What a WhatsApp AI agent actually costs
The cost layers of a WhatsApp AI agent, what makes one proposal ten times another, and how to read a quote before signing it.
Anyone asking for three quotes on a WhatsApp AI agent usually receives three numbers that do not talk to each other, and none of them explains properly what is being charged. It is not market dishonesty: the cost has four layers with different logics, and each vendor shows the layer that suits them. This piece separates the layers, explains what makes the price vary, and shows how to read a proposal before signing.
The four cost layers
1. Meta's bill, which is independent of who builds it
Since 1 July 2025 the WhatsApp Business Platform charges per message rather than per conversation. What gets charged are templates, split into three categories: marketing, always charged, plus utility and authentication, charged when sent outside an open service window. Service messages and replies inside an open window are not charged, and the amount varies with the country code of the receiving number.
This layer is the most predictable of all, because the table is public and does not depend on negotiation. It is also, almost always, the smallest.
2. The provider and the platform
Between your company and Meta sits an official provider, charging its own margin per message or a platform subscription. This is where the first large difference between proposals appears: two vendors using exactly the same Meta infrastructure can charge quite different amounts, because one passes the cost through and the other bundles service into it.
3. The AI model
If the agent thinks, somebody pays for the thinking. Language models charge per token processed, which in practice means long conversations and large histories cost more than short replies. The choice of model shifts the cost by orders of magnitude, and the most expensive model is not always the most appropriate for the task.
4. The build, which is where the money actually goes
Designing what the agent does, which questions it answers, when it hands over to a human, what it records in the CRM, and how it behaves when it does not know the answer. That layer is not technical, it is about process, and it is what separates an agent that resolves things from one that irritates customers and burns the brand.
Add maintenance. An agent is not a finished building: the catalogue changes, the returns policy changes, the team discovers a new question it cannot answer. A vendor who does not mention maintenance is omitting cost, not saving you money.
Why one proposal costs ten times another
In practice, large price differences almost never come from the technology. They come from three choices:
| Choice | Cheap | Expensive | What changes |
|---|---|---|---|
| Conversation scope | Answers frequent questions | Queries systems, creates orders, books appointments | Integration with what the company already uses |
| Level of autonomy | Suggests and hands over | Decides and executes | Risk, and therefore testing and governance |
| Source of content | Generic text | The company's knowledge base | The work of organising what the company knows |
An agent that only answers questions about opening hours and address is cheap because the problem is small. An agent that checks stock and closes an order is expensive because it touches the operation, and touching the operation requires method.
How to read an AI agent proposal
Ask for the proposal to separate, on distinct lines: platform cost, cost per message, model cost, initial build and monthly maintenance. If it all arrives as one number, you are not comparing proposals, you are comparing trust.
Then ask three questions:
- What happens when the agent does not know? An agent with no exit route to a human becomes a trap.
- Who owns the conversation and the data? Leaving a vendor cannot mean losing the history.
- What changes in my internal process? If the answer is nothing, be suspicious. An agent that changes no process usually just pushes the bottleneck further along.
The cost nobody puts in the spreadsheet
Automating a confused service process makes the confusion faster. If today the team cannot handle a returns request without consulting two people, the agent will not know either, and the customer will get a wrong answer in three seconds instead of a right answer in three hours.
Technology amplifies the process it finds. If the process is good, it scales quality. If it is bad, it scales the problem.
That is why, in consulting work, designing the service process comes before choosing the tool. Not out of methodological purism, but because it is the part that determines whether the investment pays back.
Frequently asked questions
Can I start small and grow later?
You can, and it is usually the cheapest path. Start with what the team answers repeatedly, measure how many conversations the agent resolves alone, and only widen the scope once that number is stable.
Does Meta's per-message cost weigh on the budget?
For most small and mid-sized operations, it is the smallest of the four layers. It only becomes significant at high volumes of marketing sends, which is the most expensive category.
Do I need a technical team to maintain it?
Not necessarily, but you do need somebody internal with clarity about the process, deciding what the agent may and may not do. Outsourcing the operation is possible; outsourcing the decisions about your own business is not.
Before asking for a quote
The best way to avoid overpaying for an agent is to arrive at the conversation knowing which problem it solves, how many times a month that problem appears, and how much time it consumes today. With that in hand, the proposal stops being a loose number and becomes comparable to the cost you already carry.
Sources
Apply
Want to apply this in your business?
A complimentary conversation to understand your context and decide, together, whether it makes sense to move forward.
How to organise your data before using AI
What needs tidying in company data before applying AI, and why skipping that step produces answers that are confident and wrong.
IA no suporte interno: o help desk da empresa
Como usar IA para responder as dúvidas internas que consomem o tempo de quem sabe, e por que o ganho aparece antes no RH que na TI.
IA no estoque: previsão e reposição
Onde a IA ajuda na gestão de estoque, por que previsão sem cadastro limpo não funciona e o que resolver antes de tentar prever demanda.