WhatsApp AI agents built into your systems

A WhatsApp AI agent built into your systems is the right build when you already sell or support on WhatsApp and the reply has to check a calendar, a CRM, an ERP, or an order, then hand the thread to a person who can see what was said. If the question is only opening hours and a catalog, a ready-made product or Meta's own agent may be enough. A samambai project starts at US$ 3,000, with scope, timeline, and price fixed before we write code.

Who it is for

It is for teams whose customers already live in WhatsApp. That is the default thread across much of Latin America, in Spain, in Israel, and in India: the way a US buyer might still assume SMS or email. A clinic confirming a visit, a hotel and a B2B wholesaler answering after hours, a tourist site, a property manager, a visa desk. The pattern is the same. Most messages are repetitive. A few need a human who is not starting from zero.

It is not for a company that does not use WhatsApp yet and wants a bot before it has a process. It is also not for a 200-seat contact center whose product is the inbox platform itself. In that case the platform is the system. If we belong there at all, we belong as the integration, not as a replacement for the platform.

The buyer we usually meet has already tried one of three things: a vertical SaaS that does not talk to their stack, a button-tree bot that dies on the exception, or a person copying WhatsApp into a calendar. The agent does the repeated part and returns the exception with context.

Typical systems we wire in, when they expose an API or a reliable export: HubSpot, Salesforce, Shopify, and calendars (Google Calendar and the booking tools that sit on top of it). We integrate with what you run. We do not migrate you onto a CRM so that our agent has an easier week.

What we deliver

The build sits on the WhatsApp Business Platform, Meta's Cloud API. Outside the customer service window, a business may only start a thread with an approved message template. Inside the window, the reply can be free-form Source [3] Source [4].

A project usually includes:

• Your number on the official API, with a business profile that has a real support contact. The policy requires that Source [4].
• An agent whose rules are written down: what it may answer, what it must look up, what it must not invent.
• An integration into the system that already decides the operation: calendar, CRM, ERP, store, or sheet. Integration, not a rip-and-replace.
• A handoff to a person inside the chat, with the history attached. When automation runs inside the 24-hour window, WhatsApp's policy requires a prompt escalation path: in-chat transfer to a human, a phone number, email, web support, an in-store visit, or a form Source [4]. We treat the in-chat transfer as the default, because that is the thread the customer is already in.
• Message templates (utility, authentication, or marketing) submitted for Meta's review. Only a template in APPROVED status can be sent. Review can take up to 24 hours, and a rejected template does not go out Source [5].
• Data care: least-access credentials, data in your accounts, client data never used to train a public model. Where the people are in Brazil, health data is sensitive personal data under the LGPD Source [7]. Where the people are in the EU, we apply GDPR practices the same way: minimum data, a purpose, a place the data actually lives.
• Monitoring of what the agent answered, what it handed off, and what Meta categorized as billable.

The project fee does not include Meta's invoice, the model provider's invoice, or n8n and hosting. Those bills stay in your name. How Meta's rate card works, including the October 2026 split across three official surfaces, is in off-the-shelf or custom and on pricing.

What Meta actually charges

Since July 1, 2025, Meta charges per delivered message, by template category and by the recipient's country Source [1]. Since October 1, 2026, a reply inside the 24-hour window (a service message, written by a person or by an AI) is charged too, at the country's utility rate, and a utility template inside the window is no longer free. Each business phone number gets 1,000 free service messages a month, with no rollover. A marketing template is charged on every delivery Source [1] Source [2].

Meta publishes 0.68 US cents per delivered message as the utility and authentication rate for Brazil Source [2]. A thousand replies past the allowance come to about US$ 6.80 in Meta fees, before the model. The United States and Canada are priced as North America. Spain, Israel, and India have their own rows Source [1]. You pay the recipient's market, not the market where your headquarters sits: a New York company messaging customers in Mexico, Brazil, or Spain pays those rows. Meta's pages do not yet describe this change the same way, so the budget closes in WhatsApp Manager, in the currency of the account.

Meta's own agent (Meta Business Agent) is a separate category: US$ 2 per 1 million tokens, roughly 4 to 5 US cents per message, as one charge Source [2]. A third-party agent, which is what we build, splits the bill: Meta charges delivery and the model provider charges tokens. The point is not a cheaper message. The point is a call into your system, under a rule we can audit.

A person who arrives from a click-to-WhatsApp ad opens a free entry point window. Its length differs between Meta's pages (72 hours on one, up to 7 days on the other), so budget the shorter one Source [1] Source [2]. The detail, with worked numbers, is in off-the-shelf or custom.

When a ready-made tool is enough

A vertical SaaS is enough when your system of record is already on its connector list and the rule is the product's rule: confirm, reschedule, remind. A helpdesk platform (shared inbox, several channels, a queue) is enough when the job is routing conversations across people and the AI is an add-on on that queue.

Meta's own agent is enough for public business facts: hours, address, catalog, tone. It is not, in the way we design, the layer that writes into your ERP with an audit trail that does not live inside Meta.

A custom agent is the right call when one of these is true:

• The answer depends on data that exists only in your stack (an open slot, an order status, an invoice, a location, a plan rule).
• The exception has to reach a named person, with what was already said, not a generic inbox.
• The thread can touch sensitive data. Health is not a FAQ. Under Brazil's LGPD, data about health is sensitive Source [7]. Processing it is allowed only in the cases in article 11, including specific, highlighted consent for a specific purpose, or health care performed by a health professional, a health service, or a health authority, in the wording introduced by Law 13,853 Source [7]. WhatsApp's own policy says not to use the channel for telemedicine, and not to request health information, where local rules forbid sending that data to a system that does not meet the heightened bar Source [4]. The agent books and routes. It does not diagnose, and it does not keep a clinical note in the chat.
• You will not run an unofficial session. WhatsApp's terms forbid unauthorized access and forbid building an API that substantially mimics the service without permission. Violations can mean the account is disabled or suspended Source [6]. On the Business Platform, repeated violations step from a one-day block up to a permanent disable Source [8].

The full decision table, including shelf prices reported by a vendor, is in off-the-shelf or custom.

What we do not do

We do not run your number through a modified client. We do not ask for a full card number, a bank account, or a national ID in the chat. WhatsApp's policy forbids requesting those identifiers Source [4]. We do not dress a marketing template up as utility to pay a lower rate. You are responsible for the category on an approved template, and you accept that category's rate at send time Source [1]. Misclassification is on the violation list that can restrict the account Source [8].

We do not promise that Meta will approve copy on a given day, that a customer will reply, or that no-shows will fall by a percentage. We do not staff emergencies. WhatsApp's own terms say the service does not connect you to police, fire, or a hospital Source [6].

If the work also needs a routine outside the chat (a nightly sheet, a morning ERP pull, an internal alert), that is n8n automation, in the same project or the next sprint. It does not require a second vendor.

Proof, with no client names

Agents in production, never with the company named: a medical clinic group, for front desk and scheduling; hospitality and B2B retail; an adventure park; a condominium administrator; a visa advisory. Behind part of that operation is our own n8n server, with dozens of flows live.

That shows the kind of problem that has already survived the first week in production. It does not show a conversion rate. We do not publish one.

When not to hire samambai

Do not hire us if a vertical product already integrates the system you use and the rule fits on its screen. Do not hire us if all you need is a catalog and opening hours, and Meta's agent is available in your country. Do not hire us if nobody can take the handoff. Do not hire us if the budget only works on an unofficial client. Do not hire us if you want medical, legal, or immigration advice inside the chat: the agent organizes the request and hands it to the person whose job that is. It does not replace them.

The usual fit is a Project, from US$ 3,000. A first slice, one system and one rule, can be a Sprint at US$ 1,200 for two weeks, if API access already exists. Ongoing work after go-live fits a Partnership, from US$ 900 a month. What each one includes, and what it leaves out, is on pricing. The first call is 30 minutes, with no commitment.