assistant resumes when the agent closes it. The policy requires the escape route, and customers expect it. The wider design is in our guide to AI customer support.

Outside the window. Every scheduled send checks the window time first and falls back to a template.

Opt-in and privacy

Record opt-in per category with its source and date, keep marketing consent separate from service messages, and act on “stop” at once. Businesses that message people who did not expect it get blocked and reported, their quality rating drops and Meta rate-limits them. New business portfolios also start with a cap of 250 unique customers in a rolling 24 hours for messages outside the window, which rises with verification and good quality.

On privacy, the chat is personal data and your language model provider is a processor of it. Sign the data processing terms, choose the region, set retention, and send the model only what the task needs. Tell customers in the first message that they are talking to an automated assistant. Send a payment link instead of taking card details in chat.

What breaks

Failure Usual cause What to do
Reply rejected by the API Window closed while a job was queued Check the window before every free-form send; fall back to a template
Template rejected or paused Wording too promotional for its category, or poor feedback Submit early; keep utility templates purely transactional
Double bookings Webhook retries processed twice Deduplicate on message ID; make writes idempotent
Assistant answers half a question Message bursts Short delay to combine messages before calling the model
Customer missing from your CRM Phone number absent from the webhook Key on the business-scoped user ID
Voice notes ignored Only text was handled Transcribe audio, or say plainly that you cannot

How to start

In the first week, set up a test number on the Cloud API, write the three templates you know you will need (a reminder, a follow-up for after the window closes, a re-engagement message) and submit them for review, since approval is the step you do not control. Pick one use case and one handoff inbox.

The first decision is whether a configurable product, Meta’s or a vendor’s, covers the use case, or whether your tools and data make a custom build worthwhile. If it is the second, that is work our conversational AI team takes on.

Questions

Asked often.

Can I use ChatGPT or another LLM in my WhatsApp business account?

Yes, when the assistant serves your own customers for your own business, such as support, bookings or order updates. Meta's terms bar AI providers from using the platform to offer a general-purpose AI assistant as the main service. You may also retain an AI provider as your solution provider. Read the current terms before you build, because they change.

What is the 24-hour customer service window on WhatsApp?

When a customer messages or calls your business number, a 24-hour timer starts, and it resets each time they contact you again. Inside that window you can send free-form replies, which is where an AI assistant does its work. Once it closes you can send only message templates that Meta has approved in advance.

How much does a WhatsApp AI assistant cost in Meta fees?

Meta charges per delivered message, by category and by the recipient's country. Since 1 October 2026 each business phone number gets 1,000 free service messages a month and pays for the rest, and utility templates sent inside the service window are charged again. Marketing and authentication templates have their own rates. Language model and hosting costs come on top.

Do customers have to opt in before we message them on WhatsApp?

Yes. Meta's messaging policy says you may contact people only if they gave you their number or username and opted in to hear from your business. Replying when a customer writes to you first is fine. Sending a template to someone who never agreed is how numbers get reported, lose their quality rating and are rate-limited.

Not on the list? Ask us directly

Keep reading

More notes from the build.

All blogs

Let’s build your stack.

Send a brief and we’ll reply with next steps.