Brazil AI Bill PL 2338 and Superintelligence: What Managers Should Track

Brazil AI Bill PL 2338 and Superintelligence: What Managers Should Track

Brazil's PL 2338 proposes a general AI law. Here is what it covers for companies, what to track in Congress and what is still undecided.

•

Want to apply this in your company?

Book a call with Juan →

PL 2338 is the bill in Brazil's Congress that aims to create a general legal framework for developing and using artificial intelligence systems. For companies, the core of the text is its risk-based logic, the rights of people affected by automated decisions, and its governance and transparency duties. Because the bill can change while it moves through Congress, the manager's job is to track the current version and organize operations around principles that are likely to last.

Key takeaways

  • The bill sets general rules for AI systems. It does not give a closed list of tools that a company must register.
  • Obligations tend to grow with the impact a system has on people, so classifying your processes is a management task.
  • People subject to automated decisions are likely to get the right to information, explanation and human review, in line with what the LGPD already provides.
  • Documentation, decision records and human oversight are central for anyone running AI agents in production.
  • The debate about superintelligence exists, but it is not the operational focus of the text for the daily work of a small or mid-sized company.

For a broader view of how the bill fits with other rules, see Brazil AI regulation: what companies should watch and do now.

What PL 2338 is and why managers should read it

PL 2338 was built from a preliminary draft prepared by a commission of legal scholars and was discussed in public hearings in the Federal Senate. It is not a law for programmers only. It addresses those who develop, supply and use AI systems, and that includes the company that simply hires a customer service agent on WhatsApp or a tool that screens job applications.

Before making any decision, check the official pages of the Federal Senate or the Chamber of Deputies to see which stage the text is in and which version is current. This article describes the themes of the bill, not its current status, which may change.

How the bill thinks about risk

The central idea of a framework like this is proportionality: the greater a system's impact on rights, safety, access to services or opportunities, the heavier the requirements. A bot that tells customers opening hours is very different from a system that declines credit, selects candidates or decides on a collection action.

For a manager, this means classifying your own processes before debating the law in detail. A few questions help:

  • What decision does the AI make or influence: does it only inform, recommend, or make the final call?
  • Who is affected: customers, candidates, employees, suppliers?
  • What is the possible harm if the system gets it wrong, and can that harm be reversed?
  • Does the flow involve personal data, and among it, sensitive data?
  • Is there a responsible person who can review the decision and pause the flow?

The answers usually show where the company needs more control, whatever final form the law takes.

Rights of people subject to automated decisions

Brazil already has a foundation on this topic. The LGPD addresses the review of decisions made solely on the basis of automated processing of personal data that affect the interests of the data subject, and the relevant article is a good starting point for any AI operation. PL 2338 discusses related themes, such as informing people that a system is in use, explaining results in understandable terms and allowing people to contest them. The final version may adjust these rights, so read the official text rather than relying only on summaries.

In practice, a flow that denies a request needs three things: a clear notice that automation was involved, a reason the human agent can repeat to the customer, and a path to request human review. If the agent cannot explain its own decision, its authority at that point should be limited. We go deeper on this in LGPD automated decision-making and the right to review (Art. 20).

Governance and documentation in daily work

The NIST AI Risk Management Framework organizes AI risk management into functions of governing, mapping, measuring and managing. It is not Brazilian law, but it offers a practical structure that fits well with what the bill is likely to require: knowing what systems exist, who is accountable for each one, how performance is measured, and what happens when something fails.

A simple plan to start:

  • List every AI system in use, including those built into third-party tools, such as AI features in a CRM or a helpdesk.
  • For each one, record its purpose, the data it uses, the person responsible and who can switch it off.
  • Classify impact in simple levels: informational, operational and decisional.
  • Define what the system may not do on its own, such as approving exceptions, denying refunds or closing a support case without review.
  • Keep records of relevant interactions and failures, and review that material at regular intervals.

On security, OWASP maintains a list of the most common risks in applications built on language models, including prompt injection and giving an agent more autonomy than its task requires. These risks show up in any serious regulatory framework because they are concrete technical failures, not theoretical concerns.

Superintelligence: what belongs in the debate and what does not

The word superintelligence appears often in news and social media, usually to describe systems with capabilities far beyond human ones in nearly every area. The topic is discussed in academic research and international forums, and it is natural for managers to wonder whether it already affects their operation.

A bill like PL 2338 focuses on existing AI systems and the concrete uses they have. If the version under discussion mentions themes of larger scale, read the official text, not third-party interpretations. For day-to-day decisions, what matters is not confusing the long-term debate with today's operations, which involve customer service, collections, scheduling and screening. To separate the concepts, read what ChatGPT superintelligence is and is not.

What to track while the bill moves forward

Some changes to the text will matter more to your company than others. Prioritize the following:

  • The bill's status in each legislative house, checked directly on the official pages.
  • The definitions of AI system and high-risk system, because they determine the scope of everything that follows.
  • Rules on informing users, explaining decisions and human review.
  • The role of the national data protection authority (ANPD) and any coordinating bodies, and how enforcement will work in practice.
  • Adaptation deadlines, which may vary by company size or type of use.
  • Rules on data used to train models and on copyrighted content, if they remain in the final text.

The ANPD's materials are the most useful source for the data protection side, and the official websites of the Chamber and the Senate are the place to check the bill's status.

Illustrative scenario: a furniture store with a WhatsApp agent

Illustrative scenario: a furniture store plans a WhatsApp agent to answer questions about delivery times, order status and return policies. The manager maps the flows and sees three levels of risk. Answering delivery times and order status has low impact and can be automated with logging. Explaining the return policy is useful, but the agent should quote the official policy text and not invent exceptions. Approving or denying a return outside the rules directly affects the customer, so that point needs human review and a record of the reason.

With this separation, the company does not need to wait for the bill to be finalized before starting. It can limit the agent to low-impact flows, give it access only to the information it needs and leave a clear path to a human agent. For the channel setup, consider the rules on the 24 hour WhatsApp customer service window.

The way we work follows four steps: map the current process and its baseline, limit data and actions, pilot a real flow with records and a success metric agreed in advance, and then hand over the operation and decide whether the pilot is expanded, adjusted or stopped. This approach does not depend on the outcome of one specific law, because it is about control, and control is required in any regulatory scenario.

Frequently asked questions

Does PL 2338 already apply to companies? This article does not state the bill's current status, because it changes during the legislative process. Check the official Senate and Chamber of Deputies pages. Until a law is approved and published, the practical reference is the LGPD and the sector rules already in force.

Do I need to change anything in my company now? Practices such as inventorying your AI systems, assigning an owner, recording relevant decisions and ensuring human review at sensitive points make sense regardless of how the law turns out. They also reduce risks under the LGPD and improve the security of the tools you already use.

Does PL 2338 deal with superintelligence? The bill focuses on current AI systems and their uses. Superintelligence is a long-term debate, separate from the routine of customer service, sales and operations. If you want shared vocabulary for your leadership team, the superintelligence glossary for executives helps separate the terms.

Where should I start without a large legal team? Start with mapping: list the processes that use AI, classify the impact of each one and limit what the system can do on its own. Then test one low-impact flow with records. For higher-impact decisions, a lawyer or your legal team should validate the design before the operation goes live.

Sources

AI regulationPL 2338LGPDAI governanceBrazil AI law