WhatsApp AI Assistant

A WhatsApp AI assistant looks like a website chat widget, yet it lives under very different constraints. Meta decides when a business may write, your number's quality rating decides how many people you can reach, and marketing law decides what a promotional message may say. Most of the important choices are therefore made in the account and the rules, before any code is written.

The sections below take those choices in the order you will face them: sorting your traffic, setting up the account, the messaging window, architecture, conversation design, integration rights, handover, compliance, the pilot and measurement. Each one ends in a checklist, a criterion or a step you can apply to your own business.

01Sort your message traffic before anything else

The scope of the assistant should come from real conversations, not from what the owner thinks customers ask. Export a few recent weeks of WhatsApp chats, give each conversation a single label, and the pattern appears quickly: which requests repeat and which ones need a human judgment.

We work with four buckets:

• Information: Opening hours, address, delivery times, product dimensions; the answer already exists in writing somewhere.
• Lookup: Free slots, order status, stock; the answer has to be read from a live system.
• Data collection: Name, date, address or measurements gathered step by step for a quote or a booking.
• People work: Complaints, negotiation, personal assessments, and anything that needs a legal or medical opinion.

If the first three buckets are a small share of the total, quick replies in the Business app plus a well written FAQ will serve you better. If the fourth bucket dominates, start with AI that drafts replies for your staff instead of answering customers directly. The same sorting also names your pilot topic: the most frequent request with the lowest cost of a mistake.

02What the assistant does in different businesses

The infrastructure is the same everywhere, but the job changes with the business, and the job decides which system you connect and which questions the assistant asks. The examples below are common need patterns, not client cases.

• Salons and studios: Lets the customer choose a service and a staff member, offers open slots from the calendar and sends an approved reminder the day before.
• Online stores: Reads the shipping stage from an order number, gives sizes from the catalog and collects a photo when a customer requests a return.
• Real estate and car dealers: Asks for the listing reference, shares the key facts, learns budget and timing, then opens a record for sales.
• Wholesale and dealer networks: Takes a repeat order line by line, reads price and stock from the ERP and passes the order to a sales rep for approval.
• Training providers: Explains term dates, placement tests and seat availability, and collects preregistration details.

In healthcare the assistant should stay with scheduling and administrative questions; symptoms and treatment advice are out of scope. If you can describe the assistant's job in one sentence, the scope is right. If you cannot, split the work and start with the most frequent piece.

03Setting up the Business Platform account and number

An official WhatsApp AI assistant runs on Meta's WhatsApp Business Platform, attached to a business portfolio opened in your company's name. In most projects messages flow through the Cloud API that Meta hosts; whether a business solution provider sits in between is a separate decision that affects the inbox and how you are billed.

The account steps we follow:

1. Open the Meta business portfolio under the company's legal name, or reuse an existing one, and give admin rights to at least two people.
2. Complete business verification with registration and tax documents; it also helps your sending capacity grow later.
3. Choose a display name that matches how your brand appears in public and submit it for Meta's review.
4. Decide on the number: a fresh line, or migrating the number customers already know.
5. Add a company payment method and agree in writing who receives the invoices.

Migrating an existing number keeps customer habits intact, but what happens to chat history and to the Business app depends on the options Meta offers for that number. Check this before migration day; it is the only way to avoid risking a line you have used for years. Fill in the profile description, address, hours and website on the same day, since customers read them before they type their first message.

04The 24 hour window, templates and sending limits

On WhatsApp, who may write and when is governed by rules, and every flow the assistant follows is drawn around them. When a customer messages you, a 24 hour customer service window opens; inside it the assistant can send free text, buttons or lists. Once it closes, only templates approved by Meta can be sent.

• Template categories: Meta separates marketing, utility and authentication templates, and rates differ by category and by the recipient's country code.
• Billing: Since 1 July 2025 Meta charges per delivered template; utility templates sent inside an open window are free.
• Chats from ads: When a conversation starts from a Click to WhatsApp ad, a free 72 hour window opens once the business replies.
• Sending limits: Messages outside the window can reach only a limited number of unique people per 24 hours; new portfolios start at 250, and the limit is shared by every number in the portfolio.

The limit rises in steps through verification and high quality template use. What drags quality down is usually unwanted marketing and blocks. Keep messages outside the window to expected ones, such as appointment reminders, during the first weeks. Watch the wording as well: a reminder with a discount line added may be classified as marketing, so keep it to date, time, address and a way to change the booking.

05Architecture in plain words, and choosing a model

A WhatsApp AI assistant has four layers, and each should be replaceable on its own. Meta delivers every incoming message through a webhook, an instant notification sent to your server. An orchestration layer catches it and assembles the chat history, the customer's record and whether the window is open.

• Routing: Decides which bucket the message belongs to; a small, fast model or simple rules are enough here.
• Retrieval: Pulls only the relevant passages from approved content and instructs the model to stay inside them.
• Tools: Calendar, store or CRM calls run through predefined functions; the model never touches those systems directly.
• Delivery and logging: The reply goes out through the Cloud API and is written to the shared inbox and the log at the same time.

Three questions settle the model choice: how good its answers are on your real questions and in your customers' languages, where the provider processes data, and whether its response time suits messaging. WhatsApp users expect a reply within seconds, so slower reasoning models should only handle the hard cases. Design the outage path too: if the provider is unreachable, the assistant sends a short notice and moves the chat into the agent queue instead of going silent.

06Conversation design built for WhatsApp

A good WhatsApp reply is short, covers one topic and makes the next step obvious. Long paragraphs that are acceptable in a web chat go unread on a phone, so the platform's interactive elements do the heavy lifting.

• Reply buttons: A few options such as "Book", "Track my order" or "Talk to a person" remove typing errors.
• List messages: Longer option groups, such as services or branches, are picked with one tap.
• WhatsApp Flows: Multi field details such as name, date and address can be collected in a form screen that opens inside the chat.
• Voice notes: Transcribed first and then handled under the same rules; the assistant confirms what it understood in one sentence.

A concrete flow: in the evening a customer leaves a voice note asking whether there is a free slot tomorrow afternoon. The assistant introduces itself as an AI assistant, transcribes the note, offers two slots as buttons, confirms the name after the choice and creates the booking. The exchange stays under five messages.

Tone follows the way your staff speak on the phone; a law firm and a neighborhood cafe should not greet people the same way. If you serve customers in several languages, the assistant can detect the language and answer in it, but the knowledge base must be approved in that language too, otherwise translation slips reach customers as official information.

07Permission boundaries for calendar, store and CRM

How much the assistant may write into your systems is the technical decision that needs the most care. Read access carries little risk; write access should be justified action by action.

The rules we apply when planning connections:

• Read only first: Orders and stock are read; refunds, cancellations and price changes are never made by the assistant.
• Confirmed writes: Before a booking or preregistration is created, the assistant shows the details in a single message and asks the customer to confirm.
• Duplicate protection: Each action carries a unique ID, so a network delay cannot write the same booking twice.
• Source tag: Every CRM record carries the ad, QR code or link the chat started from.

To keep the source tag accurate, create a separate chat link with its own prefilled message for each channel; the WhatsApp link generator does this in a minute. How your sales team then works those records is an automation topic of its own, covered on our CRM automation page. If the free slots the assistant reads are not kept reliably, you need a proper website with online booking before the assistant can help.

08Agent handover and the shared inbox

Handover is part of the design, not a sign of failure; what matters is that it happens on time and with context. The WhatsApp Business Messaging Policy expects businesses using automation to offer a clear path to a human, and that path should be visible in every conversation.

Write the triggers down from the start:

• The customer asks for a person, or repeats the same question in different words.
• The message mentions a complaint, a refund, a legal dispute or a request for a health judgment.
• The topic has no match in the knowledge base; the assistant does not guess.
• A basket or quote passes a size where your sales team wants to own the first contact.

The handover note needs three lines: what the customer wants, what the assistant has already said, and which details are missing. There is a WhatsApp specific catch: the agent can only send free text while the 24 hour window is still open. For chats handed over out of hours, the assistant tells the customer when to expect a reply, and if the window has closed by morning, the agent reopens the chat with a pre approved utility template.

Every conversation in the shared inbox needs an owner. Queues by role, such as sales, support or billing, make it clear where a handed over chat lands. When an agent takes over, the assistant stays quiet in that chat and resumes only after the agent closes it. Each switch is logged, so you can see weekly how long chats waited and with whom.

09Privacy, marketing consent and AI disclosure

A WhatsApp AI assistant touches three areas of law at once: personal data, electronic marketing and AI transparency. Each needs its own check; passing one does not cover the others.

• Privacy notice: The first message links to a notice explaining why chat data is processed, how long it is kept and who receives it.
• International transfers: If the model provider sits outside the EU or EEA and you serve EU customers, Chapter V of the GDPR requires an adequacy decision or safeguards such as standard contractual clauses, alongside a processing agreement with the provider.
• Opt in: WhatsApp's policy allows contact only with people who gave you their number and agreed to hear from you; store the time and source of each opt in.
• Opt out: A reply such as "stop" must be handled as an opt out request, and no further marketing goes to that number.
• AI disclosure: Article 50 of the EU AI Act requires that people are told when they interact directly with an AI system.

Electronic marketing laws differ by country, so have your adviser confirm whether your consent wording covers promotional WhatsApp messages in each market you serve. The policy also forbids asking for full card numbers, financial account numbers or ID numbers in the chat; when payment is needed, the assistant sends a secure payment link instead.

10Limits and risks: what the assistant must not do

Language models can produce fluent but wrong sentences, known as hallucination, and on WhatsApp that can become a written promise a customer screenshots. Defining what the assistant must not do is as important as defining what it can do.

• Price promises: The assistant quotes only prices stored in your system; no haggling, special discounts or delivery date commitments.
• Prompt injection: A customer may type "ignore your previous rules"; system instructions are kept apart from customer text and tool calls are limited by permission.
• Expert opinions: Legal, medical or financial judgment requests are handed over directly.
• Platform risk: Meta's terms can change; an assistant focused on your own business falls outside the ban on general purpose AI services, but review the terms at least once a year.
• Single channel dependence: Keep a phone line and email ready in case the number is restricted.

Each of these limits gets its own question in the test set, and the assistant should hold the line on all of them before launch.

11From pilot to full rollout, step by step

Opening the assistant to every topic at once hides mistakes; a narrow pilot that grows with the logs catches them while they are small. The order we recommend:

1. Test number: The team tries the assistant on the platform's test number; the live line is not touched at this stage.
2. Hard question set: Your staff write 40 to 60 of the trickiest questions from real chats, each with the expected answer and handover rule.
3. Template approval: Reminder and reopening templates go to Meta for review; rejected texts are simplified and resubmitted.
4. Out of hours pilot: The assistant first goes live only in the evenings and at weekends, for one topic.
5. Daily reading: Sample chats are read every day of the pilot, and every wrong or thin answer becomes a fix in the knowledge base.
6. Gradual widening: Once errors fall to an acceptable level, business hours and a second topic are added.

Each widening step gets a version number with a ready path back to the previous version, so an instruction change that misbehaves can be rolled back within minutes. Setup and upkeep through all of this run under our AI automation services.

12Measuring value: the numbers to track

The value of a WhatsApp AI assistant is measured by the work it closes correctly and the load it takes off your team, not by how many messages it answers. Write the measurement plan before the pilot starts so you can compare against the period before.

• Time to first reply: Compared with the earlier period, especially for messages arriving out of hours.
• Correct closure rate: Samples of chats closed by the assistant are read; a chat that ended on wrong information counts as a failure.
• Conversion to requests: Bookings, preregistrations and quote requests, tracked by source.
• Blocks and opt outs: Tracked weekly after template sends, together with the quality rating.
• Spend by template category: Marketing and utility usage is reported, and unnecessary marketing sends are cut.

For chats coming from ads, the real question is whether the assistant separates serious buyers from browsers, which requires your sales team to give every CRM record a short quality note. Seeing which ad setup brings qualified chats means joining assistant data with your Meta ads management. To see staff time saved, note roughly how many hours went into WhatsApp in one week before and one week after the pilot, so the effect is discussed with records instead of impressions.

13Common mistakes with WhatsApp assistants

Most mistakes made with a WhatsApp AI assistant are rule and process mistakes rather than software bugs, and each has a better alternative.

• Using unofficial connectors: Hooking a personal number to a bot through browser automation puts the line at risk of a ban; use the official Business Platform with an account in your company's name.
• Blasting an old customer list: Marketing templates sent without opt in records hurt both quality rating and legal footing; collect consent first and start with small, expected sends.
• Hiding that it is a bot: An assistant that pretends to be human erodes trust; say in the first message that it is an AI assistant and how to reach a person.
• Writing the knowledge base once: When prices or offers change, the assistant repeats stale facts; name an owner for each content area and a fixed update day.
• Handing over without the window in mind: An agent who replies the next day cannot send free text; prepare an approved reopening template for out of hours handovers.
• Long, multi topic replies: Paragraphs nobody reads on a phone lead to repeated questions; keep replies to one topic with a button for the next step.

14Choosing a partner and the next step

The right partner is not the one who builds fastest, but the one who leaves your number and accounts with you and explains the rules up front. Questions worth asking in the first call:

• In whose name are the Meta business portfolio, the WhatsApp account and the model provider opened, and what happens if we part ways?
• Which option applies to chat history and the Business app when our current number is migrated?
• How do the test set, pilot scope and version rollback work?
• Where are chat logs stored, for how long, and who can access them?
• Who is billed for Meta messaging fees and model usage?

If the assistant should also run in your website chat, we explain how one knowledge base serves both on our AI chatbot development page. We cannot yet show a live client WhatsApp assistant as a reference; our automation, software and web work is listed among our reference projects. Discovery and first workflow options are in the AI automation pricing section. Share a few recent chats and the calendar or order system you use through the contact form, and we will define the scope of your WhatsApp AI assistant and send a written quote.