
Brazil AI Bill PL 2338: What Companies Should Monitor Now
Brazil AI bill PL 2338 for business leaders: operations impact, rights of affected people, documentation, and what to monitor right now.
Want to apply this in your company?
Book a call with Juan →PL 2338 is the Brazilian bill that aims to set general rules for the development and use of artificial intelligence in Brazil, with a focus on higher-risk systems and on the rights of the people those systems affect. For a company, the useful question is not whether the AI is superintelligence. It is whether the system influences decisions about customers, employees or suppliers, and what obligations that could create. This article summarizes what is likely to matter for managers and points to what to watch, without predicting when the bill will be approved.
Key takeaways
- The bill uses a risk-based logic: the greater a system's impact on rights and opportunities, the stricter the requirements are likely to be.
- People affected by automated decisions are likely to have rights to information, explanation, challenge and human review.
- Companies that develop or operate AI systems may need documentation, records and governance proportional to the risk.
- The text provides for oversight and administrative penalties, and how they apply depends on the designated authority and the final approved version.
- The debate is not limited to companies that build models. Those who contract and operate third-party systems also need to know what those systems do.
What PL 2338 is and why managers should track it
PL 2338 is moving through the National Congress, and the text can change at each stage of discussion through amendments and substitute versions. For that reason, the reliable reference is always the current text published on the official websites of the Federal Senate and the Chamber of Deputies, not summaries from third parties, including this one.
The bill deals with AI as it exists today: systems that sort requests, answer customers, prioritize queues and recommend decisions. Superintelligence comes up in debates about the future, but it is not what your operation needs to solve this week. Keeping these two conversations separate prevents leadership from treating a hypothetical topic as an immediate obligation.
Who is affected: the company that uses AI, not only big tech
The practical impact shows up in ordinary routines: a WhatsApp support agent that declines a request, a résumé screening tool, a credit analysis system, or an AI feature built into a management system the team already uses. Many mid-sized companies use AI without having developed anything themselves, simply by contracting ready-made functions.
Hiring a tool does not transfer all responsibility to the vendor. The company remains accountable for how it uses the system, for the data it feeds into it, and for the decisions it lets through without review. Because the LGPD already applies to this data, it is also worth reviewing the materials from ANPD, the Brazilian national data protection authority.
How to tell whether a system weighs heavily in the risk classification
Risk classification is the most important part of the bill for day-to-day operations. Before discussing with lawyers or vendors, the team can answer a simple checklist for each system in use:
- Does the system decide or influence anything that affects access to a service, credit, employment, health, education or another right?
- Can an affected person reverse the result without meaningful cost or embarrassment?
- Does the system use sensitive personal data or information about groups in vulnerable situations?
- Does the decision take effect before any person reviews it?
- Is the party responsible for an error clearly identified?
If the answer is yes to more than one item, the system deserves a closer look before it keeps running without clear rules. This is not a legal verdict. It is a way to decide where the company should look first.
Rights of affected people and automated decisions
A significant part of the debate concerns the right of a person who is affected by an automated decision to know that automation was involved, to understand the reason in understandable terms, and to request a review. The LGPD already covers automated decisions and the right to review them, and our explanation of Article 20 of the LGPD goes deeper on this point.
In practice, this becomes a matter of process design. A customer declined by an AI agent should know that an automated system was involved, have a channel to reach a person, and receive a response that explains the general criteria. If the workflow does not provide for that path, the company will struggle to meet these rights, whatever the final text of the bill says.
Transparency, records and documentation
Documentation requirements tend to grow with the risk of the system. For a company, this means being able to answer a few basic questions with evidence: which model or tool was used, with which instructions, with which data, what actions it can perform, who approved the configuration, and what errors were observed.
Decision and execution logs are the foundation of those answers. Observability concepts, such as those described in the OpenTelemetry documentation, help the technical team think about tracing calls, errors and behavior. For governance vocabulary, the NIST AI Risk Management Framework organizes the work into functions such as govern, map, measure and manage. It can serve as an internal reference even before any formal obligation applies.
What to monitor without predicting a timeline
Because it is not possible to predict when the bill will be voted on or signed into law, the most useful step is to define what to watch and who follows each item. A monitoring list can look like this:
- The current text of the bill on the official Senate and Chamber websites, including amendments and substitute versions.
- The definition of a high-risk system and the list of examples, since this determines the scope of the obligations.
- The rules on explanation, challenge and human review, and how they are expected to work in practice.
- The documentation, registration and impact assessment requirements, and whether they vary by company size or sector.
- The authority responsible for oversight and how it will coordinate with ANPD.
- The rules on protected content used for training, if your company creates, publishes or licenses content.
- Guidance from regulators in your sector, which often details how the general law applies.
Illustrative scenario: a distributor uses an AI agent to prioritize credit requests from its customers. Even without knowing the bill's timeline, the team can map that an automated refusal affects access to a service. It then decides that every denial goes through human review, that the customer is told automation was involved, and that each decision is logged with the version of the rules used. If the final text requires more, the company already has the foundation to adjust.
Does superintelligence belong in this conversation?
Superintelligence describes hypothetical systems with capabilities far beyond human ability across nearly all tasks, and it remains a subject of academic and philosophical discussion. It is not the category most companies operate today. Teams that use support agents, automations and writing assistants are working with narrow AI, which has a defined scope and limits that need to be tested.
It is worth separating what your company runs from what is speculation. A superintelligence glossary for executives helps align terms, and ChatGPT, superintelligence and what changes for companies is a useful starting point for discussing the topic with your team. Following PL 2338 is a governance task about real systems, not about future scenarios.
How to apply this in your operation
In our method, the first step is to map the current process and the points where AI touches decisions, customers or data. Next, we limit the data and actions the system can take, pilot a real workflow with logs and a success metric agreed in advance, and then hand the operation to the team with a decision to expand, adjust or stop. This sequence does not replace a legal analysis, but it puts the company in a position to answer regulatory questions with facts about its own use. Our overview of Brazil's AI regulation and what companies should do now offers a complementary view of what to keep track of.
Frequently asked questions
Do I need to wait for PL 2338 to be approved before organizing how we use AI? No. The LGPD already applies to the processing of personal data today, and practices such as a system inventory, access control, decision records and human review are useful regardless of how the bill ends. Getting organized now reduces the adaptation work later.
Does PL 2338 apply to a small company that only uses off-the-shelf tools? It depends on how the tools are used and how much impact their decisions have, not only on the size of the company. Using a ready-made tool does not transfer all responsibility to the vendor, so you need to know what it decides, with which data, and with what level of review.
Does the bill cover superintelligence? The text focuses on current AI systems and on concrete risks tied to rights and automated decisions. Superintelligence is a speculative topic. For running a company, what matters is the scope of the systems already in use.
What is the first concrete step for a manager? List the AI systems in use, identify who operates them, which decisions they touch, and where the records are kept. Then decide who follows the official text of the bill and how often the team will review that inventory.