AI for small business: where to actually start, in order, without burning money

What you get from this article

• According to Sebrae, 52% of small business owners in Brazil had used AI in the two weeks before the 2026 survey.
• The first task should be repetitive, with an answer that already exists in writing and mistakes that are cheap to spot.
• Measure before automating: one week noting where the time goes solves more than any tool.
• Start in suggestion mode, with a person reviewing, and only then increase autonomy.
• The real cost includes the time of whoever organizes and reviews, not just the subscription.
• Customer service with full autonomy is the worst place to start, and the most tempting.

The real starting point, with a Brazilian number

The discussion about AI in small business usually happens as if it were the future. In Brazil it is not. According to the survey Transformação Digital nos Pequenos Negócios (Digital Transformation in Small Business), run by Sebrae in partnership with the Meta Research Institute between 24 March and 8 May 2026, with 7,182 respondents among sole traders, micro and small companies, 52% of owners had used AI in the two weeks before the interview.

For scale, the same comparison cited by Sebrae puts the figure at 21% among small businesses in the United States, according to a U.S. Census Bureau survey. Brazilian entrepreneurs are adopting these tools faster than American ones, and that is a good number and a dangerous number at the same time.

Dangerous because fast adoption without method is how most of the money gets lost. The typical scene: somebody signs up for a plan, uses it for three weeks to write copy, finds it interesting, forgets about it. The following month the subscription still goes out and nothing in the operation has changed.

The difference between gaining time and merely paying a subscription is not the tool you chose. It is the order in which the decisions were made, and that order is what this article is about.

A subscription signed before the task is defined is the most expensive known way of changing nothing in your operation.

Step 1: one week measuring, before any tool

The right question is not "which AI should I use". It is "where is my team time going". And almost nobody can answer that precisely, because the owner intuition is usually wrong about it.

The exercise costs one week and no money: ask each person to write down, on a sheet of paper, what they did in each half-hour block. No detail, no system, no app. At the end of the week, look at what shows up many times.

What you are looking for is a specific pattern: a task that repeats several times a day, has a similar answer every time and requires no judgment. Answering the same question by message. Moving information from one place to another. Filling in the same form. Looking for a document that exists and nobody can find.

Companies that do this week almost always discover the same thing: a large share of the team time goes to half a dozen tasks nobody considered important enough to complain about. That is where the money is, not in the complex task that looks impressive to automate.

Step 2: choosing the first task with three filters

With the list in hand, apply three filters. If the task does not pass all three, it is not the first one, however tempting it looks.

Filter 1, does the correct answer already exist in writing? If the information is in a document, a catalog, a procedure or the system history, the AI job is to find and rephrase it, which is what it does best. If the answer lives only in somebody head, the first job is writing it down, and that is not an AI project.

Filter 2, is the mistake cheap and visible? If it comes out wrong, somebody notices immediately and the cost is doing it again. A bad summary is noticed at once. A wrong tax calculation can go months before surfacing, and that is why it does not qualify as a first task.

Filter 3, does it happen several times a day? Automating something that happens once a month never pays for itself, because the effort to set it up is the same and the return is thirty times smaller. Volume is what turns minutes saved into hours saved.

Tasks that usually pass all three: triaging and drafting the first reply to a message, turning meeting audio into action items, filling in a form from loose text, a first draft of a proposal from a template, finding information in an internal document.

• Does the correct answer already exist in writing somewhere?
• If it goes wrong, does somebody notice quickly and is the cost just redoing it?
• Does it happen several times a day, rather than once a month?

Step 3: two weeks in suggestion mode

This is where most people skip ahead, and where the project is decided. Before automating anything, run the task with the AI suggesting and a person deciding, for two weeks.

That looks like a waste of time and is the opposite: it is the cheapest test there is. Two things need to be recorded, and they fit in a simple spreadsheet. How many suggestions were used unchanged and how long the task used to take and takes now.

The numbers say what to do next, without debate. A high and stable acceptance rate means the task was well chosen and you can increase autonomy. A low rate means the problem is not the tool: either the task requires judgment nobody mapped, or the instruction is incomplete, or the supporting material does not exist.

And there is a side benefit that is usually the most valuable: over those two weeks, the team discovers where the company information is a mess. Questions the AI cannot answer are exactly the ones your documentation does not answer. That report is worth having on its own.

Step 4: only then choose the tool

Choosing a tool before defining the task is the most reliable way to spend with no return, and it is where almost everyone starts.

With the task defined, the choice becomes objective. If the work happens inside email and spreadsheets, the assistant already there wins simply by not requiring a context switch. If it is answering customers based on in-house material, the design is RAG, and the decision is more about architecture than about brand. If it is programming, that is a different discussion.

For most first tasks any of the main assistants will do, and insisting on picking the best one before starting is just delay. The structured comparison is in ChatGPT, Claude or Gemini, and the criteria for not being fooled by rankings are in which AI is best today.

One decision that has to be made now and not later: a business plan, not a personal account. The difference is not features, it is the contract: use of your content in training, retention period and access control. If customer data is going through it, that is an obligation and not a preference, as detailed in LGPD and AI.

What it really costs

The subscription is the small, visible part. The real cost has three components, and two of them appear in no proposal.

The tool. Per-seat subscriptions for day-to-day use, or usage-based billing when the AI lives inside a system of yours. For a small team starting out, that is usually the smallest line in the bill.

Organizing the material. This is the cost that surprises people. Before the AI can answer based on your documents, somebody has to decide which version of each document counts, convert whatever is a photo of a scanned page, and write down what was never written. That work is human, it is tedious, and it is what determines the result.

The review in the first weeks. Somebody has to look at the outputs and adjust. That decreases over time and does not go to zero, and anyone who budgeted assuming supervision costs nothing budgeted wrong.

The rule of thumb that prevents disappointment: if the project only pays off on the optimistic estimate, it does not pay off. Project the cost, multiply by two, and see whether it still makes sense. If it does, it is a good project. If it does not, choose another task before spending.

The five mistakes that cost the most

Starting with customer service at full autonomy. It is the most tempting because it is what shows up in demos, and it is the worst place to start: the mistake is public, the cost is the customer relationship and the variety of situations is at its maximum. Start inside, where mistakes stay at home.

Automating a bad process. If the process is already confusing, AI will execute the confusion faster and at greater scale. It is worth fixing first, and sometimes the simple exercise of describing the process in order to automate it already reveals that half the steps do not need to exist.

Buying out of fear of falling behind. Signing up for five tools because a competitor mentioned using them is not a strategy. One task solved is worth more than five tools subscribed.

Not telling the team. An AI project rolled out with no conversation produces silent resistance and passive sabotage, because people assume it is about redundancies. Saying clearly what changes and what does not is part of the project, not a courtesy.

Not measuring anything. Without the before number there is no way to know whether it improved, and the discussion becomes opinion. Writing down how long the task used to take is the cheapest and most ignored item on the whole list.

What to do in the next thirty days

A concrete plan, for a company that has not done anything structured yet.

Week 1: measure. Write down where the team time is going, changing nothing and hiring nothing.

Week 2: choose a task with the three filters, write the correct answer for its five most common cases, and decide who will follow it.

Weeks 3 and 4: run it in suggestion mode, with one person reviewing, recording the acceptance rate and the time. At the end you have a number, not an impression.

With that number, the next decision makes itself: scale the same task, increase the autonomy, or switch tasks. And most importantly, you learned to run the cycle, which is the asset that serves every task after this one.

If you would rather have that diagnosis done by people who have already built this in production, with the task chosen and the cost projected before you hire anything, that is exactly the first stage of ROO3 AI consulting. And if your bottleneck is earlier than that, in not enough people arriving, the subject is a different one: paid media and SEO.

Frequently asked questions

How many Brazilian small businesses already use AI?

According to the Digital Transformation in Small Business survey, run by Sebrae with the Meta Research Institute between 24 March and 8 May 2026 with 7,182 respondents, 52% of small business owners had used AI in the two weeks before the interview. The survey cites 21% for small businesses in the United States.

What should be the first task I automate with AI?

The one that passes three filters at once: the correct answer already exists in writing somewhere, the mistake is cheap and quickly visible, and the task happens several times a day. Message triage, filling in forms from loose text and searching internal documents usually pass all three.

What does it cost to start using AI in a small company?

The subscription is the smallest part. The real cost includes organizing the material the AI will use, which means deciding which version of each document counts and writing down what was never written, plus the review time in the first weeks. A useful rule is to project the cost and multiply by two before deciding.

Do I need to hire a developer?

For the first tasks, usually not: off-the-shelf assistants and configuration will do. Programming comes in when the AI needs to talk to your systems, answer based on a large body of documents, or run without somebody watching each execution.

Will AI replace my team?

In practice it absorbs routine tasks with clear rules and verifiable results, and it still requires people at the judgment calls and the exceptions. The common mistake is sizing the saving as if supervision cost nothing, and it does cost, especially in the first months.

Where should I not start?

With customer service at full autonomy, which is the most tempting and the riskiest, because the mistake is public and the variety of situations is at its maximum. And with anything involving pricing, contractual deadlines or decisions with legal effect, where the mistake only surfaces months later.

Sources

• Agência Sebrae: small businesses in Brazil use AI more than American ones
• Data Sebrae: Digital Transformation in Small Business

Rodrigo Fávaro

Founder of ROO3, a marketing and technology agency in São José do Rio Preto, Brazil. Builds AI products running in production (Tobia, gerar.app, Pense Mercado) and maintains the AI Benchmark, a public ranking of AI models. See ROO3 AI consulting.

X @rodmf LinkedIn rodrigofavaro