Short answer. A WhatsApp AI agent that works has five layers: the channel, Meta’s WhatsApp Business Platform with a number in the company’s name, a webhook that delivers messages, the 24-hour window and approved message templates; understanding, which transcribes the voice note, detects the language and outputs a structured intent; knowledge, a base of validated answers and rules reviewed by the manager; tools, the back-office API the agent reads before answering and in which it acts within limits; supervision, a human who takes over when the agent doubts and reviews a sample every week. Without the fourth layer it is an answering machine; without the fifth it is a risk. Twenty weeks, two of them spent reading a month of real conversations before writing a line.
This guide is for the wedding hall whose manager answers at midnight, the wholesaler who receives orders as voice notes, the practice that triages summer patients, the real-estate agency, the carrier, the school, and the founder settled in France who wants an agent for Algeria and reads “WhatsApp chatbot” in every quote. It does not sell a use case; the guide on AI agents in companies already ranks the uses and autonomy levels. It describes how the agent is built, what Meta imposes and bills, and what it must never do.
What the figures say
The channel is already in the customer’s pocket. According to DataReportal’s Digital 2026 report, dated 8 November 2025 on October 2025 data, Algeria has 37.8 million internet users, 79.5 % of the population, and 27.5 million social media identities, with an advertising audience of 27.5 million on Facebook and 13.5 million on Instagram; the report publishes no WhatsApp figure, since WhatsApp does not sell advertising the same way, but it is the app through which your customers already write to you, and 51.63 million 4G and 5G subscriptions in the second quarter of 2026 according to the ARPCE put it in every hand. An agent on WhatsApp has nothing to get installed; it answers where the question is asked.
The platform has rules and a price, published by Meta. Its pricing page, in the version in force on 1 October 2026, says that billing is per delivered message since 1 July 2025, in four categories, marketing, utility, authentication and service, that service messages and utility messages sent within the 24-hour window are charged since 1 October 2026, and that “Meta provides 1,000 free service messages per month, which do not roll over”. A customer’s message opens a 24-hour window, reset by each message; outside it, the company can only write with an approved message template. A conversation opened from a “click to WhatsApp” ad gives a free window of up to seven days. Rates depend on the recipient’s country and are published in separate rate cards; Algeria sits in the “Rest of Africa” group, and we quote no amount, because the card changes and a reseller is not Meta. Meta also offers its own agent, billed at “$2.00 per 1 million tokens”, which gives the order of magnitude of a language model’s cost per conversation: cents.
The technology, finally, is documented and asks for nothing exotic: the WhatsApp Cloud API receives and sends messages over HTTPS, delivers incoming messages and receipts by webhook, in JSON, to a company server, and accepts 80 messages per second per number by default. Everything that gives the agent its value is therefore on the company’s side: what it reads, what it is allowed to do, and who takes over.
The life of a WhatsApp message, in six moments
The architecture is not described by its components but by what happens to a message sent at 11:40 pm, until the customer has their answer or a human has taken over.
The message arrives by webhook on the company’s server, with the number, the text or the audio file, and the time; the 24-hour window opens. Identification matches the number to a customer, a file or an order in the back office; an unknown number becomes a new contact, and the agent says so. Understanding transcribes the voice note, detects the language, Arabic, French or Darija, often mixed in the same sentence, and outputs a structure: “availability request, 14 August, about 200 guests”; that is the language model’s job, and the only moment it speaks for itself.
The grounded answer is the moment that separates an agent from a chatbot: before answering, the agent reads the hall’s calendar, the catalogue at this customer’s price, the practice’s diary, the retailer’s balance, through the back-office API, and answers with what it read; if the tool does not respond, it says so and hands over, it does not invent a date or a price. Action within limits follows: hold a date for 72 hours, prepare an order the retailer will confirm with one word, open a ticket with the photos received, send a receipt; never confirm without a deposit, never validate above a ceiling, never change a price. Handover and trace close the loop: after two misunderstandings, an angry word, an off-rule price request or an out-of-scope question, a human takes over in the same conversation with a three-line summary, and every exchange, every read and every action are logged on the company’s server. It is the mechanism the guide on the B2B ordering portal describes for a voice note turned into an order: the agent prepares, the customer confirms, the depot applies.
The architecture of a WhatsApp agent, in five layers
Behind the six moments, five layers, two of which are provided by Meta and the model, and three of which belong to the company.
The channel is Meta’s platform: an account in the company’s name, never the vendor’s; a dedicated number; message templates submitted and approved for everything that goes out outside the window, appointment reminder, due date, missing document; a webhook on a server that answers fast, because Meta retries and eventually gives up. Understanding is the language model, called through an API: it transcribes, detects the language, extracts the intent and elements, and phrases the answer from what the tools layer gave it; it does not know the company by itself, and that is deliberate. Knowledge is a base of validated answers, hours, conditions, frequent questions, and rules, what is said, what is not, versioned and reviewed by the manager; when a question comes back ten times and the agent hands over ten times, that is where the answer gets written.
Tools are the layer that creates the value and that most “chatbots” lack: a back-office API the agent calls to read, calendar, stock, balance, file, status, and to act within limits, hold, prepare, open, send a receipt, every action with a ceiling and a trace. If the back office is a notebook or an Excel file, the agent has nothing to read; it is the prerequisite the guide on AI agents in companies calls “something to read and something to do”. Supervision is the layer that creates trust: a queue of conversations handed to a human, a sample of twenty conversations reviewed every week by the manager, the resolution rate, the cost per conversation. The company’s three layers run on its server, with its data; the model is called through an API, and the guide on paying for cloud and AI from Algeria says how it is paid.
«Without a back office to read, an agent is just an answering machine; without a human who reviews, it is a risk.»
A worked example
An example on volume assumptions and facts that are not assumptions: a mid-sized company, wedding hall, wholesaler or clinic, receiving two thousand WhatsApp conversations a month.
Two thousand conversations a month is our assumption; count yours by rereading a month of WhatsApp, that is the project’s first week. 70 % resolved without a human by the third month is also an assumption, cautious for availability, listed prices, status and documents to send, optimistic for negotiations; the remaining 30 % go to a human with a summary, and that is a success, not a failure: the human receives a conversation already identified, transcribed and summarised. Under a minute of first response at any hour is not an assumption: it is what a webhook that answers does, against several hours when the manager’s phone is the only channel and the message arrives at 11:40 pm.
A thousand free service messages per month per number, under Meta’s pricing in force on 1 October 2026, are not an assumption either; beyond that, every delivered message is billed by country, and Algeria sits in the “Rest of Africa” group whose rate we do not quote. What this changes: a conversation of six company messages costs six messages, and the monthly cost is computed in delivered messages, not conversations; reminders sent by template outside the window have their own category and price; the language model adds cents per conversation, the order of magnitude Meta shows for its own agent. The assumptions to check at home: the volume, the share of voice notes, the share of questions that have an answer in the back office, and the number of reminders a month; a month of WhatsApp reread answers all four.
What the agent does, and what it never does
The right column is written before the first line of code, with the manager, and it is what turns a risk into a tool.
Answer, act, follow up are the three things the agent does, and each has its limit: it answers with what it read, never from memory; it acts by holding, preparing and opening, never by confirming without a deposit, validating above a ceiling or changing a price; it follows up by approved template outside the window, never with marketing without consent. Understand has its limit too: it transcribes and answers in the customer’s language, but it does not diagnose, does not advise on health, law or money; for a practice it triages and asks for documents, as the guide on the dental practice describes, and the practitioner speaks.
Hand over is the row customers notice: after two misunderstandings, an angry word, an off-rule price or an out-of-scope request, the agent says “I’ll get you someone” and does it, with a summary; a looping menu loses the customer, an agent that hands over keeps them. Keep the trace is the row the manager demands: every exchange, every back-office read and every action are logged on the company’s server, and conversations are not stored with a third party the company does not control. A seventh row is not in the table because it belongs to the manager: review the sample every week; an agent nobody reviews drifts within a quarter.
«An agent that says “I’ll get you someone” keeps the customer; a looping menu loses them.»
Six companies, what the agent reads and what it does
Six companies we often equip, and for each what the agent reads in the back office and what it does.
The wedding hall is the case that gave this guide its title: “is 14 August free?” at 11:40 pm; the agent reads the date calendar described in the guide on the Mariages.net equivalent, answers, offers two nearby dates, holds for 72 hours and sends the quote; it does not confirm without a deposit. The wholesaler receives “send me the same as last week plus two cases of oil” as a voice note; the agent reads the typical basket and the catalogue at the customer’s price, prepares the order, recalls the cut-off and warns if the ceiling is reached; the retailer confirms with one word. The practice receives “I’m coming in August, I have pain lower right”; the agent asks for the X-ray and dates, offers a slot, reminds the day before, and never diagnoses.
The real-estate agency receives “I paid, send me the receipt” or “the boiler is broken”; the agent reads the lease and schedule described in the guide on the flat managed from France, sends the receipt, states the covered period, opens a ticket with the photos and informs the owner. The carrier receives “where is my parcel?” a hundred times a day; the agent reads the real status and estimated time, offers a reschedule, and hands over if the delay is abnormal. The school receives certificate requests and timetable questions; the agent reads enrolments and payments, sends, reminds, and never argues a grade. In all six cases the left column must exist before the agent; when it does not, the project starts with the back office, and the guide on the specification says how to write it.
What Meta imposes, and what it changes in the design
Three platform rules shape the agent. The 24-hour window: as long as the customer has written in the last twenty-four hours, the company may answer freely; beyond that, only with a template approved by Meta, in a category, utility for a reminder or a receipt, marketing for an offer, authentication for a code. The agent must therefore know, for every message it wants to send, whether the window is open, and choose between a free answer and a template; and the company must have had its templates approved before the pilot, which takes days, not hours. Billing per delivered message: since 1 October 2026, service messages within the window are charged beyond the thousand free per month; a chatty agent answering in five bubbles costs five messages, and the design prefers one complete answer to five short ones. The account in the company’s name: the number, the Meta account, the approved templates and the history belong to whoever opened the account; an agent deployed on a vendor’s account is an agent whose key the company does not hold, as the guide on the app stuck with a vanished developer recalls for the stores.
Two common-sense rules are added. Voice is the norm, not the exception: a large share of messages arrive as audio, in Darija, and transcription is the first thing to test on real messages, not clean sentences. The webhook must answer fast and lose nothing: Meta delivers the message once and waits for an acknowledgement; the server acknowledges immediately, queues the message and processes it afterwards, with the same queue and idempotency logic the guide on the offline app describes for the phone: a message delivered twice must produce one answer.
Where the model runs, and what a conversation costs
The agent has two halves that do not run in the same place. The company’s three layers, knowledge, tools and supervision, run on its server, in Algeria or wherever its back office and data already run; conversations, logs and the customer file stay there. The language model is called through an API at a provider, for transcription, understanding and phrasing, and receives only what it needs for the question asked: the message, the context read from the back office, the rules; not the whole base. That split is what lets the company change model without changing anything else, and what makes the model’s cost countable per conversation.
A conversation’s cost has three lines. Delivered messages, billed by Meta by category and country beyond the thousand free, whose amount is read on the rate card in force and changes. The model, a few cents of a dollar per conversation for transcription and answer depending on the model chosen, the order of magnitude Meta shows for its own agent at two dollars per million tokens; a ten-exchange conversation with the context read from the back office typically consumes a few thousand tokens. The server and supervision, which do not depend on volume: hosting the webhook and the back office, and the manager’s weekly hour reviewing the sample. What a year of this whole then costs, updates included, follows the logic of the guide on keeping an app alive; paying for the model from Algeria is in the dedicated guide.
The 2027 calendar
The calendar is that of a twenty-week rollout, whose first two weeks produce no code and decide everything.
Weeks 0 to 2 reread a month of the company’s real WhatsApp conversations: the fifty recurring questions, the right answer to each, those with an answer in the back office and those without, and the column of what the agent never does, written with the manager. Weeks 2 to 5 open the channel: Meta account in the company’s name, number, message templates submitted and approved, receiving server that acknowledges and queues. Weeks 5 to 9 build understanding and the knowledge base, tested on the real voice notes of the reread month. Weeks 9 to 12 wire the tools, read-only first, then limited actions with their ceilings. Weeks 12 to 14 are acceptance: the two hundred real conversations of the inventory month are replayed, and every answer is graded right, acceptable or wrong; a wrong answer on an availability or a price blocks the pilot. Weeks 14 to 18 are the pilot with a human in the loop, live the first week, by sample afterwards. Weeks 18 to 20 review: resolution rate, handovers and their causes, cost per conversation, and the decision on the actions to open next, one autonomy level at a time, as the guide on AI agents in companies recommends.
What makes a WhatsApp agent fail
The WhatsApp agents we see fail almost always fail for one of these reasons.
- No back office to read. The agent answers generalities, the customer asks the manager again, and the agent is switched off in the second month.
- Answering from memory when the tool does not respond. One invented date or price, and the manager no longer trusts anything.
- No “never does” column. The agent confirms a date without a deposit or promises a price; the manager discovers the scope in a complaint.
- No handover. The customer loops in a menu and writes “I want to talk to someone” three times; they call the competitor.
- Templates not approved before the pilot. Reminders outside the window do not go out; you learn it at the first missed appointment.
- The account in the vendor’s name. The number, templates and history are theirs; changing vendor means changing number.
- Nobody reviews. Without the weekly sample, the agent drifts with new questions and nobody sees it before the customer does.
Frequently asked questions
The questions we receive from managers, clinics, wholesalers and founders settled in France.
- What is the difference between a WhatsApp chatbot and an AI agent?
- The chatbot answers with menus and learned answers; the agent reads the company’s back office at the moment of the question, calendar, stock, balance, file, answers with what it read, acts within written limits and hands over when in doubt. The difference is not the language model; it is the tools layer and the column of what it never does.
- Does the agent understand Darija and voice notes?
- Yes, and it is the first thing to test on real messages: the transcription of a Darija voice note mixed with French, then the extraction of the intent. Current models do it well on common requests and less well on local product names; the knowledge base and acceptance tests fix that, with a handover when confidence is low.
- What does a WhatsApp agent cost per month?
- Three lines: delivered messages billed by Meta by category and country beyond the thousand free service messages per month per number, whose rate is read on the card in force; the language model, a few cents of a dollar per conversation depending on the model; the server and the weekly review hour, which do not depend on volume. Count a month of conversations and reminders before pricing.
- Can it message my customers on its own?
- Only through a message template approved by Meta, in a category, and only for what the “does” column allows: appointment reminders, due dates, missing documents, receipts. No marketing without the customer’s consent, and no message outside the window without an approved template.
- Where do conversations and data go?
- To the company’s server, with the back office and the logs. The language model receives through the API what it needs for the question asked, the message and the context read, not the whole base; the WhatsApp account, the number and the history are in the company’s name, never the vendor’s.
- How long before the first automatic answer?
- Twenty weeks for an agent that reads the back office and acts within limits: two weeks of inventory, three for the channel and approved templates, four for understanding, three for tools, two of acceptance on real conversations, four of pilot with a human, two of review. An agent answering only on hours and conditions can be in pilot by week nine.