---
title: "How much does AI automation cost in 2026? Our actual prices"
url: https://samambai.com/blog/how-much-does-ai-automation-cost/
date: 2026-10-03
author: samambai
---

AI automation in 2026 is our fee plus the vendor bills in your name. The public table is a sprint at US$ 1,200, a project from US$ 3,000, and a partnership from US$ 900 a month. WhatsApp, the model, and n8n are not inside those numbers.

AI automation in 2026 costs our work plus the invoices that stay in your name. The work has three public prices, in US dollars: a sprint at US$ 1,200 for 2 weeks, a project from US$ 3,000, and a partnership from US$ 900 a month. WhatsApp, the model, n8n, and hosting are not inside those numbers. A quote that stops at the first line is half a budget.

## The table we publish

| Way | Price | What you are buying | What the line does not cover |
| --- | --- | --- | --- |
| Sprint | US$ 1,200 per 2-week cycle | One focus. Cancel whenever you want | A second system in the same cycle. Project support after the cycle. A vendor invoice |
| Project | From US$ 3,000 | Scope, timeline, and price locked. A delivery every sprint. Support after delivery, for the length written into that scope | A change outside the scope. An open roadmap. A vendor invoice. A revenue promise |
| Partnership | From US$ 900 a month | A continuous roadmap. Priority in the queue. A monthly check-in | A whole product in month one. A dedicated full-time person. A response time in minutes. A vendor invoice |

The page that spells out what is in and what is out is [pricing](/pricing/). Where the table comes from is in [meet samambai](/blog/meet-samambai/).

Priority, on a partnership, is queue order. It is not an on-call shift. Support on a project is not a default 30 days or 90 days. It is whatever the locked scope says. We do not fill that blank with a round number.

What fits in a sprint, when access to the system already exists: one n8n flow (a store order into a spreadsheet and a WhatsApp notice), one agent cut with one rule and a human handoff, one Lovable path moved to an environment where you can charge, or a GEO diagnosis on a small site. Several rules, several numbers, or a data migration push the work to a project. The pages are [WhatsApp](/whatsapp-ai-agents/), [n8n](/n8n-automation/), [Lovable](/lovable-to-production/), and [GEO](/geo/).

Most automations go live in 2 to 4 weeks. Full projects, in 4 to 8 weeks. That is our delivery window. It is not Meta's window for approving a message template.

What moves the price up from the floor, and can be named before the proposal: how many systems, whether they have an API or only a screen, how many numbers and languages, how many rules and exceptions, whether the data is health, an identity document, or a payment, and whether handoff needs a queue and business hours. What does not move the price: the buyer's city. São Paulo, New York, Madrid, and Tel Aviv use the same table.

## What the market charges for AI support

Nação Digital, a Brazilian vendor that also sells consulting (including Salesforce), published a comparison of nine support tools. The prices on the page are labeled July 2026. The set is their selection, not a census. The figures below are what that page claims. We did not re-open each checkout in October [5]. Several rows are in Brazilian reais because the article is written for Brazil. Do not convert them with a guessed exchange rate. The dollar rows are dollar rows.

| Tool | How they charge, as Nação Digital reports it | Figure the page publishes |
| --- | --- | --- |
| JivoChat | Per operator, billed annually, plus modules | Basic R$ 63 per operator. AI agent R$ 343 a month. WhatsApp module R$ 175 a month. A free plan covers up to 5 operators |
| Zenvia | A monthly band, channel billed apart | Specialist R$ 600 a month (10 users). Starter R$ 0. Expert R$ 1,800. Professional R$ 3,900. Channels from R$ 100 a month. WhatsApp setup from R$ 649 |
| Octadesk | A monthly fee with a conversation allowance | One R$ 2,499 (3,000 conversations). Flow R$ 4,399 (5,500). Extra user R$ 100 to R$ 130 |
| Blip | Base monthly fee not shown, plus overage | Overage R$ 1.40 (Plus) or R$ 1.25 (Super) per conversation. Extra agent R$ 150 or R$ 100 |
| Zendesk | Per agent on the Brazilian site, annual, plus resolutions | Suite Team R$ 359 per agent. Support Team R$ 135. Suite Professional R$ 755. Copilot R$ 350 per agent. Automated resolutions billed per use, with no closed total |
| Freshdesk | Per agent, annual, plus AI sessions | Growth US$ 19. Pro US$ 55. Enterprise US$ 89. 500 sessions included, then US$ 49 per 100 sessions |
| Tidio | A monthly fee plus AI conversations | Starter US$ 24.17 a month. Growth from US$ 49.17. Plus from US$ 300. Lyro from US$ 32.50 a month (50 conversations) |
| Intercom Fin | Per resolution | From US$ 0.99 per resolution. Nação Digital says the seat price was not clear on the official page |
| Salesforce Agentforce | The page lists three models and says it could not confirm them on the official page | Do not use US$ 2 per conversation, Flex Credits, or US$ 125 per user as fact |

A clinic SaaS that other guides put under a few hundred reais is not in this table. We did not check those checkouts ourselves. The decision that matters is in [off-the-shelf or custom](/blog/whatsapp-ai-agent-off-the-shelf-or-custom/): if the product already confirms a time on the calendar you use, the monthly fee is the right buy and a project is waste. If the product cannot see the calendar, the low fee buys a FAQ.

Those bands and a project from US$ 3,000 are not the same job. They only look comparable when someone drops both into one cell of a spreadsheet. A Freshdesk seat at US$ 19 is a helpdesk seat. An agent that writes an order back into Shopify, or a deal back into HubSpot or Salesforce, is a different build. Typical targets, not partnerships.

## The bill a software quote leaves out

### WhatsApp

Since July 1, 2025, Meta charges per delivered message, by category and by the recipient's country. A message from the person to the business is not charged [1].

The HTML pricing page, updated September 30, 2026 and read on October 2, says the October 1, 2026 change is in effect. A service message (a person or a third-party AI) is charged. A utility template inside the 24-hour window is charged too. Each business phone number gets 1,000 free service messages a month. Unused messages do not roll over. The charge starts at message 1,001. The allowance does not cover utility or marketing [1].

The Markdown export of the same document, read the same day, still says free-form text and in-window utility are free, and still treats October 1 as a future update [6]. The page on non-template messages still titles itself as an upcoming update, describes the same charge, and publishes 0.68 US cents as the utility and authentication rate for a message delivered in Brazil. On that page, a service message uses the same rate and has no volume tier [2]. Three official texts, three readings. WhatsApp Manager closes the invoice. A budget that has to ship today counts the charged message and treats the 1,000 as the HTML page's allowance, not as a reply that stays free forever.

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. If one number delivers 4,000 service messages in a month and the HTML rule is what the invoice uses, 1,000 sit in the allowance and 3,000 times 0.68 cents is US$ 20.40. Two numbers double the allowance, because it is per phone, not per company. Marketing is a different rate, and the dollar amount for each country was not in the rate-card section we read: Meta publishes that card as a CSV. Open the CSV for your market before you forecast a volume discount. Volume tiers exist for utility and authentication, at the business-portfolio level, and only charged messages count. Service, on the newer page, has no such ladder [1] [2].

You pay the recipient's country, not headquarters. A company in New York writing to Brazil pays the Brazil row. Spain, Israel, and India have their own rows. The United States and Canada are priced together as North America [1]. A team in Madrid texting Spanish numbers pays Spain. A team in Tel Aviv texting Israeli numbers pays Israel. Latin America is not one rate.

An eligible account billed in Brazil creates new WhatsApp Business accounts in reais from July 1, 2026, and has until June 30, 2027 to move the others. From July 1, 2027, Meta stops delivering messages from a non-reais account for that eligible customer [6]. India has until December 31, 2026 to move to rupees [1].

There is a free entry point window when the person arrives from a click-to-WhatsApp ad and the business replies inside the customer service window. The HTML page says that window may stay open for up to 7 days, and that marketing, utility, authentication, and service are not charged inside it. Meta Business Agent is still charged [1]. The other page still says 72 hours [2]. Do not budget the longer window until the invoice agrees.

Meta Business Agent, Meta's own category, is US$ 2 per million tokens, as a single charge. A typical message is 20,000 to 25,000 tokens, about 4 to 5 US cents. Their table for 10,000 replies in Brazil, with a lower-complexity third-party AI, estimates about US$ 268 in total (about 2 cents of model plus the 0.68-cent delivery). Higher complexity, about US$ 968. Meta's own agent, at the higher complexity, about US$ 400 to US$ 500. That estimate is Meta's, not a measurement of ours [2].

### The model

The OpenAI API pricing docs, read on October 2, 2026, show no publication date. gpt-6-luna, short context (up to 272,000 input tokens), standard rate: US$ 0.10 per million input tokens, US$ 0.01 cached, US$ 0.50 output. Above 272,000 input tokens, the same row goes to US$ 0.20 input and US$ 0.75 output. That is not "double across the board." Long-context output is not double US$ 0.50 [4].

A worked example, not a client measurement. 1,500 input tokens and 250 output tokens on gpt-6-luna, standard rate, short context:

- Input: 1,500 / 1,000,000 × 0.10 = US$ 0.00015
- Output: 250 / 1,000,000 × 0.50 = US$ 0.000125
- One reply: US$ 0.000275
- 5,000 replies: US$ 1.375, near US$ 1.40

Next to Meta, once delivery is charged, the token on this row is the small line. A stable prompt that hits the cache can drop the input. The cache hit rate is not a promise. Long context moves output to US$ 0.75 per million. Web search, on the same page, is US$ 10 per 1,000 calls, plus the search-content tokens at the model rate. A preview tool on the same page is priced differently. The tool name on the invoice is the one that counts. An agent that only calls your API does not need that search [4].

Batch and Flex, on the same table, cut gpt-6-luna in half (US$ 0.05 input and US$ 0.25 output on short context). That does not fit a WhatsApp reply that has to leave now. A data-residency endpoint adds 10% for a model released on or after March 5, 2026. Check the release date of the model you will actually call before you lock the budget. This page is not the price of a chat seat in a browser. It is an API token rate [4].

### n8n

On the reading of October 2, 2026, with the annual toggle on (the page labels annual as 17% below monthly):

- Starter: €20 a month, 2,500 executions, 5-minute maximum.
- Pro: €50 a month, 10,000 executions, 40-minute maximum. The plan card says up to 50 concurrent executions. The comparison table on the same page says 20 / 50. Confirm at checkout.
- Business: €667 a month, 40,000 executions, self-hosted only. Overage is €4,000 per extra block of 300,000 executions. A company under 20 employees can look at the Startup plan, 50% off Business, if they qualify. That discount is n8n's, not ours.

One execution is a full workflow run, not each step. In the FAQ, a chatbot is conversations per week times the average number of messages: each message that fires the workflow counts. Their example of a schedule every 5 minutes lands around 8,600 to 8,900 executions a month. That blows past Starter and nearly fills Pro, on a single flow. A daily schedule is 30 or 31. Hosted data sits in the European Union, in Frankfurt. If you need the data in another region, that cloud plan is the wrong region. The Community Edition is the self-hosted version on GitHub: no cloud fee of that kind, and the server bill is yours. We did not open the license terms. Do not read "it is on GitHub" as "every commercial use is free" [3].

Five thousand messages a month that fire the workflow blow past Starter and fit in Pro. If each conversation fires the flow four times, 5,000 messages are 1,250 conversations, not 5,000. The wrong sum here is counting a step, or counting a conversation when n8n counts a message.

### Hosting

A VPS, Supabase, Cloudflare, or whatever you already pay. We do not resell that invoice, and we do not publish a server price, because it belongs to the provider. Three currencies sit together: our work in dollars, n8n in euros, Meta in dollars or in reais. A cross-border card fee is not in this piece, because we did not open a central-bank table for it. Read the statement.

None of those bills is inside the US$ 1,200, the US$ 3,000, or the US$ 900. They are split out on [pricing](/pricing/).

## How to estimate the return without a promise

The method fits on one sheet. A percentage from us does not fit at the bottom.

1. Pick one repetitive task the team already does on WhatsApp: confirm a time, "where is my order," opening hours.
2. Count the hours per week you already measure. If you do not measure them, the estimate stops. Do not invent the hours.
3. Multiply by the loaded hourly cost you already know (wage plus burden, not a market average we did not research).
4. Subtract Meta (after the 1,000 free service messages per number, if the HTML page is what the invoice uses), the model, and n8n or the server.
5. Compare what is left each month with one sprint at US$ 1,200, or with the project floor of US$ 3,000.
6. Show the range. If the hours are wrong, the range is wrong. If a person still takes the handoff, subtract only the hours that actually left the week, not the whole inbox.

An illustration, assumptions written down, not a client result. Ten hours a week. 10 × 52 / 12 is about 43.3 hours a month. At a loaded US$ 25 an hour, about US$ 1,083. At US$ 40, about US$ 1,733. Those two hourly figures are placeholders for you to replace. They are not a samambai average.

Vendor bills in the same illustration, also assumptions: 1,000 utility messages in Brazil at 0.68 cents (US$ 6.80), plus 3,000 service messages charged after the allowance (US$ 20.40), plus about US$ 1.40 of model on 5,000 short gpt-6-luna replies, plus €50 of n8n Pro on the annual plan. Meta plus the model land near US$ 29. The euro stays in euros until the card converts it.

If those 43 hours actually leave the week, a sprint of US$ 1,200 sits inside one month of that US$ 1,083 to US$ 1,733 band, after the vendor bills. A US$ 3,000 project sits inside about two to three months of the same band. The "if" is the whole sentence. Without it, the sum is a wish.

The other end, so we do not sell the wrong project. Two hours a week at US$ 25: 2 × 52 / 12 × 25 is about US$ 217 a month. A US$ 3,000 project is more than a year of that labor. If a Jivo seat at R$ 343 plus the WhatsApp module at R$ 175 already writes to your calendar, that monthly fee is the right buy, not the project. We do not convert reais to dollars here, because we did not open an exchange rate.

What we will not sign: a savings percentage, a revenue lift, a count of conversations resolved, a citation in ChatGPT or Gemini, your number's quality rating, or template approval on a date. Be wary of a contract that locks those as if they were a software delivery.

## Mistakes that wreck the budget

- Treating a reply inside the window as free forever. The HTML page updated September 30, 2026 says otherwise, with a 1,000-message allowance. The Markdown export still says it is free. Keeping only the cheap sentence is the mistake [1] [6].
- Assuming the 1,000 free messages cover a utility reminder or a campaign. They do not. The allowance is service messages [1].
- Paying headquarters' country instead of the recipient's country.
- Filing an offer as utility. You are responsible for the category on the approved template [1].
- Counting an n8n execution per step, or forgetting that each chatbot message fires one execution [3].
- Using OpenAI's Batch column on a live reply [4].
- Comparing a Freshdesk seat at US$ 19 with an agent that writes into an ERP. Those are different jobs [5].
- Putting an Agentforce price on the sheet after Nação Digital said it could not confirm the number [5].
- Adding only dollars and forgetting n8n's euros and, in Brazil, Meta's reais.
- Reading "it is on GitHub" as a settled license [3].

## What not to do

Do not buy a project to replace a SaaS that already confirms a time on your calendar. Do not run an unofficial WhatsApp session to "come in cheaper": the risk of losing the number sits in the terms of service and on the enforcement page, and the walkthrough is in [off-the-shelf or custom](/blog/whatsapp-ai-agent-off-the-shelf-or-custom/). Do not ask for a medical diagnosis, a full card number, or a national ID inside the chat. Do not put a guaranteed return in the contract. Do not ask for a twelve-month program whose first software shows up in month six.

Client data: least access, data in your accounts, LGPD and GDPR practice, no training of a public model. That does not show up as a compliance surcharge. It changes the design, and the design is already in the scope.

## When not to hire

Do not buy the sprint if the problem needs five systems at once. Do not buy the US$ 900 partnership expecting a complete app in month one. Do not hire us if the hours saved are smaller than a SaaS that already fits the rule: buy the SaaS. Do not hire us if what you want is the tool license, a talk, or the cheapest hourly freelancer with no human review. Do not hire us if the condition is a savings percentage with our name on it.

Hire us if there is a flow, an agent, or an app to put live, and you want the table before the story. The first conversation is 30 minutes, with no commitment. The three numbers are already on [pricing](/pricing/).