What a WhatsApp AI assistant actually costs to run for a year

Every quote we send for a WhatsApp assistant gets the same question back: fine, but what does it cost me every month after you leave? It is the right question, and most vendors answer it badly — either with one number that hides everything, or with a “starting from” price that only covers their own software.

So here is the full stack, priced, for a small Kuwaiti business. And a warning: the WhatsApp part of this is about to change on 1 October 2026.

The four things you actually pay for

A WhatsApp assistant is not one product. You are paying four separate suppliers, and only one of them is us:

Meta, per message, for the WhatsApp Business Platform.
An AI provider, per token, for the model that writes the replies.
A BSP or hosting, monthly, for the pipes in between.
Whoever maintains it when your price list changes and the assistant starts quoting last year’s prices.

People budget for the first two and forget the last two. The last two are usually the bigger number.

1. What Meta charges

Meta moved from per-conversation to per-message pricing on 1 July 2025. You pay per delivered template message, priced by the recipient’s country and by category. For Kuwait, the rates published by Business Solution Providers at the time of writing are:

Marketing (promotions, offers)	$0.0375 per message
Utility (order updates, receipts, reminders)	$0.0100 per message
Authentication (OTP codes)	$0.0100 per message

Saudi Arabia and the UAE are meaningfully more expensive for marketing — around $0.055 per message in both — so a Gulf-wide campaign does not cost what a Kuwait-only campaign costs.

Until now, the replies your assistant sends inside the 24-hour customer service window have been free. That has been true since November 2024, and it is the single reason support assistants have been cheap to run: the customer messages you first, and everything you say back for the next 24 hours costs nothing.

This is the part that changes on 1 October

Meta is reportedly ending that. From 1 October 2026, each business phone number gets 1,000 free service messages per month, and everything beyond that is charged at the destination country’s utility rate. Utility templates sent inside the 24-hour window also become billable.

Every major BSP has published the same summary. We have not been able to confirm the exact figures against a page on Meta’s own domain, so treat the 1,000-message allowance as reliable-but-unconfirmed and check Meta’s pricing calculator before you sign anything. What is not in doubt is the direction: the free lunch inside the service window is ending.

For a business handling 3,000 assistant replies a month, that is roughly 2,000 chargeable messages at $0.01 — about $20 a month where it used to be zero. Not a catastrophe. But if you were sold an assistant on the basis that “the conversations are free,” your quote is now wrong.

2. What the AI model charges

This is the part everyone worries about and it is almost always the smallest line on the bill.

Support assistants do not need a frontier model. They need something that can read a paragraph of Arabic, check it against your price list or FAQ, and write four sentences back. The cheap tiers from OpenAI, Anthropic and Google all handle this, and they currently sit somewhere between $0.05 and $0.30 per million input tokens, and $0.40 to $2.50 per million output tokens.

A realistic support exchange — customer question, your retrieved product data, a short answer — runs a few thousand tokens end to end. At those rates you are into fractions of a fil per exchange. Even at high volume, most of our clients’ monthly model spend lands in the tens of dollars, not the hundreds.

The exception is when someone insists on using a large reasoning model for a job a small one does fine, or when the assistant is re-sent the entire product catalogue on every single message because nobody set up retrieval properly. Both are architecture mistakes, not pricing problems.

3. The pipes

You need something between WhatsApp and the model. Three routes:

Meta Cloud API direct — no platform fee, you pay only Meta’s message charges, but someone has to build and host the integration.
A BSP — 360dialog publishes tiers from €49/month per number; Twilio charges no monthly fee but adds roughly half a cent of markup per message.
Your own server — a small VPS from $4–6/month is genuinely enough for a support assistant. This is not a workload that needs scale.

The honest answer for most Kuwaiti SMEs is Cloud API direct on a $6 droplet. The BSP fee buys you a dashboard and a support contract, which is worth it for some teams and pure overhead for others.

4. The part nobody budgets for

Your prices change. Your opening hours change during Ramadan. A supplier stops carrying a product and the assistant keeps confidently recommending it. Someone asks a question in Kuwaiti dialect that the assistant mishandles in a way that embarrasses you in a screenshot.

An assistant left untouched for a year does not stay accurate — it decays quietly, and you find out from a customer. Budget for someone to own it: reviewing flagged conversations, updating the knowledge it draws on, and adjusting where it hands over to a human. This is the line that separates a system that pays for itself from one that quietly costs you customers.

So, the year

For a small Kuwaiti business running a genuine support line — not a marketing blast — the infrastructure is cheap: message fees in the low tens of dollars a month after October, model costs in the same range, hosting under ten. The software is not what costs money.

What costs money is the maintenance, and what makes it worth paying is that it replaces hours of someone re-typing the same six answers. That is the calculation to run — not the token price.

Sources

Meta, WhatsApp Business Platform pricing documentation
Per-country rates as published by Plivo
October 2026 service-message changes as reported by BSPs including Wati and YCloud
Model pricing: OpenAI, Anthropic and Google published API pricing pages

Prices checked 9 September 2026. Rates move — Meta revises them quarterly.