- On WhatsApp the cost of an AI reply depends almost entirely on who spoke first, which means the bot's design and the bill are the same decision.
- Since 1 July 2025 Meta charges per delivered template message, with the rate set by the template category and the country code of the recipient's number.
- Ordinary replies that are not templates have cost nothing since 1 November 2024, but they can only be sent inside an open customer service window.
- That window lasts 24 hours and it reopens every time the customer sends another message, so a conversation that keeps moving stays free to answer.
- From 1 October 2026 every business phone number gets 1,000 free service messages a month, with charging starting at the 1,001st and no rollover of unused ones.
- Since 2 August 2026 the EU requires people to be told they are dealing with an AI system from the first interaction, unless it is genuinely obvious.
A shop owner asking whether to put AI on WhatsApp is usually asking two questions at once. Can it answer customers well enough, and what will it cost. The second question has a much more precise answer than most people expect, because Meta publishes the pricing mechanics in detail, and those mechanics should determine how the bot is built.
The short version: answering is cheap, initiating is not. Every design decision follows from that one asymmetry, and a shop that builds its assistant without knowing it will produce a bill that looks arbitrary and is not.
How does WhatsApp actually charge a business?
Per delivered template message, since 1 July 2025. Meta's pricing documentation for the WhatsApp Business Platform states that you are charged only when a template message is delivered, and that the rate varies by the template's category and by the country calling code of the recipient's WhatsApp number. This replaced an older model that billed per conversation.
Two carve outs matter more than the rates themselves. Since 1 November 2024, Meta does not charge for non template messages, the ordinary text and image replies a conversation is made of. Those can only be sent inside an open customer service window. And since 1 July 2025, utility templates sent in response to a user inside that open window are also free.
The window is the whole game. It runs for 24 hours, and Meta's documentation is explicit that it opens and resets with each user message. A customer writes at nine in the morning, and for the next 24 hours you can answer them as many times as the conversation needs at no messaging cost. If they reply at four in the afternoon, the clock restarts.
There is also a free entry point window, during which all messages including templates are free for 72 hours. It opens when a customer reaches you through a qualifying entry point, an advert that clicks to WhatsApp being the common one for a small shop.
The country code detail deserves more attention than it gets from anyone selling a chatbot. Because rates are set by the recipient's country calling code, the same campaign costs different amounts depending on where your customers live, and a shop with buyers spread across several markets cannot reason about cost per message as a single figure. If most of your list sits in one country, look up that country. If it does not, the number you need is a weighted average, and nobody will calculate it for you.
Volume changes it too. Meta's documentation notes that lower rates for utility and authentication templates can be unlocked based on messaging volume, which is the usual shape of platform pricing: the businesses least able to absorb the cost pay the most per message. Worth knowing before a vendor quotes you a rate they pay at their scale rather than one you will pay at yours.
What are the four template categories?
Marketing, utility, authentication and service. The category is not a label you choose freely, it follows from what the message says, and it sets the price.
| Category | What it carries | When it costs | Sensible AI role |
|---|---|---|---|
| Service | Ordinary conversation replies | Only inside an open 24 hour window, free tier arriving 1 October 2026 | Almost all of it, this is where a bot belongs |
| Utility | Order confirmations, dispatch and appointment updates | Free in response to a user inside an open window, charged otherwise | Trigger from your systems, never from a generated guess |
| Authentication | Codes and login verification | Charged per delivery, volume tiers available | None, this should never touch a language model |
| Marketing | Offers, re engagement, anything promotional | Charged per delivery at the highest rates | Drafting, with a human approving the send |
Read that table as an instruction about where to point the AI. The cheapest category is the conversational one, and it is also the one where a language model is genuinely useful. The expensive categories are the scripted ones, where a model adds cost and risk without adding much.
An assistant that generates an order status is a different thing from one that reads an order status out of your database and puts it into a sentence. The first will eventually invent a tracking number. Templates for utility messages exist precisely so the facts come from your system.
What changes on 1 October 2026?
Service messages get a free monthly allowance. Meta's documentation states that from that date every business phone number receives 1,000 free service messages per month, and charging begins with the 1,001st delivered in that month. The allowance resets monthly and unused messages do not roll over.
For most independent shops that number is larger than their entire monthly volume, which makes conversational support on WhatsApp close to free in practice. For a busier shop it sets a clear planning figure, and it rewards exactly the behaviour that produces good service: answering inside the window rather than pushing people to email.
It also changes the arithmetic of automation. If service messages are effectively free up to a thousand a month, the reason to automate replies stops being cost per message and becomes response time and your own hours. That is a healthier reason, because it points the tool at the problem it is actually good at.
Where does an AI reply earn its place?
In the gap between a question arriving and a human being available. For a shop run by one or two people, that gap is most of the week, and it is where customers quietly give up.
The questions that repeat are the ones worth automating, and they are always the same shape: where is my order, do you ship to X, is this in stock in size Y, what is your returns window, are you open on Sunday. All of them have exact answers that live in a system you already own. None of them require creativity, and all of them are answerable inside the free window because the customer opened it by asking.
The questions to escalate are equally consistent. Anything about a refund amount. Anything where the customer is unhappy rather than curious. Anything involving an exception to a policy. We worked through that triage in which support tickets an AI should handle before any others, and the WhatsApp version differs only in that the medium feels more personal, so a wrong answer lands harder.
Do you have to tell people they are talking to a bot?
In the European Union, yes, since 2 August 2026. The Commission's FAQ on the transparency obligations in Article 50 states that providers must design systems interacting directly with people, naming chatbots and AI agents specifically, so that those people are informed they are dealing with an AI. The notification has to come at the start of the first interaction, in a clear and distinguishable way, and in line with accessibility requirements.
There is an exception where the AI involvement is obvious, and the FAQ is careful about it. The obligation applies on cumulative criteria: the thing has to qualify as an AI system, it has to be designed for a genuine two way exchange rather than merely collecting data or firing automated responses, and the interaction has to be direct. A menu of numbered options is not a chatbot in this sense. A system that converses is.
Practically this is one line at the top of the first reply, and it costs nothing because it rides inside a message you were sending anyway. The version that works commercially names the machine and offers the exit in the same breath, something close to: this is an automated assistant, reply with the word human at any point. Customers tolerate a bot that is honest about being one far better than they tolerate discovering it. The platform expects that exit to exist too: the messaging policy's list of acceptable escalation paths opens with in chat transfer to a human agent, alongside a phone number, an email address, web support and a support form. The broader set of rules across jurisdictions is in when a shop has to declare that a customer is talking to AI.
What does the messaging policy require before you send anything?
Permission, and a specific kind of it. The WhatsApp Business Messaging Policy sets two conditions that both have to hold before you may contact someone: they gave you their mobile number, and you received opt in permission confirming that they wish to receive subsequent messages or calls from you. Having somebody's number because they ordered from you is condition one only.
The policy also puts the method squarely on the business. You are solely responsible for determining how opt in is collected, for making sure that method complies with the laws that apply to your communications, and for giving whatever notices those laws require. There is no template handed down, which means a checkbox at checkout that says you agree to be contacted on WhatsApp, stored with a timestamp, is the minimum workable record.
Coming off the list is broader than most shops implement. The policy requires respecting all requests to block, discontinue or opt out of your messages, made on or off WhatsApp, and says that includes removing the person from your contacts list. A customer who tells you in an email to stop messaging them on WhatsApp has opted out, and a system that only listens for a keyword inside the chat will miss it.
Two data rules are worth writing into whatever tool you use. Information obtained about a person through WhatsApp, other than the content of the message thread itself, may not be used for anything beyond supporting messaging with that person. And information from one customer's chat may not be forwarded or shared with another customer. An assistant that quotes one buyer's message to reassure another has broken the second rule while being helpful.
What can go wrong that costs real money?
Four things, in rising order of expense.
The one that is easiest to miss is throttling. The policy states that people can block or report businesses, and that if a business's quality tier stays low for a sustained period the platform will limit how many messages it can send or how many calls it can initiate. That is a real operational risk from an assistant that annoys people: the punishment is not a fine, it is a reduced ability to reach the customers who do want to hear from you. And it arrives slowly enough that the cause is hard to spot.
The first is a bot that reopens conversations to sell. A marketing template is charged at the highest rate, and an assistant that decides on its own to re engage a quiet customer is spending money per person with no human looking at the list. Sending decisions belong to a person or to a rule, never to a model's judgement about whether someone seems interested.
The second is the promise your assistant makes. A message from your business number carries your business's word, and a bot that says a refund will be processed has committed you in the customer's mind whatever your policy says. That exposure is the subject of what happens when your chatbot promises something you did not, and WhatsApp raises the stakes because the transcript sits in the customer's pocket permanently.
The third is quieter and more common. A bot that answers everything badly trains customers to stop asking, and the cost shows up as orders that did not happen rather than as a line on an invoice. Nothing in your dashboard will attribute it. The only defence is reading real transcripts weekly, which takes twenty minutes and is the single highest value habit in this whole area.
How should a small shop actually start?
Narrow, and inside the window. Pick the three questions you answer most often, let the assistant handle only those, and have it hand over on anything else with a sentence that does not pretend. Measure two things for a month: how many conversations it closed without a human, and how many customers asked for a human immediately.
Keep every initiated message manual at first. Order updates can be automated because they are triggered by real events in your system, but anything promotional stays a decision a person makes, both because of the cost and because the tone of an unsolicited WhatsApp message is unforgiving.
Watch the number of conversations rather than the number of messages. Because replies inside the window are free and the window resets on every customer message, a long helpful exchange costs the same as a short unhelpful one. That is an unusual incentive structure and it is worth exploiting: there is no financial reason to rush a customer off the channel.
One structural warning. Whatever tool you use sits between you and your customers, and the conversation history it holds is a business asset. Check what happens to those transcripts if you leave, whether you can export them, and whether the vendor uses them to train anything. The same question applies to every part of a shop's stack, which is why we publish what our own platform costs and what it keeps on the plans and credits page rather than leaving it to be discovered later.
Is it worth it at all?
For a shop whose customers already message, clearly yes, and the reason is not cost. It is that the reply arrives in four seconds at eleven at night, when the alternative was a form that gets answered on Tuesday.
For a shop whose customers do not message, adding the channel is a bigger project than adding the bot, and the honest sequence is to open the channel manually first and see whether anyone uses it. A quiet WhatsApp number with an AI on it is still a quiet WhatsApp number.
The pricing rewards the right behaviour for once. Let people come to you, answer them properly while the window is open, and initiate rarely and deliberately. That is also how you would want to be treated as a customer, which is not always how these things work out.