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

The question never comes first. It comes after the demo, once someone has watched an assistant read an enquiry, check availability and answer in Arabic without a human touching it. Then the room goes quiet and somebody asks the real question: what does this cost me every month?

It is the right question, and it is usually answered badly — often with numbers that were correct two years ago. WhatsApp changed how it charges in July 2025, and a great deal of the advice still circulating describes a pricing model that no longer exists.

Here is how the cost actually works now, why the surprising part is rarely the AI, and what to ask before you sign anything.

First, the model that changed

Until mid-2025, WhatsApp billed by conversation: a 24-hour window, one charge, however many messages passed through it. Most cost estimates you will find online still assume this.

Since 1 July 2025 the platform charges per message, by category. That is not a small adjustment — it changes which parts of an assistant cost money and which are free, and therefore changes the entire shape of the bill. Meta’s own pricing documentation is the source of record, and it is revised more often than most articles about it.

The four message categories

Everything your assistant sends falls into one of four buckets, and they are priced very differently:

• Service messages — ordinary, non-template replies sent inside the customer service window. Free since November 2024.
• Utility templates — order updates, delivery notices, appointment reminders. Free inside the service window, charged outside it.
• Authentication templates — one-time passcodes and verification. Free inside the window, charged outside.
• Marketing templates — promotions and campaigns. Always charged, and they cannot be sent inside a service window at all.

The customer service window opens when a customer messages you and stays open for 24 hours from their last message. Inside it, your assistant’s replies cost nothing in platform fees.

What this means for a support assistant

Read those rules together and a conclusion falls out that surprises most people: an assistant that only answers inbound customer questions has no per-message platform cost at all. A customer messages you, the window opens, and every reply your assistant sends inside it — clarifying questions, product details, availability — is a free service message.

This is why the old advice misleads: budget a support assistant on conversation rates and you over-estimate platform cost dramatically, while missing where the money actually goes.

So where does the money go?

Four places, in roughly this order of size for a support-focused assistant:

• Your provider’s platform fee — the monthly charge from whoever gives you official API access. For a support-only assistant this is often the largest fixed line.
• The language model — billed by text processed, in both directions.
• Hosting — server, database, monitoring.
• Maintenance — the human hours that keep the answers correct.

Notice what is not on that list: per-message charges, as long as you stay inside the service window. Add outbound messaging and the picture changes completely — which is the next section, and the one that produces surprise invoices.

Where the bills actually appear: outbound

The moment you initiate contact, you leave the free zone.

A marketing template is always charged, in every market, with no free window. A single campaign to your whole customer list can cost more than a year of answering support questions, because support answers are free and campaigns are not.

Utility templates — “your order shipped”, “your appointment is tomorrow” — are free inside an open window and charged outside it. Most land outside, because you are messaging someone who has not written in 24 hours. That cost scales with order volume, not support volume.

The honest budgeting rule: model inbound support at close to zero platform cost, and model every outbound message by category at the current rate for your market.

Rates move, and they move by market

Rates are set per country and revised regularly — October 2025 brought increases in some markets and decreases in others. Utility and authentication messages carry volume tiers, so the rate falls as monthly volume rises, resetting each month.

A currency and billing shift also runs through 2026, with several regional currencies including AED and SAR arriving from April 2026.

We deliberately do not print rate tables: they vary by market, they change, and a stale number is worse than none. Check the current rate for the country your customers message from, on the day you build the model.

The language model is cheaper than you think

For most customer-service work, the AI is the smallest line on the bill. A short question and a short answer consume a fraction of a fils. Even at thousands of conversations a month, model cost on a well-built assistant is usually modest.

What inflates it is how much context gets sent with each request. Push your entire catalogue and the full chat history into every call and you multiply the cheapest line item until it is no longer cheap. The fix is not a cheaper model — it is sending only what the question needs.

This is where engineering discipline shows up directly on an invoice. Retrieval that fetches the three relevant paragraphs instead of the whole manual is the difference between a model bill you ignore and one you argue about. It is the same discipline we apply when building custom AI assistants generally, not just on WhatsApp.

The line item nobody quotes

Maintenance never appears in the proposal, and it decides whether the project is still worth anything a year from now.

Your prices, products and policies change, and customers ask things nobody scripted for. An assistant not reviewed monthly degrades — within a quarter it is a confident source of outdated answers, which is worse than having none, because customers believe it.

Budget for it explicitly: a few hours a month of someone reading real transcripts, correcting wrong answers and updating source material. A proposal with no maintenance line is not cheaper. It is incomplete.

A worked example

Take a retailer receiving roughly a thousand customer enquiries a month — availability, price, delivery timing — and sending order-status updates to about four hundred customers.

The inbound thousand cost nothing per message — service messages inside open windows. The four hundred order updates are utility templates, most landing outside a window, so charged at the utility rate with volume tiering applied.

Against that sit the fixed lines: platform fee, model usage, a modest server, and maintenance. In that shape of business the platform fee and maintenance usually dominate, templates are visible but manageable, and the model is smallest.

Now change one assumption — the retailer starts sending weekly promotions to fifteen thousand customers. Those are marketing templates, always charged, no free window. That single decision can outweigh everything else on the bill combined. Same assistant, same support traffic, entirely different invoice.

Four mistakes we see repeatedly

Budgeting on conversation rates — the model retired in 2025, producing estimates unrelated to a current invoice. Treating inbound and outbound as one number — they follow completely different rules. Sending everything as context, which turns the cheapest component into the most expensive for no gain. And no maintenance budget, so the assistant quietly rots.

What to ask before you sign

Ask for the lines broken out separately, not one bundled figure. Ask which country’s rates it uses. Ask how many outbound messages are assumed and in which category — that is the number that moves the bill. Ask what the maintenance retainer covers, in hours and scope.

And ask the awkward one: if the relationship ends, who owns the assistant, the prompts, the conversation logs and the customer data? The answer should be you.

How we scope it

We separate inbound from outbound first, because one is close to free and the other is not, then add a conservative model estimate, hosting and a fixed maintenance allowance. Less exciting than a vendor headline, considerably closer to the invoice.

Then we test it against reality rather than trusting it. That is the point of starting narrow: one process, on your real data, with the real cost measured over a month instead of estimated in a spreadsheet.

Frequently asked questions

Is it true that WhatsApp replies are free now?

Replies sent inside the 24-hour customer service window — opened when a customer messages you — are free of platform charges, and have been since November 2024. Messages you initiate outside that window are charged by category, and marketing templates are always charged.

Is a WhatsApp AI assistant cheaper than hiring someone?

For high-volume, repetitive enquiries, usually yes — but the honest comparison is not assistant versus employee. It is assistant plus a smaller team versus a larger team. The assistant absorbs the repetitive majority; people handle the exceptions, which is the work worth paying people for.

What is the minimum volume that justifies building one?

Below a few hundred enquiries a month, the platform fee and maintenance are hard to justify against answering messages yourself. The economics improve sharply with volume, because the fixed costs spread while the human alternative rises in a straight line.

Can it answer in Kuwaiti dialect, not just formal Arabic?

Yes, and it should. Customers do not write in Modern Standard Arabic. Handling dialect well is mostly a matter of testing against real transcripts rather than idealised sample questions, and correcting what comes back wrong.

What happens when it does not know the answer?

It should say so and hand over to a person. An assistant that invents an answer to avoid admitting ignorance is a liability, not a feature. This is a build decision, and the most important one for anything customer-facing.