WhatsApp AI agents: off-the-shelf tool or custom-built? How to decide

Buy the off-the-shelf tool when the system you already run is on that product's connector list and the rule (confirm, reschedule, remind) is the product's rule. Build a custom agent when the reply has to read your calendar, your CRM, or your ERP, apply an exception the product does not have, and hand the thread to a person with the history attached. The software subscription is not the cost of the operation. Meta bills the message on the side, and so does the model provider.

If your customers are in Brazil, Mexico, Spain, Israel, or India, WhatsApp is the thread they already answer. A team in New York selling into those countries is not "adding a channel." It is showing up where the buyer already is. The decision is still the same one a São Paulo clinic makes.

Decide this before you look at a price

Four questions catch most bad purchases. Answer them with the name of a system, not with "we'll see."

Where does the fact that changes the answer live? If it lives on a public catalog, a website, and the hours on the door, Meta's agent or a FAQ product will do. If it lives in "Tuesday at 3 is open," in an order status, or in a plan rule, something has to read your system. A product with no connector to that system does not read it by charm.

Who takes the exception? If nobody does, you do not have an agent. You have a chat that stalls. WhatsApp's Business Messaging Policy, updated September 23, 2026, allows automation inside the 24-hour window and requires an escalation path: an in-chat transfer to a person, a phone number, email, web support, an in-store visit, or a form Source [4].

What happens after 24 hours? A customer who goes quiet cannot be nudged with free-form text. A reminder, an order update, and a promotion are not the same template category, and they are not the same price Source [1] Source [3].

Is the data sensitive? Health, a national ID, a card number. Each one has a stop, below. If the honest answer is "the bot asks for the ID and the symptoms so we can get ahead," the design is wrong before you pick a vendor.

With those four answers, the table is a filter, not a menu.

Four paths

We build the fourth row. The first three are the better buy when they actually fit. What a project includes is on WhatsApp AI agents.

HubSpot, Salesforce, Shopify, and a calendar are the usual targets when a US or European team describes the stack. A vertical clinic product in Brazil will name a local practice-management system instead. The test is identical: can the product read the record, and can it write the result back? A "WhatsApp integration" that only notifies a human, without the slot or the order, has not passed the test.

Shelf prices: a vendor's reading, not a census

Nação Digital, a Brazilian content vendor that also sells consulting (including Salesforce), published a comparison of nine customer-service AI tools on July 31, 2026, and updated it on September 24. The author says the prices were read on the vendors' own pages in July 2026. The set is their selection, not a market census. The figures below are what that page claims. We did not check each checkout ourselves Source [9].

• JivoChat: an AI agent at R$ 343 a month, on top of a per-operator plan (Basic at R$ 63 per operator on annual billing) and a WhatsApp module at R$ 175 a month.
• Zenvia: Specialist at R$ 600 a month (10 users). Channels are a separate pack, from R$ 100 a month, and WhatsApp activation has a setup from R$ 649.
• Octadesk: One at R$ 2,499 a month (3,000 conversations), with AI included in the price they report.
• Blip: base subscription not shown on the page. Overage of R$ 1.40 per conversation on Plus and R$ 1.25 on Super, in their reading.
• Zendesk: Suite Team at R$ 359 per agent on annual billing, on the Brazilian site as Nação Digital reports it, with automated resolutions billed on use. The page does not close the total.
• Intercom Fin: from US$ 0.99 per resolution. Nação Digital itself says the seat price was not clear on the official page.
• Freshdesk: Growth at US$ 19 per agent on annual billing, with 500 AI sessions included and a paid pack after that.
• Salesforce Agentforce: Nação Digital says, in so many words, that it could not confirm the figures on the official page. Do not use those numbers.

Prices are in Brazilian reais because that is how those vendors publish them. A buyer in Madrid or Tel Aviv paying a global helpdesk will see dollars or euros instead. The shape is what travels: a low monthly seat, a conversation pack, and a WhatsApp line item that was not in the first screenshot.

A monthly tool at a few hundred reais and a US$ 3,000 project are not competing for the same job. They compete only when someone pretends they do. If the R$ 343 agent already confirms inside the calendar you use, buying a project is waste. If it cannot see that calendar, the low subscription buys a FAQ.

What Meta bills on top of all of them

Since July 1, 2025, the charge is per delivered message. Template category plus the recipient's country Source [1]. The HTML pricing page, updated September 30, 2026 and read on October 2, says the October 1 change is in effect: a service message and in-window utility are charged. Each business phone number gets 1,000 free service messages a month, with no rollover. The charge starts at message 1,001. The allowance does not cover utility or marketing. You are responsible for the category on an approved template, and you accept that category's rate at send time Source [1].

The page on non-template messages describes the same October 1, 2026 change and still presents it as upcoming Source [2]. The Markdown export of the pricing document, read the same day, still describes free-form replies and in-window utility as free Source [12]. A serious budget keeps all three sentences and checks WhatsApp Manager. It does not keep only the sentence that makes the slide look cheaper.

For Brazil, that second page publishes 0.68 US cents as the current utility and authentication rate per delivered message Source [2]. That is not the marketing rate. It is not Spain, Israel, India, or North America. Those markets have their own rows. The United States and Canada are priced together as North America Source [1]. You pay the recipient's row. A company in New York messaging customers in Brazil pays the Brazil row.

Higher volume lowers utility and authentication rates, by market and by category, and the counter resets each month. Marketing is not on that ladder on the main page. Service messages, on the newer page, have no volume tier either Source [1] Source [2].

A worked example, assumptions in the open. One thousand utility reminders to Brazilian numbers, at 0.68 cents: US$ 6.80. Utility does not draw down the service allowance. That excludes marketing, the model, and the software. If the HTML rule is what the invoice uses and one number delivers 4,000 service messages, 1,000 sit in the allowance and 3,000 times 0.68 cents is US$ 20.40, before tokens. Without the allowance, the same 4,000 would be US$ 27.20. The Markdown export still describes those messages as free Source [12].

Meta Business Agent, Meta's own category, uses a different meter: US$ 2 per 1 million tokens, a typical message at 20,000 to 25,000 tokens, about 4 to 5 US cents, as a single charge Source [2]. Their comparison for 10,000 messages in Brazil, with a lower-complexity third-party AI, estimates about 2 to 3 US cents a message once a model and the 0.68-cent delivery are added. That estimate is Meta's, not a measurement of ours.

Eligible businesses billed in Brazil have created new accounts in BRL since July 1, 2026, and have until June 30, 2027 to migrate the rest. On July 1, 2027, Meta stops delivering from unmigrated accounts for those customers Source [12]. India has its own cutoff, December 31, 2026, for accounts that must move to INR Source [1].

The model and n8n, which the SaaS screenshot leaves out

A custom agent calls a model. On the OpenAI API pricing docs, read on October 2, 2026, one of the low-cost models on the table (gpt-6-luna, short context, standard rate) is US$ 0.10 per 1 million input tokens and US$ 0.50 per 1 million output tokens Source [11]. The page shows no publication date.

An illustration, not a client measurement: 1,500 input tokens and 250 output tokens per reply. That is about US$ 0.0003 a reply. Five thousand replies in a month land near US$ 1.40 of model cost. Next to Meta, once delivery is billed, the token line is usually the small one. Cache, long context, and web search change it. Web search on the same page is US$ 10 per 1,000 calls, plus the retrieved tokens at the model's rate Source [11]. An agent that only calls your API does not need that search.

If the agent's core runs on n8n Cloud, the official pricing page read the same day shows Starter at €20 a month on annual billing (2,500 executions) and Pro at €50 a month on annual billing (10,000 executions). A monthly toggle exists. Annual is labeled 17% cheaper. One execution is a full workflow run, not each step. The FAQ's chatbot example says to multiply conversations by messages: each message that fires the workflow counts Source [10]. Five thousand messages in a month blow past Starter and fit in Pro. The published max duration is 5 minutes on Starter and 40 on Pro. The same page says hosted data sits in the EU, in Frankfurt. Community Edition is the version on your own server, published on GitHub: that software does not carry this subscription, the server does, and the server bill is yours.

None of the three invoices (Meta, the model, n8n) is inside the US$ 1,200 sprint or the US$ 3,000 project. They are split out on pricing and in how much AI automation costs.

Failures we treat as production defaults

An unofficial session. A library glued to WhatsApp Web, a modified app, a "cheap API" that is not the Cloud API. WhatsApp's terms, which still display an effective date of January 4, 2021 on the page we read in 2026, forbid unauthorized access, forbid reverse engineering, and forbid creating software or an API that functions substantially the same as the service and offering it to third parties without permission. A violation can disable or suspend the account Source [6]. On the official platform, a repeated violation (spam, a template in the wrong category) steps from a 1-day or 3-day block to 5, 7, or 30 days, then an indefinite lock, then a permanent disable if nothing changes Source [7]. The number you lose is the number on the card, the site, and Google. Rebuilding on the official API after that is slower than starting there.

A 24-hour window counted wrong. The window opens when the person messages you or calls you, and it resets when they do that again. A message from the business does not reset the clock. Window closed: approved templates only Source [3] Source [4]. An agent that "replies" 26 hours later in free text does not reply. The API refuses it. Tomorrow's reminder has to be a utility template, written and approved ahead of time, not a sentence improvised at 9 p.m.

The wrong template category, or a template still in review. Marketing, utility, and authentication are not interchangeable. Utility is an update on something the person asked for (an order, a time, an account). Marketing is an offer. Review runs when you create the template. Anything other than APPROVED does not send. In WhatsApp Manager, review can take up to 24 hours Source [5]. Submitting Monday's reminder on Friday afternoon is how the date slips. Low quality can pause or disable a template even after approval Source [5].

A handoff with no history. The customer stated the request three times, and the person who picks up sees an empty thread. That is not a UI nicety. It is why the policy requires a clear escalation Source [4]. The handoff we ship carries the thread. If your team only watches email, the design has to say so in week one, not after the complaint.

A Brazil or Mexico number with no country code. Cloud API docs warn that if the request goes out without the plus sign, the business number's country code is prepended, and the message can land on the wrong person. For Brazil and Mexico, the extra prefix may be modified by the API itself. The docs call that standard behavior, not a bug Source [3]. Test with a real number in both countries before you call it done. A spreadsheet export that drops a digit, or duplicates one, is the same bug from another door.

Sensitive data in the wrong place. Brazil's LGPD defines health data as sensitive personal data. Processing it fits only the cases in article 11. The wording from Law 13,853 limits the health-care basis to a procedure performed by a health professional, a health service, or a health authority Source [8]. Consent, when that is the basis, has to be specific and highlighted, for a specific purpose. WhatsApp's policy says not to use the channel for telemedicine and not to request health information where local rules require a system that meets a heightened bar. It also says not to request a full card number, a financial account number, or a personal ID number Source [4]. A clinic agent confirms, reschedules, and routes when the person describes a symptom. It does not run the consult. A visa agent organizes the document that was requested. It does not give immigration advice in the chat. A team in the EU should treat the same fields as special-category data under GDPR practice: collect less, keep it in your system, do not paste the chart into a prompt "so the model has context."

No opt-in. You may only contact someone who gave you their number and opted in to later messages. You pick the method, and it has to satisfy the law that governs that communication. A stop request, on WhatsApp or off it, ends the contact Source [4]. A purchased list is not opt-in. It is the short path to an account restriction.

How to choose in an afternoon

1. Write the sentence the customer actually sends. "I need to move my appointment" is one. "The order is late and the invoice expired" is another.
2. For each sentence, name the system that holds the answer. If the name is "none," you either do not need an agent, or you need to fix the data first.
3. Name who receives the sentence the agent must not close, and during which hours that person exists.
4. Split replies (window open) from follow-ups (template). Count templates. Do not count "the bot will figure it out."
5. If steps 2 and 3 fit a product you would already pay for, buy the product. If they do not, a custom build is the path. The first call with us is 30 minutes, with no commitment.

When not to hire samambai

Do not hire us if a vertical product already talks to your system and the exception fits on its screen. Do not hire us if a catalog and opening hours are the whole job, and Meta's agent is available for your number. Do not hire us to replace a fifty-person contact center: that is a platform, and the platform bills like a platform. Do not hire us if there is no human for the handoff. Do not hire us if the condition is an unofficial API. Do not hire us for a diagnosis, a legal opinion, or visa advice.

Hire us when WhatsApp is already the channel and your system is the source of truth. A clinic group, hospitality and B2B retail, an adventure park, a condominium administrator, and a visa advisory already run agents we built, with no client name on this page. A project starts at US$ 3,000. What a third-party subscription does not cover is on pricing.