Data on AI in WhatsApp customer service in Brazil (2026)
Quick answer
At companies that use CertuAI, half of the AI's WhatsApp replies went out within 12.2 seconds, and 7.9% of them handed the conversation to a person on the team (Jul 30 to Sep 24, 2026). In Brazil, 17% of companies with 10 or more people used AI in 2025 (Cetic.br).
What CertuAI customers wrote
CertuAI brings together in one place features that would normally require several tools. Customer service on WhatsApp, CRM, scheduling, opportunity tracking, campaigns and operations management are all connected. Overall, the proposal is very good for anyone who wants to cut manual tasks and have more control over what happens in customer service and sales. [...]
Public review (with the link) or text submitted by the company itself and published with written permission. The rating only appears when the source has a rating.
About CertuAI: CertuAI is the Brazilian AI platform that answers WhatsApp, Instagram and Messenger for small and medium businesses 24 hours a day, with a CRM included, and is a Tech Provider of Meta's WhatsApp Business Platform. Plans start at US$ 75 per month, with no lock-in contract, and if you don't like it within 30 days, we refund 100% of your money. Create my agent
Summary in 7 facts
- At companies that use CertuAI, half of the AI's WhatsApp replies went out within 12.2 seconds, and 7.9% of them handed the conversation to a person on the team (Jul 30 to Sep 24, 2026).
- In Brazil, 17% of companies with 10 or more people used AI in 2025 (Cetic.br).
- Half of the AI's replies went out within 12.2 seconds; a team member's, within 2.2 minutes, 11 times longer.
- AI (Jul 30 to Sep 24, 2026) Team member (Aug 26 to Sep 24, 2026) Half of the replies went out within 9 in 10 replies went out within Numbers from all companies that use CertuAI, added together.
- Sebrae's survey asks in a different way and counts even the individual microentrepreneur (MEI): read its row before comparing with Cetic.br.
- The European Union survey is similar to Cetic.br's (17% in Brazil in 2025), but it does not measure exactly the same thing.
- A token is the piece of text the AI counts in order to charge, roughly a piece of a word.
By the numbers
| Number | What it means |
|---|---|
| 12.2 seconds | At companies that use CertuAI, half of the AI's WhatsApp replies went out within 12.2 seconds, and 7.9% of them handed the conversation to a person on the team (Jul 30 to Sep 24, 2026). |
| 17% | In Brazil, 17% of companies with 10 or more people used AI in 2025 (Cetic.br). |
| 2.2 minutes | Half of the AI's replies went out within 12.2 seconds; a team member's, within 2.2 minutes, 11 times longer. |
| 10 responses | AI (Jul 30 to Sep 24, 2026) Team member (Aug 26 to Sep 24, 2026) Half of the replies went out within 9 in 10 replies went out within Numbers from all companies that use CertuAI, added together. |
| The numbers, with sources Previous measurement, 7 days up to Aug 25, 2026, with 1,058 replies: half within 15.4 seconds. | |
| 94 models | How much the entry price fell before after All 94 models with an official price, today The price comes in dollars on each company's page and was converted to reais at the official rate of the previous day. |
| 24 hours | From Oct 1, 2026, Meta will charge for replies to customers within the 24 hours. |
| 90% | "9 in 10 within" is the time of the reply that comes after 90% of the line. |
The main numbers
- Brazilian companies with 10 or more people that used AI
- 17%
- Cetic.br, 2025 (interviews from Feb 2025 to Jan 2026)
Each number has, in the table of its section, the sample, the period, the source with a link and the day it was checked.
How long does the AI take to answer on WhatsApp?
Half of the AI's replies went out within 12.2 seconds; a team member's, within 2.2 minutes, 11 times longer.
- AI (Jul 30 to Sep 24, 2026)
- Team member (Aug 26 to Sep 24, 2026)
Half of the replies went out within
9 in 10 replies went out within
Numbers from all companies that use CertuAI, added together. The comparison with the team is not fair one to one: a person only enters the count when the AI did not answer first.
The numbers, with sources
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| Half of the AI's replies went out withinTime between the customer's last message and the AI's reply. Each reply counts once. Waits of more than 10 minutes are left out. | 12.2 seconds | Jul 30 to Sep 24, 202632,230 replies from 57 accounts | CertuAIchecked on |
| 9 in 10 AI replies went out withinSame count as above: 9 in 10 replies went out in that time or sooner. | 21.1 seconds | Jul 30 to Sep 24, 202632,230 replies from 57 accounts | CertuAIchecked on |
| Half of a team member's replies went out withinTime between the customer's first message still unanswered and a team member's reply, in the same conversation. It only counts when the AI did not answer first: the conversation was with the team or the person was faster. Note: This is not a fair, one-to-one comparison with the AI: the person only enters the count in conversations where the AI did not answer. The measurement starts on Aug 26, 2026, when the system began tagging which team member replied. | 2.2 minutes | Aug 26 to Sep 24, 20269,639 replies from 37 accounts | CertuAIchecked on |
| 9 in 10 of a team member's replies went out withinSame count as above. That is almost 2 hours and 20 minutes. | 139.9 minutes | Aug 26 to Sep 24, 20269,639 replies from 37 accounts | CertuAIchecked on |
| AI replies that handed the conversation to a team memberTimes the AI called someone on the team, divided by the total number of AI replies. Counted per reply, not per conversation. | 7,9% | Jul 30 to Sep 24, 202637,951 replies from 57 accounts | CertuAIchecked on |
| Replies the AI sent on WhatsAppCounted by the system that answers: one reply each time the customer writes. | 37.951 | Jul 30 to Sep 24, 202657 accounts | CertuAIchecked on |
| Messages customers sent through WhatsAppNo imported history, no test contacts and without CertuAI's sales chats. | 101.038 | Jul 30 to Sep 24, 202662 accounts | CertuAIchecked on |
| Customer conversationsA customer's messages count as one conversation until they go more than 24 hours without writing. | 16.404 | Jul 30 to Sep 24, 202662 accounts | CertuAIchecked on |
| Customer messages that arrived outside business hoursBusiness hours are Monday to Friday, 8 a.m. to 6 p.m., in the account's time zone (the default is Brasília time). | 32,5% | Jul 30 to Sep 30, 2026122,720 messages from 67 businesses | CertuAIchecked on |
| How much each AI reply takes from the plan's balanceEach reply the AI sends on a paid plan takes this much from the month's balance. The same balance pays for images, campaigns and sites. | 120 credits | since Sep 23, 2026system rule, not a measurement | CertuAIchecked on |
Previous measurement, 7 days up to Aug 25, 2026, with 1,058 replies: half within 15.4 seconds. Both hold for their own period. The newer measurement, from Jul 30 to Sep 30, 2026, is in the CertuAI WhatsApp Customer Service Index.
How many companies and people use AI in Brazil?
Sebrae's survey asks in a different way and counts even the individual microentrepreneur (MEI): read its row before comparing with Cetic.br.
Brazilian companies that used AI, by size
- 2024
- 2025
The numbers, with sources
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| Brazilian companies with 10 or more people that used AICetic.br is the study center of the committee that oversees the internet in Brazil. It counts a company that used at least one AI technology. | 17% | 2025 (interviews from Feb 2025 to Jan 2026)companies with 10 or more people, sample survey | Cetic.brchecked on |
| The same measure, one year earlierSame question and same type of company as the row above. | 13% | 2024companies with 10 or more people, sample survey | Cetic.brchecked on |
| Companies with 10 to 49 people that used AIThe smallest companies in the survey. Businesses with fewer than 10 people are not included. | 15% | 2025 (interviews from Feb 2025 to Jan 2026)companies with 10 to 49 people, sample survey | Cetic.brchecked on |
| Companies with 10 to 49 people that used AI, one year earlierSame question as the row above, in the previous edition. | 10% | 2024companies with 10 to 49 people, sample survey | Cetic.brchecked on |
| Companies with 50 to 249 people that used AICounts a company that used at least one AI technology. | 32% | 2025 (interviews from Feb 2025 to Jan 2026)companies with 50 to 249 people, sample survey | Cetic.brchecked on |
| Companies with 250 people or more that used AICounts a company that used at least one AI technology. | 50% | 2025 (interviews from Feb 2025 to Jan 2026)companies with 250 people or more, sample survey | Cetic.brchecked on |
| People who use the internet in Brazil and have used AI that writes and creates (like ChatGPT)First time the survey asked this. It can't be compared with the previous year yet. | 32% | 2025 (interviews from Mar to Aug 2025)people aged 10 or older who use the internet, sample survey | Cetic.brchecked on |
| Brazilians who have used AI that writes and createsEstimate for the population aged 10 or older. The round number is Cetic.br's own; the exact percentage is in the row above. | 50 million people | 2025 (interviews from Mar to Aug 2025)24,535 people interviewed in 27,177 households | Cetic.brchecked on |
| Industrial companies with 100 people or more that used AICompanies that said they used AI in the year, divided by the total number of industrial companies of that size. Calculation by CertuAI with the official table. | 16,9% | 2022industrial companies with 100 people or more, sample survey | IBGEchecked on |
| Industrial companies with 100 people or more that used AI, two years laterCalculation by CertuAI with the official table: 4,261 divided by 10,167. It does not cover retail, services or companies with fewer than 100 people. | 41,9% | 20244,261 of 10,167 industrial companies with 100 people or more (numbers estimated by the survey) | IBGEchecked on |
| Small businesses that used AI in the 2 weeks before the interviewThe owner says whether they used AI in the business. The question copies one from a United States government survey, which found 21% there. Note: It is the owner who reports, and it counts even free tools. That is why the number is much higher than Cetic.br's, which asks in a different way and only hears from companies with 10 people or more. | 52% | interviews from Mar 24 to May 8, 20267,182 small business owners, from the individual microentrepreneur (MEI) to the small company | Sebrae and Meta Research Institutechecked on |
How does Brazil compare with the world?
The European Union survey is similar to Cetic.br's (17% in Brazil in 2025), but it does not measure exactly the same thing.
Companies that used AI
The numbers, with sources
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| Private money invested in AI companies worldwideInvestment by private funds and investors in AI companies. Up 127.5% over 2024, according to the report. | US$ 344.7 billion | 2025compilation of the year's investments | Stanford Universitychecked on |
| How much the price of an AI at GPT-3.5 level fellThe price per 1 million tokens went from US$ 20 to US$ 0.07. Note: It compares the cheapest model that scores the same on each date, not the same model getting cheaper. | 280 times cheaper | Nov 2022 to Oct 2024the cheapest model that passes the same test | Stanford Universitychecked on |
| Notable AI models released by institutions in the United StatesNotable is an editorial criterion from Epoch AI: influence, citations and being among the best of the time. | 59 models | 2025Epoch AI model database | Stanford Universitychecked on |
| Notable AI models released by institutions in ChinaSame criterion as the row above. | 35 models | 2025Epoch AI model database | Stanford Universitychecked on |
| European Union companies with 10 or more people that used AICounts a company that used at least one AI technology from a fixed list (reading text, recognizing speech, generating text, recognizing images, and others). Note: It can be set next to Cetic.br's 17%, but it is not the same measure: the list of technologies and the sectors counted are similar, not identical. | 19,95% | 2025annual survey in the 27 countries, companies with 10 or more people | Eurostatchecked on |
| United States companies that used AI to produce their goods or servicesThe company says whether it used AI in the 2 weeks before the interview. Between Dec 2025 and May 2026 the number stayed between 17% and 20%. | 19,8% | cycle ended May 3, 2026sample of companies with employees, weighted for the whole country | United States Census Bureauchecked on |
How much does it cost to use AI, and how much has the price fallen?
A token is the piece of text the AI counts in order to charge, roughly a piece of a word. Input is what you send the AI to read; output is what it writes.
How much the entry price fell
- before
- after
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| OpenAI's GPT-4, first version with a date in the tableList price in dollars per 1 million tokens. Input is the text the AI reads; output is what it writes. | US$ 30 (input) and US$ 60 (output) | Jun 13, 2023price table, not a measurement | OpenAIchecked on |
| OpenAI's GPT-5List price in dollars per 1 million tokens. | US$ 1.25 (input) and US$ 10 (output) | Aug 7, 2025price table, not a measurement | OpenAIchecked on |
| How much the input price fell from GPT-4 to GPT-5From US$ 30 to US$ 1.25 per 1 million input tokens: 1 minus 1.25 divided by 30. It became 24 times smaller. Note: They are different models, and GPT-5 does more than GPT-4. | 95.8% drop | Jun 2023 to Aug 2025calculation on the price table | CertuAI's calculation on OpenAI's official tablechecked on |
| Anthropic's Claude 3 Opus, at launchLaunch price in dollars per 1 million tokens. | US$ 15 (input) and US$ 75 (output) | Mar 4, 2024price table, not a measurement | Anthropicchecked on |
| Anthropic's Claude Opus 5.5, today's priceList price in dollars per 1 million tokens. Note: The newest Opus models count about 30% more tokens for the same text, so the drop per text is smaller than the drop per token. | US$ 4 (input) and US$ 20 (output) | in effect on Sep 25, 2026price table, not a measurement | Anthropicchecked on |
| How much the input price of Anthropic's Opus line fellFrom US$ 15 to US$ 4 per 1 million input tokens: 1 minus 4 divided by 15. | 73.3% drop | Mar 2024 to Sep 2026calculation on the price table | CertuAI's calculation on Anthropic's official priceschecked on |
All 94 models with an official price, today
R$ 0,18the cheapest, Amazon Nova Micro
R$ 6,26the median
R$ 751,44the most expensive, o1-pro
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| AI chat models in CertuAI's catalogText models only, with data taken from the company's official page or from Hugging Face, where companies publish open models. | 262 models from 27 companies | prices as of Oct 13, 2026262 models from 27 companies | CertuAI AI model catalogchecked on |
| Models with an official price published by the companyAPI list price, no discount. 6 of them also have open weights. | 94 of 262 | prices as of Oct 13, 2026262 models from 27 companies | CertuAI AI model catalogchecked on |
| Open-only models, with no official priceAn open model is one anyone can download and run. It has no list price: the cost depends on where it runs. | 168 of 262 | prices as of Oct 13, 2026262 models from 27 companies | Hugging Facechecked on |
| Lowest price per 1 million input tokens (Amazon Nova Micro)A token is the piece of text the AI counts in order to charge, roughly a piece of a word. Input is what you send the AI to read. On the official page: US$ 0.035. | R$ 0,18 | prices as of Oct 13, 202694 models with an official price | Amazon's official pricing pagechecked on |
| Highest price per 1 million input tokens (OpenAI o1-pro)Same calculation as the row above. On the official page: US$ 150. | R$ 751,44 | prices as of Oct 13, 202694 models with an official price | OpenAI's official pricing pagechecked on |
| Lowest price per 1 million output tokens (Mistral AI Ministral 3 (3B))Output is the text the AI writes. On the official page: US$ 0.1. | R$ 0,50 | prices as of Oct 13, 202694 models with an official price | Mistral AI's official pricing pagechecked on |
| Highest price per 1 million output tokens (OpenAI o1-pro)Same calculation as the row above. On the official page: US$ 600. | R$ 3.005,76 | prices as of Oct 13, 202694 models with an official price | OpenAI's official pricing pagechecked on |
The price comes in dollars on each company's page and was converted to reais at the official rate of the previous day. All the models, from cheapest to most expensive, are in the AI price table in reais; model by model, with a calculator, at AI models and their prices.
What is the best AI model today?
It depends on who measures: LMArena collects votes from people and has no Portuguese category; Prosa uses only real conversations in Brazilian Portuguese. The method is in each one's row.
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| First place in LMArena's text ranking: claude-fable-5 (Anthropic)People compare two answers without knowing which AI they come from and vote for the better one. The score comes from those votes. A small difference between two models can be a tie. Note: It measures the taste of the people who vote on the site, mostly in English, and not whether the answer is correct. There is no Portuguese category. | 1,505.7 points | cutoff Sep 13, 202630,057 votes on this model | LMArenachecked on |
| Best open model in the same ranking: kimi-k3-max (Moonshot), 17th placeAn open model is one anyone can download and run on their own computer. | 1,484.8 points | cutoff Sep 13, 202620,987 votes on this model | LMArenachecked on |
| People's votes in LMArena's text rankingEach vote is a blind comparison between two answers. | 8,146,274 votes | cutoff Sep 13, 2026all the votes in the ranking | LMArenachecked on |
| Models in LMArena's text rankingHow many models have a score in the ranking. | 402 models | cutoff Sep 13, 2026all the models in the ranking | LMArenachecked on |
| First place in the Prosa test, of real conversations in Brazilian Portuguese: GPT-5.2The score is how many of the criteria for a good answer the model met. The scores are given by other AIs. Note: The test was made by Maritaca AI, which also makes the Sabiá model. Only the 16 models from May 2026 are included. | 88,6% | paper of May 2, 2026981 real conversations and 16 models | Maritaca AIchecked on |
| Sabiá-4, Maritaca AI's Brazilian model, on the same test (9th of 16)Same score as the row above. | 70,2% | paper of May 2, 2026981 real conversations and 16 models | Maritaca AIchecked on |
How much does AI on WhatsApp cost with Meta's new charge?
From Oct 1, 2026, Meta will charge for replies to customers within the 24 hours.
How much Meta charges per delivered message
The numbers, with sources
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| Reply to the customer within 24 hours after their message (service message)Meta's list price per message delivered to a Brazilian number. During this period Meta does not charge for the service message. | R$ 0 (free) | from Jul 1, 2026 to Sep 30, 2026price table, not a measurement | Metachecked on |
| Free service messages per business number, per monthReplies within the 24 hours the customer opened. The 1,000 free are per number and do not carry over to the next month. | 1,000 messages | from Oct 1, 2026price table, not a measurement | Metachecked on |
| Each service message after the month's 1,000 freeMeta's list price per message delivered to a Brazilian number, before tax. The company that connects WhatsApp may charge more on top. | R$ 0,035 | from Oct 1, 2026price table, not a measurement | Metachecked on |
| What Meta charges for 10 thousand replies in the month, on a single number(10,000 minus 1,000 free) times R$ 0.035 = R$ 315. Note: This is only Meta's part: it does not include the cost of the AI, of the company that connects WhatsApp or tax. | R$ 315.00 per month | from Oct 1, 2026calculation on the price table | CertuAI's calculation on Meta's official tablechecked on |
| Utility message (order notice, appointment reminder)Meta's list price per message delivered to a Brazilian number, before tax and before the volume discount. | R$ 0,035 | since Jul 1, 2026price table, not a measurement | Metachecked on |
| Marketing message (advertising, promotion)Meta's list price per message delivered to a Brazilian number, before tax. | R$ 0,3217 | since Jul 1, 2026price table, not a measurement | Metachecked on |
| Meta One Essencial subscription, the cheapest plan for businessesMonthly price "starting at" in Brazil, according to Meta. The business plans include more use of Meta's AI agent. Note: Meta does not say how many conversations each plan covers. | R$ 59.00 per month | announced Sep 15, 2026announced price, not a measurement | Metachecked on |
| Companies already using Meta's AI agent on WhatsApp and MessengerWorldwide number, with Brazil not broken out. | more than 1 million companies | as of Jun 3, 2026number stated by Meta | Metachecked on |
| People with a cell phone who were served by a chatbot in the last 12 monthsInternet survey of people who have a cell phone, not of the whole population. | 79,4% | March 2026 (interviews from Mar 4 to 24, 2026)4,138 people with a cell phone, aged 16 or older | Mobile Time and Opinion Boxchecked on |
| Average score people give to chatbot customer serviceAverage of scores from 0 to 10. By age: 6.2 for ages 16 to 29, 5.7 for 30 to 49 and 5.0 for 50 or older. | 5.6 out of 10 | March 2026 (interviews from Mar 4 to 24, 2026)4,138 people with a cell phone, aged 16 or older | Mobile Time and Opinion Boxchecked on |
| People who use WhatsApp and talked to a company through the app in the last 12 monthsTalked to a brand or company on WhatsApp at least once in 12 months. | 79,5% | March 2026 (interviews from Mar 4 to 24, 2026)4,138 people with a cell phone, aged 16 or older | Mobile Time and Opinion Boxchecked on |
The math: with 10 thousand replies in a month, Meta charges (10,000 minus 1,000) times R$ 0.035, which is R$ 315.00 per month. The first 1,000 messages of the month are free.
Where does the AI law stand in Brazil?
The AI bill (PL 2338/2023) has not become law yet.
- Bill 2338/2023, the AI Legal Framework, introduced in the Senate
- Bill approved on the Senate floor
- Bill reached the Chamber of Deputies
- ANPD publishes the method for testing AI in this environment
- Status of the bill in the Chamber of Deputies: awaiting an opinion
The numbers, with sources
| Data point | Value | Period and sample | Source |
|---|---|---|---|
| Bill 2338/2023, the AI Legal Framework, introduced in the SenateDate the bill was introduced. | 03/05/2023 | May 3, 2023official record | Federal Senatechecked on |
| Bill approved on the Senate floorAfter that, the bill went to the Chamber of Deputies. | 10/12/2024 | Dec 10, 2024official record | Federal Senatechecked on |
| Bill reached the Chamber of DeputiesArrival date recorded by the Chamber. | 17/03/2025 | Mar 17, 2025official record | Chamber of Deputieschecked on |
| Status of the bill in the Chamber of DeputiesThe bill has not become law yet. The status changes with each step in the Chamber. | awaiting an opinion | last movement on Sep 2, 2026official record | Chamber of Deputieschecked on |
| Companies testing AI under the supervision of the National Data Protection Authority (ANPD)It is ANPD's test environment (called a sandbox), which checks whether the company explains well how the AI decides. It began in March 2026. | 3 companies | 2026 (report of Jul 2, 2026)number stated by ANPD | National Data Protection Authoritychecked on |
| ANPD publishes the method for testing AI in this environmentMethod to assess whether the AI is transparent, whether what it does can be explained and whether it protects personal data. | 19/08/2026 | Aug 19, 2026official record | National Data Protection Authoritychecked on |
How were the numbers measured?
See the method, source by source
CertuAI's numbers: come from the system itself, counting the messages and replies of 62 accounts of companies that use CertuAI, from Jul 30 to Sep 24, 2026. Nobody read a conversation: the program only adds up and measures time. Left out are the history imported from before the account, test contacts and CertuAI's sales chats.
- Conversation: messages from the same customer until they go more than 24 hours without writing.
- Half and 9 in 10: put all the replies in a line, from fastest to slowest. "Half within" is the time of the reply in the middle of the line. "9 in 10 within" is the time of the reply that comes after 90% of the line.
- Business hours: Monday to Friday, 8 a.m. to 6 p.m., in the account's time zone.
- Limit: a single account reaches 30.4% of the total messages. By type of business the sample is still small, which is why we do not publish a number by type.
Numbers from other sources: copied as the source published them, with the link and the day we checked. When the row says "CertuAI's calculation", the calculation is written in it. The full name of each source appears when you hover over the link and is in the files to download.
- AI use in Brazil and worldwide: Cetic.br (TIC Enterprises and TIC Households surveys), IBGE, Sebrae, Stanford University, Eurostat and the United States Census Bureau. Most publish once a year; the US Census, every two weeks.
- Prices: each company's official pricing page, and Hugging Face for open models, which have no list price. It is the list price, with no volume discount.
- Rankings: LMArena, which changes every day with new votes, and the Prosa test, published in May 2026.
- WhatsApp: Meta's official price table, which usually changes at the start of each quarter, and Meta's press room in Brazil.
- Law: open data from the Chamber of Deputies and the Senate and news from ANPD.
Left out are numbers we could not find again on the source's page, surveys that require sign-up to read and surveys we did not recheck. The page is reviewed when a source publishes a new number; the "checked on" under each source says when the row was checked.
How to cite this data?
Certu, Dados de IA no atendimento pelo WhatsApp no Brasil, set/2026, certu.com.br/dadosCreative Commons Attribution license, version 4.0 (CC BY 4.0): you can use and republish, including in articles and papers, as long as you cite the source. For a number from another source (Meta, a survey, the model catalog), also cite the original source on its row.
Where to download the data?
All the numbers on this page, with full source, period, sample and date. The CSV opens with accents displayed correctly in Excel.
When was this data updated?
Last updated: . Each number also has its own date, in the "checked on" under the source. The quarterly data on when customers write and how fast the AI answers, by line of business, is in the CertuAI WhatsApp Customer Service Index. CertuAI also runs an open survey of small business owners about WhatsApp customer service, at /pesquisa.
Frequently asked questions
- Can I use these numbers in an article or a paper?
- Yes. The license is CC BY 4.0: just cite "CertuAI, AI data in WhatsApp customer service in Brazil, Sep 2026, certu.com.br/dados". For a number from another source, also cite the original source.
- Does the data show any company or customer?
- No. CertuAI's numbers are sums of all accounts together. No company, customer or conversation appears.
- Why is there no number by type of business?
- Because the sample by type is still small: in each one, a single account sends a good share of the messages. A number like that shows the account, not the type of business.
In short
At companies that use CertuAI, half of the AI's WhatsApp replies went out within 12.2 seconds, and 7.9% of them handed the conversation to a person on the team (Jul 30 to Sep 24, 2026). In Brazil, 17% of companies with 10 or more people used AI in 2025 (Cetic.br).
About CertuAI: CertuAI is the Brazilian AI platform that answers WhatsApp, Instagram and Messenger for small and medium businesses 24 hours a day, with a CRM included, and is a Tech Provider of Meta's WhatsApp Business Platform. Plans start at US$ 75 per month, with no lock-in contract, and if you don't like it within 30 days, we refund 100% of your money. Create my agent
