Brazil Tax Incentives for AI Projects: What to Check First

Brazil Tax Incentives for AI Projects: What to Check First

To find out whether an AI project can use Brazil's Lei do Bem or a funding line, a company needs to confirm three things in the official sources. First, does it meet the requirements for who can use the incentive. Second, is the work research and development with real novelty and technical uncertainty. Third, does documentation exist to prove it. Using AI does not qualify anything on its own. For Brazil tax incentives for AI projects, the final answer comes from the text of the law, the responsible agencies and a tax accountant or tax lawyer, not from a blog article.

Key takeaways

• The incentives look at the type of work and the profile of the company, not at the word "AI". Configuring an off-the-shelf tool is almost never what they are meant to support.
• The Lei do Bem sets requirements on the company itself, including its tax regime and tax compliance. Check that before you discuss the project.
• Funding lines have their own calls, deadlines and rules. Read the current call on the agency's own page, because conditions change.
• Documentation created during the project is worth more than reconstruction afterward. Record hypotheses, tests, failures and decisions from the start.
• No internal check guarantees eligibility. It only tells you whether the case is worth taking to a specialist.

What does the Lei do Bem require from an AI project?

The Lei do Bem is the popular name for the technological innovation incentives chapter of Law 11.196. It lets companies that meet the stated conditions use tax benefits tied to research, development and innovation activities. The full text is on the Brazilian government's official legislation portal, and that is where your reading should start, not with third-party summaries.

There are two blocks of questions. The first is about the company: which tax regime it is under, whether it has taxable profit in the period, whether its tax obligations are current, and whether it has a routine for reporting to the agency responsible for the program. The second is about the project: what exactly was developed, what the technical doubt was at the start, and why the solution was not obvious to a professional in the field.

For AI, that second block is usually the harder one. Subscribing to a virtual assistant tool and configuring its answers is use of existing technology. Creating a new method to classify technical documents in a specific sector, testing approaches that fail, measuring and iterating may have the character of development. The line between the two cases is a matter of interpretation, which is why the decision should not be made by the technology team alone.

It is also worth checking what the law's regulation and the responsible agency's manuals say about concepts such as technological innovation, basic research, applied research and experimental development. They define the vocabulary your report will have to use.

Which funding lines exist besides the Lei do Bem?

A tax incentive is one path. Another is financing or innovation grants, which depend on public calls. Some official starting points:

• Finep, which brings together financing, calls and programs aimed at innovation in companies.
• BNDES, which maintains credit lines and programs, some focused on technology and digitalization.
• State research support foundations, which often run programs for small, innovative companies in their state.
• Research institutions and innovation networks that work in partnership with companies, such as Embrapii, when the project involves cooperation with a research institution.

Do not treat this list as definitive. Programs open, close and change their rules. The reliable source is the agency's own page with the call currently in force, read in full, including the annexes on criteria and eligible expenses.

One practical detail: each line has its own concept of an eligible project. One call may require a partnership with a scientific institution, another may require the company to contribute its own funds, another may limit the size of the applicant. Reading the call before you build the proposal prevents you from writing a polished project that fits no rule.

How do I know if my project is development or just tool usage?

An honest way to test it is to answer simple questions in writing, without marketing language:

• What technical problem had no known solution for us when we started?
• Which alternatives did we test, and why did some not work?
• What did we build that is new, beyond buying and configuring a product?
• How did we measure whether it improved, and against what baseline?
• Who worked on it, during which period and on which activities?

If the answers are vague, the project is probably technology adoption. That is legitimate and useful for the business, but it is probably not the focus of these incentives. If the answers are concrete and there is a record of tests and failures, it is worth taking the material to a specialist.

The opposite risk exists too: describing routine work as innovation. People who review reports know this pattern. An honest text about a smaller project tends to be more defensible than an inflated text about an ordinary one.

Checklist before you talk to an accountant or consultant

Use this list to arrive at the conversation with material, not just an idea.

• Company data: tax regime, tax standing, and history of prior incentive use, if any.
• Problem description: one page, no jargon, on what the project tries to solve and for whom.
• State of the art: what already exists on the market and why it does not solve your case.
• Technical hypotheses: what you intend to test and how you will know it worked.
• Team: who participates, under what relationship (employee, partner, contractor) and for how long.
• Records: code repository, meeting notes, test results, model and prompt versions, discarded decisions.
• Data and privacy: which data enters the project, the legal basis under LGPD, and whether anything is sent abroad. If so, see international data transfer under LGPD when using AI providers.
• Expenses: what is payroll, what is contracted services, what is infrastructure, separated by project.
• Expected result: which success metric was agreed before you started.

Step by step: checking the official sources

This promises nothing. It is a verification routine that can be done in a few weeks.

• Step 1: read the innovation incentives chapter of Law 11.196 and note the requirements that concern the company.
• Step 2: locate the regulation and the manuals of the agency that receives program information. Note the definitions of research and development and the format of the annual report.
• Step 3: run the company check with your accountant. If the company does not meet the starting requirements, the project does not change that.
• Step 4: look at the open calls at Finep, BNDES and your state's research support foundation. Save each call as a PDF with the date you read it.
• Step 5: compare your project with the criteria of each source, item by item, in a simple spreadsheet with three columns: meets, does not meet, not sure.
• Step 6: take the "not sure" items to a specialist. Ask what evidence they would need to see.
• Step 7: decide whether the project continues as is, changes scope to fit better, or goes ahead without an incentive.

Step 7 matters. An AI project that makes sense for the business does not need an incentive to exist. The incentive is a possible consequence, not a reason to invent a project.

Illustrative scenario: a component manufacturer and visual inspection

This is an illustrative scenario, unrelated to any real company. A parts manufacturer wants to reduce errors in visual inspection of surface finish. The team tests an off-the-shelf vision model, notices it fails on specific glossy surfaces, builds its own set of labeled images, compares three approaches and documents where each one fails.

In this case there is a concrete technical problem, hypotheses, tests with negative results and a method built by the team. That gives a specialist material to assess whether the project may fit. Compare it with another company that buys a ready-made chatbot and configures frequently asked questions: it is a good customer service project, but the report would have little to say about technical uncertainty.

Neither situation is better for the business by definition. The difference is only in how well each fits what the incentives describe. For ideas on where to start in industry, see how to start with AI in manufacturing: maintenance and quality first.

How map, limit, pilot and hand over helps with documentation

The working method for AI in companies starts by mapping the current process, then limiting data and actions, piloting a real flow and handing over the operation. It produces, with no extra effort, a good part of the evidence any assessment asks for.

• Mapping creates the baseline: how the process worked before, at what volume and with what failures.
• Limiting records scope decisions, the data used and the risks accepted, which connects to LGPD and to the Brazil AI regulation: what companies should watch and do now.
• Piloting produces logs, versions and a success metric agreed in advance. To structure this, see how to measure an AI automation pilot.
• Handing over leaves operating documentation, owners and a rule for when to expand, adjust or stop.

This does not make any project eligible. It only avoids the most common problem: discovering, at the moment of justification, that nobody kept a record of what was tried.

Common mistakes when seeking incentives for AI

• Starting from the incentive and only then looking for a project that fits.
• Confusing tool usage with development.
• Failing to separate project hours and expenses from routine ones.
• Forgetting that project data needs a legal basis and careful access control, including when the team measures productivity. There is a guide on how to measure AI productivity without surveillance on your team.
• Using third-party summaries, this one included, as the final source. Always cite the official text.
• Waiting until the end of the year to gather evidence.

Frequently asked questions

Do AI projects automatically qualify for the Lei do Bem? No. The law does not treat AI as a category. What matters is whether the company meets the requirements and whether the work counts as research, development or technological innovation under the rule. The analysis is case by case, with support from an accountant or tax lawyer.

Can deploying an AI agent on WhatsApp qualify? It depends on what is done. Buying a platform and configuring flows tends to be technology adoption. If there is development with real technical uncertainty and documented testing, a specialist analysis may be worthwhile. Do not assume eligibility without that assessment.

Do I need an incentives consultant to get started? For the initial check, no: reading the law, the regulation and the calls already helps you filter. To decide eligibility and prepare reports, support from an accountant or tax specialist is advisable, because the risk of misreading the rules falls on the company.

What if the project does not fit any incentive? Go ahead with it if it makes sense for the business. A well-measured pilot, with clear limits on data and actions, delivers value without depending on a tax benefit. If the scope changes later, the documentation you have already kept makes a new analysis easier.

Sources

• Planalto: Lei 11.196, incentivos à inovação tecnológica
• Finep: financiamento e chamadas para inovação
• BNDES: linhas e programas de financiamento