How much does a WhatsApp AI assistant cost?
Much less of the price depends on the AI model than people expect. Most of it depends on what the assistant is allowed to do, and whether it has to read or write in your own system. Here are the four levels we see, what moves the cost within each, the running costs that are not the build, and what a proposal should contain so you can compare quotes.
The short answer
The price of a WhatsApp AI assistant is set mostly by two things: what it is allowed to do, and whether it has to read or write in your own system. An assistant that answers from your information and passes everything else to your team is a contained piece of work. One that looks up a customer’s file, opens a request in your platform or updates a record has to be built and tested against that system, and that is where most of the effort goes.
The AI model matters far less than people expect. It is a component you pay for per message once the assistant is live, not the bulk of the build.
Ask what the assistant will be allowed to do before you ask what it costs. The second answer follows from the first.
Four levels, by what the assistant is allowed to do
Each level includes everything in the one before it. The jump that matters most is from the second to the third, when the assistant starts working inside your own system.
Answers and refers
It answers customers’ questions from your own information: services, requirements, documents, timelines. When a question falls outside that, it tells the customer how to reach your team.
Your information written down, or a few sessions with the people who hold it. A list of what it must never answer.
Hands the live conversation to a person
When a case needs judgment, it passes the conversation to someone on your team with the full history attached, then picks up again when they are done. The customer stays in the same chat.
Who takes over, and in which tool your team already works. A written rule for when it must hand over.
Reads and writes in your own system
It identifies the customer, checks the status of a file or order, and opens or updates requests in your platform through its API, so nobody re-keys what the customer already said.
Access to your system’s API, a test environment, and someone on your side who knows which fields matter.
Adds the operational layer
Tickets in your team’s tools, voice notes transcribed, photos and documents read, bursts of messages grouped into one turn, and messages it sends first, such as reminders or status updates.
Your ticketing tool, the outbound messages you want and the reason for each, and agreement on what it may send without a person.
At Fixlab, the assistant we built answers from their knowledge base, files requests through their platform, reads photos and voice notes, and hands cases to a person. Most businesses do not need all of that on day one.
What moves the cost within a level
Two assistants at the same level can take very different amounts of work. Four things explain most of the difference.
- How written down your rules already are. If your criteria live in a document, we build from it. If they live in two people’s heads, part of the work is writing them down with you, and that takes time on both sides.
- How many conversation paths there are. One service and one kind of request means a few paths. Several services, customer types and exceptions mean many, and each path has to be built and tested.
- The state of your system’s API. A documented API with a test environment is straightforward. One that is missing, undocumented or has no test environment adds work before the assistant can use it. This only matters from the third level on.
- How fast your team reviews test rounds. We test against real scenarios and your team checks the answers. Reviews that come back in a couple of days keep the project moving. Reviews that take weeks stretch it, and a longer project costs more.
The model you choose barely moves the build. How clear your rules are and how ready your system is moves it a lot.
Costs that are not the build
Once it is live, an assistant has three running costs. We set each one up in your name, on your accounts, so you pay the provider directly and see every invoice. None of them is a fee to us.
- WhatsApp messages. WhatsApp charges per message, and from 1 October 2026 that includes service messages, the replies a business sends in a conversation the customer started. Rates depend on the country and the type of message. We explain the change in what WhatsApp charges for service messages from October 2026.
- The AI model. Each reply calls a model once or several times, and the provider charges for the text it reads and writes. Long conversations, a large knowledge base and images or voice notes cost more per message.
- Hosting. The messaging layer and the agent run on a server or cloud account. It is a fixed monthly line that grows slowly with volume.
How to estimate them before you commit
- 01Count your customer conversations per month today, from WhatsApp, your inbox or your phone logs.
- 02Estimate the messages per conversation, counting the assistant’s replies and anything it sends first.
- 03Multiply by WhatsApp’s published rate for your country and message type.
- 04Ask for the model cost per conversation measured during testing, not guessed. We measure it in the test rounds and put it in writing.
- 05Add hosting as a fixed monthly line.
Run the numbers at today’s volume and at double it. If the assistant only makes sense at the lower figure, you want to know that before anything is built.
What a good proposal should contain
Quotes for what sounds like the same assistant can look very different, usually because they describe different things. A proposal you can compare with another one should contain:
- A fixed price where there is no uncertainty. If the scope is clear, the price should be fixed, not an open hourly estimate.
- A named variable where there is. If something is still unknown, such as the state of your API, the proposal should name it, say how it will be resolved and say how it affects the price.
- Milestones with a written definition of done. Each payment tied to something you can check, such as the assistant passing an agreed set of test scenarios.
- Correction rounds. How many rounds of changes are included after each review, and what happens after that.
- Warranty. How long fixes are covered after launch, and what counts as a fault rather than a new request.
- Who owns what. The WhatsApp number, the accounts, the knowledge base and the code: whose name each one is in, and what you keep if you stop working with the studio.
A quote without these is not cheaper or more expensive than another. It is just impossible to compare.
FAQ: the cost of a WhatsApp AI assistant
How much does a WhatsApp AI assistant cost?
Why don’t you publish prices?
Is the AI model the main cost?
What does it cost to run once it is live?
How do I compare two quotes?
Can we start small and add more later?
To see what a finished assistant does, read about our WhatsApp customer service assistant. For the same question about any other tool, see how much a custom AI tool costs.