Half of Brazilian companies already use AI, but only 21% say they're ready for agents

What AWS announced at the São Paulo Summit

On September 3, 2026, at AWS Summit São Paulo, AWS presented the study "Unlocking the Potential of AI in Brazil 2026". The headline that ran across the tech news sites: 50% of Brazilian companies, roughly 12 million businesses, have already adopted artificial intelligence. Baguete describes this as a 40% increase over the previous year.

Half of the country's businesses using AI sounds like a lot. And it is, if we accept the word "using" the way the survey accepted it. Further down you'll see that the same term, "adoption", shows up in three different surveys with very different results, and that this isn't a contradiction, it's a difference in the question being asked.

The figure that matters most to anyone running a small business comes right after the headline: only 21% of companies say they are ready for technologies like agentic AI, the kind where the software doesn't just answer but carries out multi-step tasks. In other words, many people have started, but few feel prepared for the next step up.

In this article we separate what the survey says, what it doesn't say, how it compares with other studies, and what you, as a business owner, can do with it without having to become an expert.

Who ran the survey, and why does that matter for reading the numbers?

The study was commissioned by AWS itself. That doesn't invalidate anything, but it changes how to read it: this is not an independent survey, it's research from a company that sells cloud infrastructure and AI services, released at that company's own event.

Another point: in the three news reports we consulted (Baguete, Convergência Digital and Teletime), none gives the sample size or the methodology. We don't know how many companies responded, of what size, in which sectors, or how the question was asked. Without that, the numbers can be used as an indication, but not as an exact measure.

Vendor-commissioned research is often useful, which is why we cite it here. But it tends to highlight what suits the vendor's narrative: adoption growing, investment rising, a readiness gap the vendor itself can help close. Read it with that filter and compare it with other sources, as we'll do below.

One data point from AWS's side illustrates the context: the company plans around US$220 billion in global AI investment in 2026 and wants to train 2 million people in Latin America by 2028, 1 million of them in Brazil. The Summit gathered 18,000 attendees. In short, AWS has a lot of interest in seeing the Brazilian market move forward.

What we know and what we don't

We know the name of the study, who commissioned it, who presented it, when, and the percentages released. We don't know the sample size, the methodology, the breakdown by sector, or the exact criterion for classifying a company as an "adopter". Keep this list in mind, because it weighs on the comparison with the other surveys.

What do the three levels of AI maturity mean?

The survey divides companies into three levels: 58% are at the basic level, 27% at intermediate and 15% at advanced. The advanced level rose from 12% to 15% in one year. Agentic AI sits within that 15% group. In other words, most companies are still at the bottom of the ladder.

Basic: 58%

This is use for gradual efficiency gains, almost always with chatbots and off-the-shelf tools. Think of the support agent who uses an AI assistant to draft replies, or the finance person who asks for a report summary. Useful and cheap to start, but usually loose: each person uses it their own way, with no process.

Intermediate: 27%

Here AI is already built into the company's products or services. It's no longer just a tool an employee opens in a browser tab: it's part of what the company delivers to customers. This takes more planning, integration with systems, and someone responsible for keeping it running.

Advanced: 15%

This is the group at the cutting edge, and where the survey places agentic AI. An AI agent receives a goal and carries out a sequence of steps to reach it, instead of answering one question at a time. Only 15% of companies are here, which helps explain why only 21% say they're ready for this kind of technology.

And agents, how many companies already use them?

This calls for care. Only Convergência Digital gives a specific percentage: according to the outlet, 6% already use AI agents. Teletime, which covered the same event, notes that no separate percentage for agents was released. Since it's a single source, we treat the 6% as information to be confirmed, not as a firm figure.

What do the investment and productivity numbers indicate?

Among the companies surveyed, 68% increased their AI investment in the past year, and 72% of those that adopted it report productivity gains. In addition, 79% expect AI to drive growth over the next 12 months, and 76% rate AI as a high or top strategic priority.

These are encouraging numbers, and at the same time largely self-reported. "Productivity gain" here is the perception of the people who answered, not a measurement of hours or revenue. That's normal in opinion surveys, but it helps explain a number that appears further on and that almost nobody comments on.

That number is this: only about a third of companies say they trust their own ability to measure the financial return on AI. Stop and put the pieces together. More than two-thirds say they gained productivity, but only about 33% believe they can measure the return. Many people feel it works, few can prove by how much.

Why this matters for a small business

When money is tight, investing in a tool you can't evaluate is a double risk: you pay and you don't know if it was worth it. For a large company, that gets diluted. For a small one, a forgotten subscription or a project with no clear result weighs on the month's cash flow.

Why do only 21% feel ready for agents?

The main obstacle named is people: 48% of companies cite a lack of digital and AI skills. Only 21% feel ready for technologies like agents, and only about a third are confident they can measure the return on AI. The problem, then, seems to lie less in the tool and more in who is going to use it.

Baguete reinforces this reading with another number: 72% agree that AI tools are available. In other words, access is not the bottleneck. Any company can subscribe to an AI assistant today. What's missing is knowing how to apply it properly: for what, with what care, and with what expected result.

It makes sense. Think back to when you bought your first management system. Having the software solved nothing until someone understood the company's process, entered the data correctly and trained the team. AI works the same way, except the tool looks so easy to open that we underestimate the work that comes afterward.

What "being ready" means in practice

The survey doesn't detail the criterion, but we can translate it into questions you can answer yourself. Does anyone in the company understand what AI can and can't do? Are the processes clear enough to be automated? Is there a way to know whether it worked? If the answer to any of them is "I don't know", you're in the same boat as most.

What about startups?

Baguete notes that startups show more advanced adoption than small and medium-sized companies and than large ones. That fits the idea that the bottleneck is one of habit and skill: companies born in a digital environment tend to have less friction when it comes to experimenting and adjusting.

Why do surveys give such different numbers?

Because each one measures something different. AWS speaks of 50% of companies with AI adopted. Cetic.br, which is more structured, reports 17% adoption among all companies. Sebrae with Meta finds 52% of small businesses using AI. All three use the word "adoption", but they ask different things.

Cetic.br: structural adoption, 17%

Cetic.br's TIC Empresas 2025 survey, with a page dated June 2026, interviewed 4,174 companies by phone between February 2025 and January 2026. AI adoption among all companies went from 13% to 17%. Among small ones (10 to 49 employees), it went from 10% to 15%. Among large ones, from 38% to 50%.

The survey also shows how companies use AI. Among those that use it, natural language generation, such as text and conversations, rose from 20% to 30%. In addition, 80% of adopters bought off-the-shelf software and 60% hired external vendors. In other words, those using AI are buying ready-made solutions and asking for outside help, which matches the lack of in-house skills cited by AWS.

Sebrae and Meta: recent use, 52%

The survey "Digital Transformation in Small Businesses 2026", released on August 17, 2026, interviewed 7,182 small businesses between March 24 and May 8. The result: 52% used AI in the two weeks before the interview. In the United States, the comparable figure is 21%. The intention to adopt or expand use reaches 60% in Brazil and 24% in the US.

Notice the yardstick: "used it in the last two weeks". Opening an AI assistant once to answer a customer is enough to count. It isn't necessarily structured adoption, with a process and someone in charge. It's a snapshot of habit, and that's why the number is high.

AWS: vendor survey, 50%

AWS's 50% comes from a commissioned survey with no published methodology. We can't say whether the question was about structured use, occasional use or intention. We also don't know which companies were in the base. That's why we can't confidently put it side by side with the other two.

Side by side

| Survey | Who measures | Question | Result | |---|---|---|---| | AWS, "Unlocking the Potential of AI in Brazil 2026" | Study commissioned by AWS | Companies that adopted AI (criterion not disclosed) | 50% | | Cetic.br, TIC Empresas 2025 | 4,174 companies, by phone | AI adoption by companies | 17% (all); 15% (small); 50% (large) | | Sebrae and Meta, 2026 | 7,182 small businesses | Used AI in the previous two weeks | 52% |

So which number is right?

Probably all of them, within their own yardstick. What changes is what each one sees. A curiosity: Cetic's 50% for large companies matches AWS's 50% for the total, but these are different groups and methods, so they can't be treated as the same measurement.

The gap between 17% and 52% is nobody's mistake. It's the distance between "having structured AI in the company" and "someone has used AI recently". In our reading, a good share of small companies sits in that space in the middle.

In practice, here's what this means for you: the "half already use it" figure is a snapshot of habit, not of maturity. It's quite possible that a good part of the use is employees using a free tool on their own, without the company having decided anything. That leads us to the next point.

What does the Sebrae survey show about how small businesses use AI?

According to Sebrae and Meta, 28% of small businesses use AI to improve tasks employees already did, and 19% to create new functions. In addition, 90% did not change their headcount in the six months before adopting it. The use, then, is more about supporting what exists than restructuring.

This picture matches the basic level in the AWS survey. AI comes in as a helper: it speeds up a text, organizes information, answers a question. Almost nobody is firing or hiring because of it, at least at this first stage. That eases a common fear among business owners, that adopting AI requires rebuilding the whole team.

On the side of those who don't yet plan to use it, the reason is also similar to AWS's: 38% of those who don't plan to adopt cite a lack of knowledge about what AI can do. It's not a lack of money or access, it's a lack of reference points. If you recognize yourself in this group, know that you are far from an exception.

What this says about the next step up

If 28% improve existing tasks and 19% create new functions, the natural path is to start with the first and only then move to the second. Creating a new function with AI, such as automated support or a new feature in your product, requires a good understanding of what already works today. Whoever skips this stage tends to automate a mess.

Which step of the ladder are you on? A quick checklist

Answer honestly, just for yourself. You don't have to show it to anyone. The idea is to place your company within the survey's three levels and find the next realistic step, without comparing yourself to a headline.

Signs you're at the basic level

• Employees use AI tools on their own, with no agreed rules.
• You can't say how much the company spends on AI subscriptions.
• Nobody has measured how much time AI saves, it just "seems to help".
• There's no one responsible for deciding which tools come in.

If you ticked two or more, you're at the basic level, along with 58% of the companies in the AWS survey. It's not a problem, it's the most common starting point.

Signs you're moving toward intermediate

• There is at least one company process with AI in the middle, and someone looks after it.
• AI shows up in what you deliver to customers, not just behind the scenes.
• You can say, at least roughly, what improved after it.
• There is a simple rule about which data can or can't be typed into the tools.

Signs you're close to advanced

• You already have well-described processes that an agent could run in steps.
• Someone with AI skills, in-house or hired, tracks the results.
• You measure the return with some concrete number, even a simple one.
• AI errors have a clear review path before they reach the customer.

Few small companies have all of this, and that's fine. What matters is knowing where you're starting from so you can pick the right next step.

How do you measure the return on AI before spending more?

If only about a third of companies are confident they can measure the return on AI, starting there already puts you ahead. Pick a task, write down how long it took before, use AI for a few weeks and compare. Without that before and after, any conclusion is a guess.

A hypothetical example

This is a hypothetical example, only to illustrate the reasoning. Imagine a store that answers customer questions by message. Before using AI, the owner notes that each reply takes a certain amount of time and that the attendant spends part of the day on it. After a few weeks using an assistant to draft replies, he repeats the notes.

If the time dropped, great: you can estimate what that represents in freed-up hours. If it didn't, or if reviewing the replies consumed what was gained, he found out early and cheaply that this use isn't worth the subscription. In both cases he decides with information, not with a feeling.

What to write down, simply

• The chosen task and who does it.
• The time or cost before AI.
• The time or cost after, including human review.
• The cost of the tool over the period.
• Whether the quality of the result got worse, stayed the same or improved.

A spreadsheet with five columns does the job. No sophisticated dashboard needed. The point is to turn "it seems to help" into something you can show a partner, your accountant, or yourself six months from now.

Which AI uses are safest to start with?

The best first use of AI is one where a mistake costs little and a human reviews it before anything leaves the company. Think of internal, repetitive tasks that are easy to check. We have no survey number for this: it's common-sense guidance, based on what usually works.

Low-risk ideas

• Drafting emails and messages that a person reviews before sending.
• Summarizing meetings, long documents and reports for internal use.
• Organizing messy lists, tables and notes.
• Suggesting text ideas for posts and product descriptions, with human editing.
• Putting together a first draft of a procedure or internal manual.

Uses that call for more care

• Anything involving customer data, contracts or financial information.
• Automatic replies sent straight to the customer, without review.
• Pricing, credit or hiring decisions based only on the tool's answer.
• Agents that take actions in real systems, such as sending, paying or deleting.

These aren't forbidden. They just call for process, review and, preferably, a small test first. And this is where the topic of agents comes back: they do things on their own, so the cost of a mistake is higher. It makes sense that most don't feel ready.

What are the most common mistakes when adopting AI?

The most common mistakes aren't technical. They're about order: buying the tool before understanding the process, not measuring the result, leaving each employee to fend for themselves, and jumping straight to agents without having the basics organized. Almost all of them are solved with a conversation and a spreadsheet, not with more technology.

1. Buying the tool before designing the process

Whoever automates a confusing process gets a faster confusion. Before choosing software, write down in simple steps how the task is done today. If you can't describe it, the AI won't be able to do it properly either.

2. Not measuring the return

This is the mistake the AWS survey itself lets show through: many people perceive a gain, few can quantify it. Without measurement, you don't know when it's worth investing more and when it's time to cancel.

3. Leaving use unstructured

Each person using a different tool, with different data, with no rules. This is the basic level in its most fragile form. A one-page policy saying what can and can't be typed in already helps a lot.

4. Ignoring the team's skills

If 48% point to a lack of skills as a barrier, it's because a tool without training yields little. Set aside time for the team to learn to ask well, to review the answer, and to be suspicious of what looks too right.

5. Wanting to jump straight to agents

Agents are the highest step on the ladder. They work best when the process is already clear, the data is organized and someone is monitoring. Without that, the agent just repeats your problems at higher speed.

6. Comparing yourself to the headline

"Half of companies use AI" can make you think you've fallen behind. As we saw, it depends on the yardstick: by Cetic.br's, adoption among all companies is 17%, and among small ones it's 15%. You're probably in better company than you think.

If I don't use AI yet, should I be worried?

Not necessarily. According to Sebrae and Meta, 38% of those who don't plan to adopt AI say they lack knowledge about what it can do. That's a legitimate and solvable reason. The bigger risk isn't taking a little longer, it's adopting without criteria just because everyone seems to be doing it.

At the same time, the intention to adopt or expand use reaches 60% among the Brazilian small businesses surveyed by Sebrae and Meta. The trend is clear. So the useful question isn't "should I use it?", but "where does it make sense, at what pace and with what care for my business?".

A simple path, in order: take stock of what is already happening informally in the company; pick one low-risk task; measure before and after; set a basic rule of use; and only then think about expanding. Each step is small, and together they take you from "it seems to help" to "I know how much it helps".

Quick questions

Do half of Brazilian companies really use AI?

It depends on the survey. AWS says 50%, but the survey was commissioned by AWS and doesn't release the sample or methodology. Cetic.br reports 17% adoption among all companies. Sebrae and Meta find 52% among small businesses, measuring use in the two weeks before the interview.

Is the AWS survey independent?

No. The study was commissioned by AWS. That doesn't make it false, but it's a vendor survey presented at the vendor's own event. It's worth reading alongside sources like Cetic.br, which discloses the number of companies surveyed and the period.

What is agentic AI?

It's AI that receives a goal and carries out a sequence of steps to reach it, instead of just answering one question at a time. In the AWS survey, it falls within the advanced level, which gathers 15% of companies. Only 21% say they're ready for this kind of technology.

How many companies already use AI agents?

There's no firm number. Convergência Digital says 6% already use agents, but it's the only source with that percentage. Teletime, which covered the same event, notes that no separate percentage was released. We treat the 6% as a figure to be confirmed.

Where should I start if I've never used AI?

Pick an internal task that's repetitive and easy to review, such as drafting messages or summarizing documents. Write down how long it took before, use the tool for a few weeks and compare. Avoid entering sensitive customer data and set a simple rule about what can be typed in.

Do I need to hire a specialist to use AI?

Not to start. For the basic level, a curious person on the team is enough. But a lack of skills is the most cited barrier in the AWS survey (48%), and Cetic.br shows that 60% of adopters hired external vendors. Asking for help at some point is common.

Where Diglion comes in

Diglion helps small and medium-sized businesses move from loose AI use to something they can measure. In practice, that means mapping where the tool is already being used, choosing one or two low-risk tasks, setting up a simple way to measure before and after, defining a one-page rule of use and training the team, before thinking about agents.

Sources consulted

• Baguete, 50% das empresas brasileiras já adotaram IA, retrieved 2026-09-08.
• Convergência Digital, Doze milhões de empresas usam IA no Brasil, mas somente 6% já usam agentes de IA, retrieved 2026-09-08.
• Teletime, AWS e IA agêntica, retrieved 2026-09-08.
• Cetic.br, Uso de inteligência artificial por empresas brasileiras avança e atinge 17%, retrieved 2026-09-08.
• Brasil em Folhas, Sebrae aponta que 52% de pequenos negócios usam IA, retrieved 2026-09-08.