Your company does not need to wait for an AI law to have obligations around generative AI. Brazil's data protection law, the LGPD, already applies to any personal data that goes through a prompt, a knowledge base used for RAG, or an agent with access to your CRM. The useful question is not "can we use it?" but "could we explain tomorrow where every piece of personal data that went into these tools ended up?"
This article is management guidance, not legal advice. Decisions on legal basis and contracts should go through your data protection officer and legal counsel.
What is in force and what is still a bill
Brazil's AI framework bill, PL 2338/2023, passed the Senate in December 2024 and reached the Chamber of Deputies in March 2025. As of September 2026, the Chamber still lists it as "awaiting the rapporteur's opinion in the special committee," with dozens of related bills attached. None of its provisions creates an obligation today.
Meanwhile, the data protection authority got stronger. Law 15.352/2026, signed in February 2026, turned the ANPD into a regulatory agency with financial autonomy and 200 new specialist positions. And it did not wait for a new law to act on AI. In July 2024 it ordered a preventive suspension of Meta's use of Brazilian users' data to train its AI, with a daily fine of R$ 50,000 for non-compliance. The issues it raised were an inadequate legal basis, unclear information to data subjects, excessive limits on the exercise of data subject rights, and processing of children's data without safeguards. Meta resumed only after meeting the ANPD's requirements at the end of August that year.
That case is the best map of what the ANPD looks at when AI is involved: legal basis, transparency and data subject rights. None of these depends on PL 2338.
Where personal data gets in without anyone noticing
Most of the exposure does not come from a formal AI project. It comes from everyday use.
In the prompt. In April 2023, Samsung staff uploaded internal source code to ChatGPT, and in May the company temporarily banned generative AI tools. The everyday version is the analyst who pastes a spreadsheet of overdue customers to "summarize the most critical cases." That is sharing personal data with a third party, with no adequate contract and no record. Microsoft and LinkedIn's 2024 Work Trend Index, based on 31,000 workers in 31 countries, found that 78% of AI users bring their own tools to work. A ban on paper does not change that behavior.
In the knowledge base. An internal RAG assistant indexes contracts, tickets and emails. If nobody filtered them first, the index now holds tax IDs, employees' health information and confidential clauses, and anyone with access to the assistant can retrieve them with a well-phrased question.
In an agent that can take actions. An agent that reads the support inbox and updates the CRM processes personal data on every run. It needs a defined purpose, minimal access and a trail of what it did.
At the vendor. Where the provider stores prompts and outputs, for how long, and whether it trains on them. Business plans usually change the answer: OpenAI, for example, states that it does not train on Enterprise, Business or API data by default. The free plan your team uses on their phones follows different rules. And because processing usually happens outside Brazil, international transfer rules apply: ANPD Resolution 19/2024 approved standard contractual clauses, and the deadline to add them to contracts relying on that mechanism was August 23, 2025.
Legal basis: the first question you will get
For every use case, someone has to state which of the LGPD's Article 7 legal bases supports the processing. In generative AI the usual candidates are performance of a contract (the agent answers a customer about their own order) and legitimate interest (internal analysis of tickets to improve service).
Legitimate interest is not a blank check. The ANPD's guidance, published in February 2024, calls for a three-step balancing test: legitimate purpose, necessity of the data, and weighing against the data subject's expectations, plus safeguards. It also does not cover sensitive data. If your use case touches health, biometrics or racial origin, the conversation is different and narrower.
A practical test: if the data subject knew their data was used this way, would they be surprised? If the honest answer is yes, the legal basis is probably weak, and the transparency certainly is.
Article 20: a review, yes; a human one, not necessarily
There is a mistake that shows up in many presentations on this topic. Article 20 of the LGPD gives data subjects the right to request a review of decisions made solely through automated processing that affect their interests, including consumer and credit profiling. The original text required review "by a natural person." That requirement was removed by Law 13.853/2019, and Congress upheld the veto that kept it out.
That is less of a relief than it sounds. Paragraph 1 requires the controller to provide, on request, clear information about the criteria and procedures behind the decision. Paragraph 2 allows the ANPD to audit for discriminatory aspects when a company refuses to explain on trade secret grounds. And regulating this article is on the agency's agenda: Technical Note 12/2025 consolidated 124 contributions to its public consultation on AI and automated decision review. Contributors converged on impact assessments, transparency and governance; no regulation has been published yet.
In practice, if an agent denies a refund, prioritizes a collections queue or ranks a job applicant, you need to be able to explain the criteria. Human review is still the simplest way to answer a review request well, even where it is not mandatory.
An example: the agent that reads customer emails
Picture a distributor that wants an agent to read the support inbox, spot exchange requests and open a ticket in the ERP. Before writing any code, these questions need written answers:
- Purpose: open exchange tickets. The agent does not reuse the content for anything else, such as marketing profiles.
- Legal basis: performance of the contract with the customer who asked for the exchange. Emails about other topics that land in the same inbox must be discarded by the agent, not stored.
- Minimal data: name, order number, product, reason. Tax ID only if the ERP requires it, and masked in logs.
- Model provider: business plan, no training on your data, a contract with international transfer clauses and a defined retention period.
- Decision: the agent opens the ticket, but any denial goes to a person. That takes Article 20 off the table for this flow.
- Records: every run stores which email was read, what data was extracted, what action was taken and which agent version acted. This also covers the duty to keep records of processing operations under Article 37.
None of this blocks the project. It takes one or two meetings and saves rework when legal or a customer starts asking questions.
A 30-day checklist
- Inventory what is already in use: contracted generative AI tools, browser extensions, assistants built into software you already pay for. Ask the teams, without a witch hunt.
- Classify each use by the data it touches: no personal data, regular personal data, sensitive data, or data about children and adolescents.
- Choose one official tool on a business plan and publish a short rule on what can and cannot go into it. One page, with examples from your company's daily work.
- Review contracts with AI providers: training use, retention, processing location and standard clauses for international transfer.
- For every use that decides something about people, document the criteria and the review path. If the risk is high, produce a data protection impact assessment.
- Add AI to privacy training. The analyst with the overdue-customers spreadsheet means no harm; nobody showed them where the line is.
Banning costs more than it seems
The instinctive reaction, like Samsung's in 2023, is to block everything. It works for a few weeks. Then usage moves to personal phones, where the company has no contract, no logs and no control. Rolling out a business tool with clear rules costs licenses and a few hours of governance, and it brings usage back inside a perimeter you can audit. In our view it is the cheapest option for almost every mid-sized company.
This is what we call cyborg architecture applied to data: people decide where AI may touch personal data, and AI operates inside those limits. It does not decide on its own where the line sits.
Next step
If nobody in your company can list today which AI tools touch customer data, start with item 1 of the checklist this week. If you want help building the inventory, the policy and the review criteria, that is what Governance As a Service does. Our trust page shows how we handle personal data in our own operation.